diff --git a/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c b/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c
index 4090db8e45a..5d501b6ac4b 100644
--- a/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c
+++ b/src/gallium/auxiliary/gallivm/lp_bld_sample_aos.c
@@ -454,7 +454,7 @@ lp_build_sample_image_nearest(struct lp_build_sample_context *bld,
    LLVMValueRef s_float, t_float = NULL, r_float = NULL;
    LLVMValueRef x_stride;
    LLVMValueRef x_offset, offset;
-   LLVMValueRef x_subcoord, y_subcoord, z_subcoord;
+   LLVMValueRef x_subcoord, y_subcoord = NULL, z_subcoord;
 
    lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32, bld->vector_width));
 
@@ -778,7 +778,7 @@ lp_build_sample_image_linear(struct lp_build_sample_context *bld,
    LLVMValueRef y_offset0, y_offset1;
    LLVMValueRef z_offset0, z_offset1;
    LLVMValueRef offset[2][2][2]; /* [z][y][x] */
-   LLVMValueRef x_subcoord[2], y_subcoord[2], z_subcoord[2];
+   LLVMValueRef x_subcoord[2], y_subcoord[2] = {NULL, NULL}, z_subcoord[2];
    unsigned x, y, z;
 
    lp_build_context_init(&i32, bld->gallivm, lp_type_int_vec(32, bld->vector_width));
diff --git a/src/gallium/auxiliary/util/u_sse.h b/src/gallium/auxiliary/util/u_sse.h
index cae4138ba01..e372d3b6b64 100644
--- a/src/gallium/auxiliary/util/u_sse.h
+++ b/src/gallium/auxiliary/util/u_sse.h
@@ -1,6 +1,6 @@
 /**************************************************************************
  *
- * Copyright 2008 VMware, Inc.
+ * Copyright 2008-2021 VMware, Inc.
  * All Rights Reserved.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
@@ -38,6 +38,8 @@
 #define U_SSE_H_
 
 #include "pipe/p_config.h"
+#include "pipe/p_compiler.h"
+#include "util/u_debug.h"
 
 #if defined(PIPE_ARCH_SSE)
 
@@ -296,6 +298,408 @@ transpose2_64_2_32(const __m128i * restrict a01,
 #define SCALAR_EPI32(m, i) _mm_shuffle_epi32((m), _MM_SHUFFLE(i,i,i,i))
 
 
+/*
+ * Implements (1-w)*a + w*b = a - wa + wb = w(b-a) + a
+ * ((b-a)*w >> 8) + a
+ * The math behind negative sub results (logic shift/mask) is tricky.
+ *
+ * w -- weight values
+ * a -- src0 values
+ * b -- src1 values
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_lerp_epi16(__m128i w, __m128i a, __m128i b)
+{
+   __m128i res;
+
+   res = _mm_sub_epi16(b, a);
+   res = _mm_mullo_epi16(res, w);
+   res = _mm_srli_epi16(res, 8);
+   /* use add_epi8 instead of add_epi16 so no need to mask off upper bits */
+   res = _mm_add_epi8(res, a);
+
+   return res;
+}
+
+
+/* Apply premultiplied-alpha blending on two pixels simultaneously.
+ * All parameters are packed as 8.8 fixed point values in __m128i SSE
+ * registers, with the upper 8 bits all zero.
+ *
+ * a -- src alpha values
+ * d -- dst color values
+ * s -- src color values
+ */
+static inline __m128i
+util_sse2_premul_blend_epi16( __m128i a, __m128i d, __m128i s)
+{
+   __m128i da, d_sub_da, tmp;
+   tmp      = _mm_mullo_epi16(d, a);
+   da       = _mm_srli_epi16(tmp, 8);
+   d_sub_da = _mm_sub_epi16(d, da);
+
+   return  _mm_add_epi16(s, d_sub_da);
+}
+
+
+/* Apply premultiplied-alpha blending on four pixels in packed BGRA
+ * format (one/inv_src_alpha blend mode).
+ *
+ * src    -- four pixels (bgra8 format)
+ * dst    -- four destination pixels (bgra8)
+ * return -- blended pixels (bgra8)
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_blend_premul_4(const __m128i src,
+                         const __m128i dst)
+{
+
+   __m128i al, ah, dl, dh, sl, sh, rl, rh;
+   __m128i zero = _mm_setzero_si128();
+
+   /* Blend first two pixels:
+    */
+   sl = _mm_unpacklo_epi8(src, zero);
+   dl = _mm_unpacklo_epi8(dst, zero);
+
+   al = _mm_shufflehi_epi16(sl, 0xff);
+   al = _mm_shufflelo_epi16(al, 0xff);
+
+   rl = util_sse2_premul_blend_epi16(al, dl, sl);
+
+   /* Blend second two pixels:
+    */
+   sh = _mm_unpackhi_epi8(src, zero);
+   dh = _mm_unpackhi_epi8(dst, zero);
+
+   ah = _mm_shufflehi_epi16(sh, 0xff);
+   ah = _mm_shufflelo_epi16(ah, 0xff);
+
+   rh = util_sse2_premul_blend_epi16(ah, dh, sh);
+
+   /* Pack the results down to four bgra8 pixels:
+    */
+   return _mm_packus_epi16(rl, rh);
+}
+
+
+/* Apply src-alpha blending on four pixels in packed BGRA
+ * format (srcalpha/inv_src_alpha blend mode).
+ *
+ * src    -- four pixels (bgra8 format)
+ * dst    -- four destination pixels (bgra8)
+ * return -- blended pixels (bgra8)
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_blend_srcalpha_4(const __m128i src,
+                           const __m128i dst)
+{
+
+   __m128i al, ah, dl, dh, sl, sh, rl, rh;
+   __m128i zero = _mm_setzero_si128();
+
+   /* Blend first two pixels:
+    */
+   sl = _mm_unpacklo_epi8(src, zero);
+   dl = _mm_unpacklo_epi8(dst, zero);
+
+   al = _mm_shufflehi_epi16(sl, 0xff);
+   al = _mm_shufflelo_epi16(al, 0xff);
+
+   rl = util_sse2_lerp_epi16(al, dl, sl);
+
+   /* Blend second two pixels:
+    */
+   sh = _mm_unpackhi_epi8(src, zero);
+   dh = _mm_unpackhi_epi8(dst, zero);
+
+   ah = _mm_shufflehi_epi16(sh, 0xff);
+   ah = _mm_shufflelo_epi16(ah, 0xff);
+
+   rh = util_sse2_lerp_epi16(ah, dh, sh);
+
+   /* Pack the results down to four bgra8 pixels:
+    */
+   return _mm_packus_epi16(rl, rh);
+}
+
+
+/**
+ * premultiplies src with constant alpha then
+ * does one/inv_src_alpha blend.
+ *
+ * src 16xi8 (normalized)
+ * dst 16xi8 (normalized)
+ * cst_alpha (constant alpha (u8 value))
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_blend_premul_src_4(const __m128i src,
+                             const __m128i dst,
+                             const unsigned cst_alpha)
+{
+
+   __m128i srca, d, s, rl, rh;
+   __m128i zero = _mm_setzero_si128();
+   __m128i cst_alpha_vec = _mm_set1_epi16(cst_alpha);
+
+   /* Blend first two pixels:
+    */
+   s = _mm_unpacklo_epi8(src, zero);
+   s = _mm_mullo_epi16(s, cst_alpha_vec);
+   /* the shift will cause some precision loss */
+   s = _mm_srli_epi16(s, 8);
+
+   srca = _mm_shufflehi_epi16(s, 0xff);
+   srca = _mm_shufflelo_epi16(srca, 0xff);
+
+   d = _mm_unpacklo_epi8(dst, zero);
+   rl = util_sse2_premul_blend_epi16(srca, d, s);
+
+   /* Blend second two pixels:
+    */
+   s = _mm_unpackhi_epi8(src, zero);
+   s = _mm_mullo_epi16(s, cst_alpha_vec);
+   /* the shift will cause some precision loss */
+   s = _mm_srli_epi16(s, 8);
+
+   srca = _mm_shufflehi_epi16(s, 0xff);
+   srca = _mm_shufflelo_epi16(srca, 0xff);
+
+   d = _mm_unpackhi_epi8(dst, zero);
+   rh = util_sse2_premul_blend_epi16(srca, d, s);
+
+   /* Pack the results down to four bgra8 pixels:
+    */
+   return _mm_packus_epi16(rl, rh);
+}
+
+
+/**
+ * Linear interpolation with SSE2.
+ *
+ * dst, src0, src1 are 16 x i8 vectors, with [0..255] normalized values.
+ *
+ * weight_lo and weight_hi should be a 8 x i16 vectors, in 8.8 fixed point
+ * format, for the low and high components.
+ * We'd want to pass these as values but MSVC limitation forces us to pass these
+ * as pointers since it will complain if more than 3 __m128 are passed by value.
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_lerp_epi8_fixed88(__m128i src0, __m128i src1,
+                            const __m128i * restrict weight_lo,
+                            const __m128i * restrict weight_hi)
+{
+   const __m128i zero = _mm_setzero_si128();
+
+   __m128i src0_lo = _mm_unpacklo_epi8(src0, zero);
+   __m128i src0_hi = _mm_unpackhi_epi8(src0, zero);
+
+   __m128i src1_lo = _mm_unpacklo_epi8(src1, zero);
+   __m128i src1_hi = _mm_unpackhi_epi8(src1, zero);
+
+   __m128i dst_lo;
+   __m128i dst_hi;
+
+   dst_lo = util_sse2_lerp_epi16(*weight_lo, src0_lo, src1_lo);
+   dst_hi = util_sse2_lerp_epi16(*weight_hi, src0_hi, src1_hi);
+
+   return _mm_packus_epi16(dst_lo, dst_hi);
+}
+
+
+/**
+ * Linear interpolation with SSE2.
+ *
+ * dst, src0, src1 are 16 x i8 vectors, with [0..255] normalized values.
+ *
+ * weight should be a 16 x i8 vector, in 0.8 fixed point values.
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_lerp_epi8_fixed08(__m128i src0, __m128i src1,
+                            __m128i weight)
+{
+   const __m128i zero = _mm_setzero_si128();
+   __m128i weight_lo = _mm_unpacklo_epi8(weight, zero);
+   __m128i weight_hi = _mm_unpackhi_epi8(weight, zero);
+
+   return util_sse2_lerp_epi8_fixed88(src0, src1,
+                                      &weight_lo, &weight_hi);
+}
+
+
+/**
+ * Linear interpolation with SSE2.
+ *
+ * dst, src0, src1, and weight are 16 x i8 vectors, with [0..255] normalized
+ * values.
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_lerp_unorm8(__m128i src0, __m128i src1,
+                      __m128i weight)
+{
+   const __m128i zero = _mm_setzero_si128();
+   __m128i weight_lo = _mm_unpacklo_epi8(weight, zero);
+   __m128i weight_hi = _mm_unpackhi_epi8(weight, zero);
+
+#if 0
+   /*
+    * Rescale from [0..255] to [0..256].
+    */
+   weight_lo = _mm_add_epi16(weight_lo, _mm_srli_epi16(weight_lo, 7));
+   weight_hi = _mm_add_epi16(weight_hi, _mm_srli_epi16(weight_hi, 7));
+#endif
+
+   return util_sse2_lerp_epi8_fixed88(src0, src1,
+                                      &weight_lo, &weight_hi);
+}
+
+
+/**
+ * Linear interpolation with SSE2.
+ *
+ * dst, src0, src1, src2, src3 are 16 x i8 vectors, with [0..255] normalized
+ * values.
+ *
+ * ws_lo, ws_hi, wt_lo, wt_hi should be a 8 x i16 vectors, in 8.8 fixed point
+ * format, for the low and high components.
+ * We'd want to pass these as values but MSVC limitation forces us to pass these
+ * as pointers since it will complain if more than 3 __m128 are passed by value.
+ *
+ * This uses ws_lo, ws_hi to interpolate between src0 and src1, as well as to
+ * interpolate between src2 and src3, then uses wt_lo and wt_hi to interpolate
+ * between the resulting vectors.
+ */
+static ALWAYS_INLINE __m128i
+util_sse2_lerp_2d_epi8_fixed88(__m128i src0, __m128i src1,
+                               const __m128i * restrict src2,
+                               const __m128i * restrict src3,
+                               const __m128i * restrict ws_lo,
+                               const __m128i * restrict ws_hi,
+                               const __m128i * restrict wt_lo,
+                               const __m128i * restrict wt_hi)
+{
+   const __m128i zero = _mm_setzero_si128();
+
+   __m128i src0_lo = _mm_unpacklo_epi8(src0, zero);
+   __m128i src0_hi = _mm_unpackhi_epi8(src0, zero);
+
+   __m128i src1_lo = _mm_unpacklo_epi8(src1, zero);
+   __m128i src1_hi = _mm_unpackhi_epi8(src1, zero);
+
+   __m128i src2_lo = _mm_unpacklo_epi8(*src2, zero);
+   __m128i src2_hi = _mm_unpackhi_epi8(*src2, zero);
+
+   __m128i src3_lo = _mm_unpacklo_epi8(*src3, zero);
+   __m128i src3_hi = _mm_unpackhi_epi8(*src3, zero);
+
+   __m128i dst_lo, dst01_lo, dst23_lo;
+   __m128i dst_hi, dst01_hi, dst23_hi;
+
+   dst01_lo = util_sse2_lerp_epi16(*ws_lo, src0_lo, src1_lo);
+   dst01_hi = util_sse2_lerp_epi16(*ws_hi, src0_hi, src1_hi);
+   dst23_lo = util_sse2_lerp_epi16(*ws_lo, src2_lo, src3_lo);
+   dst23_hi = util_sse2_lerp_epi16(*ws_hi, src2_hi, src3_hi);
+
+   dst_lo = util_sse2_lerp_epi16(*wt_lo, dst01_lo, dst23_lo);
+   dst_hi = util_sse2_lerp_epi16(*wt_hi, dst01_hi, dst23_hi);
+
+   return _mm_packus_epi16(dst_lo, dst_hi);
+}
+
+/**
+ * Stretch a row of pixels using linear filter.
+ *
+ * Uses Bresenham's line algorithm using 16.16 fixed point representation for
+ * the error term.
+ *
+ * @param dst_width destination width in pixels
+ * @param src_x    start x0 in 16.16 fixed point format
+ * @param src_xstep step in 16.16. fixed point format
+ *
+ * @return final src_x value (i.e., src_x + dst_width*src_xstep)
+ */
+static ALWAYS_INLINE int32_t
+util_sse2_stretch_row_8unorm(__m128i * restrict dst,
+                             int32_t dst_width,
+                             const uint32_t * restrict src,
+                             int32_t src_x,
+                             int32_t src_xstep)
+{
+   int16_t error0, error1, error2, error3;
+   __m128i error_lo, error_hi, error_step;
+
+   assert(dst_width >= 0);
+   assert(dst_width % 4 == 0);
+
+   error0 = src_x;
+   error1 = error0 + src_xstep;
+   error2 = error1 + src_xstep;
+   error3 = error2 + src_xstep;
+
+   error_lo   = _mm_setr_epi16(error0, error0, error0, error0,
+                               error1, error1, error1, error1);
+   error_hi   = _mm_setr_epi16(error2, error2, error2, error2,
+                               error3, error3, error3, error3);
+   error_step = _mm_set1_epi16(src_xstep << 2);
+
+   dst_width >>= 2;
+   while (dst_width) {
+      uint16_t src_x0;
+      uint16_t src_x1;
+      uint16_t src_x2;
+      uint16_t src_x3;
+      __m128i src0, src1;
+      __m128i weight_lo, weight_hi;
+
+      /*
+       * It is faster to re-compute the coordinates in the scalar integer unit here,
+       * than to fetch the values from the SIMD integer unit.
+       */
+
+      src_x0 = src_x >> 16;
+      src_x += src_xstep;
+      src_x1 = src_x >> 16;
+      src_x += src_xstep;
+      src_x2 = src_x >> 16;
+      src_x += src_xstep;
+      src_x3 = src_x >> 16;
+      src_x += src_xstep;
+
+      /*
+       * Fetch pairs of pixels 64bit at a time, and then swizzle them inplace.
+       */
+
+      {
+         __m128i src_00_10 = _mm_loadl_epi64((const __m128i *)&src[src_x0]);
+         __m128i src_01_11 = _mm_loadl_epi64((const __m128i *)&src[src_x1]);
+         __m128i src_02_12 = _mm_loadl_epi64((const __m128i *)&src[src_x2]);
+         __m128i src_03_13 = _mm_loadl_epi64((const __m128i *)&src[src_x3]);
+
+         __m128i src_00_01_10_11 = _mm_unpacklo_epi32(src_00_10, src_01_11);
+         __m128i src_02_03_12_13 = _mm_unpacklo_epi32(src_02_12, src_03_13);
+
+         src0 = _mm_unpacklo_epi64(src_00_01_10_11, src_02_03_12_13);
+         src1 = _mm_unpackhi_epi64(src_00_01_10_11, src_02_03_12_13);
+      }
+
+      weight_lo = _mm_srli_epi16(error_lo, 8);
+      weight_hi = _mm_srli_epi16(error_hi, 8);
+
+      *dst = util_sse2_lerp_epi8_fixed88(src0, src1,
+                                         &weight_lo, &weight_hi);
+
+      error_lo = _mm_add_epi16(error_lo, error_step);
+      error_hi = _mm_add_epi16(error_hi, error_step);
+
+      ++dst;
+      --dst_width;
+   }
+
+   return src_x;
+}
+
+
+
 #endif /* PIPE_ARCH_SSE */
 
 #endif /* U_SSE_H_ */
diff --git a/src/gallium/auxiliary/util/u_tests.c b/src/gallium/auxiliary/util/u_tests.c
index d50dc80ff48..aa381448b05 100644
--- a/src/gallium/auxiliary/util/u_tests.c
+++ b/src/gallium/auxiliary/util/u_tests.c
@@ -447,7 +447,7 @@ util_test_constant_buffer(struct pipe_context *ctx,
             "MOV OUT[0], CONST[0][0]\n"
             "END\n";
       struct tgsi_token tokens[1000];
-      struct pipe_shader_state state;
+      struct pipe_shader_state state = {0};
 
       if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
          puts("Can't compile a fragment shader.");
@@ -743,7 +743,7 @@ test_texture_barrier(struct pipe_context *ctx, bool use_fbfetch,
    }
 
    struct tgsi_token tokens[1000];
-   struct pipe_shader_state state;
+   struct pipe_shader_state state = {0};
 
    if (!tgsi_text_translate(text, tokens, ARRAY_SIZE(tokens))) {
       assert(0);
diff --git a/src/gallium/drivers/llvmpipe/lp_clear.c b/src/gallium/drivers/llvmpipe/lp_clear.c
index 3fd67b79dac..9345d910144 100644
--- a/src/gallium/drivers/llvmpipe/lp_clear.c
+++ b/src/gallium/drivers/llvmpipe/lp_clear.c
@@ -38,6 +38,7 @@
 #include "lp_setup.h"
 #include "lp_query.h"
 #include "lp_debug.h"
+#include "lp_state.h"
 
 
 /**
@@ -57,6 +58,8 @@ llvmpipe_clear(struct pipe_context *pipe,
    if (!llvmpipe_check_render_cond(llvmpipe))
       return;
 
+   llvmpipe_update_derived_clear(llvmpipe);
+
    if (LP_PERF & PERF_NO_DEPTH)
       buffers &= ~PIPE_CLEAR_DEPTHSTENCIL;
 
diff --git a/src/gallium/drivers/llvmpipe/lp_context.c b/src/gallium/drivers/llvmpipe/lp_context.c
index 361634ed1bd..35f3618e8d7 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.c
+++ b/src/gallium/drivers/llvmpipe/lp_context.c
@@ -284,6 +284,9 @@ llvmpipe_create_context(struct pipe_screen *screen, void *priv,
    draw_wide_point_threshold(llvmpipe->draw, 10000.0);
    draw_wide_line_threshold(llvmpipe->draw, 10000.0);
 
+   /* initial state for clipping - enabled, with no guardband */
+   draw_set_driver_clipping(llvmpipe->draw, FALSE, FALSE, FALSE, TRUE);
+
    lp_reset_counters();
 
    /* If llvmpipe_set_scissor_states() is never called, we still need to
diff --git a/src/gallium/drivers/llvmpipe/lp_context.h b/src/gallium/drivers/llvmpipe/lp_context.h
index b1adba61db7..22f177eee24 100644
--- a/src/gallium/drivers/llvmpipe/lp_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_context.h
@@ -156,6 +156,9 @@ struct llvmpipe_context {
    unsigned nr_fs_variants;
    unsigned nr_fs_instrs;
 
+   boolean permit_linear_rasterizer;
+   boolean single_vp;
+
    struct lp_setup_variant_list_item setup_variants_list;
    unsigned nr_setup_variants;
 
diff --git a/src/gallium/drivers/llvmpipe/lp_debug.h b/src/gallium/drivers/llvmpipe/lp_debug.h
index d76b9be7544..2262cdb0bba 100644
--- a/src/gallium/drivers/llvmpipe/lp_debug.h
+++ b/src/gallium/drivers/llvmpipe/lp_debug.h
@@ -47,6 +47,11 @@
 #define DEBUG_CS            0x10000
 #define DEBUG_TGSI_IR       0x20000
 #define DEBUG_CACHE_STATS   0x40000
+#define DEBUG_NO_FASTPATH   0x80000
+#define DEBUG_LINEAR        0x100000
+#define DEBUG_LINEAR2       0x200000
+#define DEBUG_SHOW_DEPTH    0x400000
+#define DEBUG_ACCURATE_A0   0x800000 /* verbose */
 
 /* Performance flags.  These are active even on release builds.
  */
@@ -58,6 +63,8 @@
 #define PERF_NO_BLEND       0x20  	/* disable blending */
 #define PERF_NO_DEPTH       0x40  	/* disable depth buffering entirely */
 #define PERF_NO_ALPHATEST   0x80  	/* disable alpha testing */
+#define PERF_NO_RAST_LINEAR 0x100  	/* disable linear rast */
+#define PERF_NO_SHADE       0x200  	/* disable fragment shaders */
 
 
 extern int LP_PERF;
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.c b/src/gallium/drivers/llvmpipe/lp_jit.c
index c3bfe3b15b3..3c763240a97 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.c
+++ b/src/gallium/drivers/llvmpipe/lp_jit.c
@@ -39,6 +39,7 @@
 #include "gallivm/lp_bld_debug.h"
 #include "gallivm/lp_bld_format.h"
 #include "lp_context.h"
+#include "lp_screen.h"
 #include "lp_jit.h"
 
 static LLVMTypeRef
@@ -189,6 +190,7 @@ lp_jit_create_types(struct lp_fragment_shader_variant *lp)
    struct gallivm_state *gallivm = lp->gallivm;
    LLVMContextRef lc = gallivm->context;
    LLVMTypeRef viewport_type, texture_type, sampler_type, image_type;
+   LLVMTypeRef linear_elem_type;
 
    /* struct lp_jit_viewport */
    {
@@ -314,6 +316,74 @@ lp_jit_create_types(struct lp_fragment_shader_variant *lp)
       lp->jit_thread_data_ptr_type = LLVMPointerType(thread_data_type, 0);
    }
 
+   /*
+    * lp_linear_elem
+    *
+    * XXX: it can be instanced only once due to the use of opaque types, and
+    * the fact that screen->module is also a global.
+    */
+   {
+      LLVMTypeRef ret_type;
+      LLVMTypeRef arg_types[1];
+      LLVMTypeRef func_type;
+
+      ret_type = LLVMPointerType(LLVMVectorType(LLVMInt8TypeInContext(lc), 16), 0);
+
+      arg_types[0] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
+
+      /* lp_linear_func */
+      func_type = LLVMFunctionType(ret_type, arg_types, ARRAY_SIZE(arg_types), 0);
+
+      /*
+       * We actually define lp_linear_elem not as a structure but simply as a
+       * lp_linear_func pointer
+       */
+      linear_elem_type = LLVMPointerType(func_type, 0);
+   }
+
+   /* struct lp_jit_linear_context */
+   {
+      LLVMTypeRef linear_elem_ptr_type = LLVMPointerType(linear_elem_type, 0);
+      LLVMTypeRef elem_types[LP_JIT_LINEAR_CTX_COUNT];
+      LLVMTypeRef linear_context_type;
+
+
+      elem_types[LP_JIT_LINEAR_CTX_CONSTANTS] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
+      elem_types[LP_JIT_LINEAR_CTX_TEX] =
+            LLVMArrayType(linear_elem_ptr_type, LP_MAX_LINEAR_TEXTURES);
+      elem_types[LP_JIT_LINEAR_CTX_INPUTS] =
+            LLVMArrayType(linear_elem_ptr_type, LP_MAX_LINEAR_INPUTS);
+      elem_types[LP_JIT_LINEAR_CTX_COLOR0] = LLVMPointerType(LLVMInt8TypeInContext(lc), 0);
+      elem_types[LP_JIT_LINEAR_CTX_BLEND_COLOR] = LLVMInt32TypeInContext(lc);
+      elem_types[LP_JIT_LINEAR_CTX_ALPHA_REF] = LLVMInt8TypeInContext(lc);
+
+      linear_context_type = LLVMStructTypeInContext(lc, elem_types,
+                                                    ARRAY_SIZE(elem_types), 0);
+
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_linear_context, constants,
+                             gallivm->target, linear_context_type,
+                             LP_JIT_LINEAR_CTX_CONSTANTS);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_linear_context, tex,
+                             gallivm->target, linear_context_type,
+                             LP_JIT_LINEAR_CTX_TEX);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_linear_context, inputs,
+                             gallivm->target, linear_context_type,
+                             LP_JIT_LINEAR_CTX_INPUTS);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_linear_context, color0,
+                             gallivm->target, linear_context_type,
+                             LP_JIT_LINEAR_CTX_COLOR0);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_linear_context, blend_color,
+                             gallivm->target, linear_context_type,
+                             LP_JIT_LINEAR_CTX_BLEND_COLOR);
+      LP_CHECK_MEMBER_OFFSET(struct lp_jit_linear_context, alpha_ref_value,
+                             gallivm->target, linear_context_type,
+                             LP_JIT_LINEAR_CTX_ALPHA_REF);
+      LP_CHECK_STRUCT_SIZE(struct lp_jit_linear_context,
+                           gallivm->target, linear_context_type);
+
+      lp->jit_linear_context_ptr_type = LLVMPointerType(linear_context_type, 0);
+   }
+
    if (gallivm_debug & GALLIVM_DEBUG_IR) {
       char *str = LLVMPrintModuleToString(gallivm->module);
       fprintf(stderr, "%s", str);
diff --git a/src/gallium/drivers/llvmpipe/lp_jit.h b/src/gallium/drivers/llvmpipe/lp_jit.h
index 4f00d44c800..f9ca7706bda 100644
--- a/src/gallium/drivers/llvmpipe/lp_jit.h
+++ b/src/gallium/drivers/llvmpipe/lp_jit.h
@@ -324,6 +324,94 @@ typedef void
                     unsigned depth_sample_stride);
 
 
+#define LP_MAX_LINEAR_CONSTANTS 16
+#define LP_MAX_LINEAR_TEXTURES 2
+#define LP_MAX_LINEAR_INPUTS 8
+
+
+/**
+ * This structure is passed directly to the generated fragment shader.
+ *
+ * It contains the derived state.
+ *
+ * Changes here must be reflected in the lp_jit_linear_context_* macros and
+ * lp_jit_init_types function. Changes to the ordering should be avoided.
+ *
+ * Only use types with a clear size and padding here, in particular prefer the
+ * stdint.h types to the basic integer types.
+ */
+struct lp_jit_linear_context
+{
+   /**
+    * Constants in 8bit unorm values.
+    */
+   const uint8_t (*constants)[4];
+   struct lp_linear_elem *tex[LP_MAX_LINEAR_TEXTURES];
+   struct lp_linear_elem *inputs[LP_MAX_LINEAR_INPUTS];
+
+   uint8_t *color0;
+   uint32_t blend_color;
+
+   uint8_t alpha_ref_value;
+};
+
+
+/**
+ * These enum values must match the position of the fields in the
+ * lp_jit_linear_context struct above.
+ */
+enum {
+   LP_JIT_LINEAR_CTX_CONSTANTS = 0,
+   LP_JIT_LINEAR_CTX_TEX,
+   LP_JIT_LINEAR_CTX_INPUTS,
+   LP_JIT_LINEAR_CTX_COLOR0,
+   LP_JIT_LINEAR_CTX_BLEND_COLOR,
+   LP_JIT_LINEAR_CTX_ALPHA_REF,
+   LP_JIT_LINEAR_CTX_COUNT
+};
+
+
+#define lp_jit_linear_context_constants(_gallivm, _ptr) \
+   lp_build_struct_get(_gallivm, _ptr, LP_JIT_LINEAR_CTX_CONSTANTS, "constants")
+
+#define lp_jit_linear_context_tex(_gallivm, _ptr) \
+   lp_build_struct_get_ptr(_gallivm, _ptr, LP_JIT_LINEAR_CTX_TEX, "tex")
+
+#define lp_jit_linear_context_inputs(_gallivm, _ptr) \
+   lp_build_struct_get_ptr(_gallivm, _ptr, LP_JIT_LINEAR_CTX_INPUTS, "inputs")
+
+#define lp_jit_linear_context_color0(_gallivm, _ptr) \
+   lp_build_struct_get_ptr(_gallivm, _ptr, LP_JIT_LINEAR_CTX_COLOR0, "color0")
+
+#define lp_jit_linear_context_blend_color(_gallivm, _ptr) \
+   lp_build_struct_get_ptr(_gallivm, _ptr, LP_JIT_LINEAR_CTX_BLEND_COLOR, "blend_color")
+
+#define lp_jit_linear_context_alpha_ref(_gallivm, _ptr) \
+   lp_build_struct_get_ptr(_gallivm, _ptr, LP_JIT_LINEAR_CTX_ALPHA_REF, "alpha_ref_value")
+
+
+typedef const uint8_t *
+(*lp_jit_linear_llvm_func)(struct lp_jit_linear_context *context,
+                           uint32_t x,
+                           uint32_t y,
+                           uint32_t w);
+
+/* We're not really jitting this, but I need to get into the
+ * rast_state struct to call the function we actually are jitting.
+ */
+struct lp_rast_state;
+typedef boolean
+(*lp_jit_linear_func)(const struct lp_rast_state *state,
+                      uint32_t x,
+                      uint32_t y,
+                      uint32_t w,
+                      uint32_t h,
+                      const float (*a0)[4],
+                      const float (*dadx)[4],
+                      const float (*dady)[4],
+                      uint8_t *color,
+                      uint32_t color_stride);
+
 struct lp_jit_cs_thread_data
 {
    struct lp_build_format_cache *cache;
diff --git a/src/gallium/drivers/llvmpipe/lp_limits.h b/src/gallium/drivers/llvmpipe/lp_limits.h
index 1b8a37e2bb0..2c90a4fff1e 100644
--- a/src/gallium/drivers/llvmpipe/lp_limits.h
+++ b/src/gallium/drivers/llvmpipe/lp_limits.h
@@ -69,11 +69,6 @@
 #define LP_MAX_THREADS 16
 
 
-/**
- * Max bytes per scene.  This may be replaced by a runtime parameter.
- */
-#define LP_MAX_SCENE_SIZE (512 * 1024 * 1024)
-
 /**
  * Max number of shader variants (for all shaders combined,
  * per context) that will be kept around.
diff --git a/src/gallium/drivers/llvmpipe/lp_linear.c b/src/gallium/drivers/llvmpipe/lp_linear.c
new file mode 100644
index 00000000000..999e7781a9b
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_linear.c
@@ -0,0 +1,363 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "pipe/p_config.h"
+
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_pack_color.h"
+#include "util/u_rect.h"
+#include "util/u_sse.h"
+
+#include "lp_jit.h"
+#include "lp_rast.h"
+#include "lp_debug.h"
+#include "lp_state_fs.h"
+#include "lp_linear_priv.h"
+
+
+#if defined(PIPE_ARCH_SSE)
+
+
+/* For debugging (LP_DEBUG=linear), shade areas of run-time fallback
+ * purple.  Keep blending active so we can see more of what's going
+ * on.
+ */
+static boolean
+linear_fallback(const struct lp_rast_state *state,
+                unsigned x, unsigned y,
+                unsigned width, unsigned height,
+                uint8_t *color,
+                unsigned stride)
+{
+   unsigned col = 0x808000ff;
+   int i;
+
+   for (y = 0; y < height; y++) {
+      for (i = 0; i < 64; i++) {
+         *((uint32_t *)(color + y*stride) + x + i) = col;
+      }
+   }
+
+   return TRUE;
+}
+
+
+/* Run our configurable linear shader pipeline:
+ */
+static boolean
+lp_fs_linear_run(const struct lp_rast_state *state,
+                 unsigned x, unsigned y,
+                 unsigned width, unsigned height,
+                 const float (*a0)[4],
+                 const float (*dadx)[4],
+                 const float (*dady)[4],
+                 uint8_t *color,
+                 unsigned stride)
+{
+   const struct lp_fragment_shader_variant *variant = state->variant;
+   const struct lp_tgsi_info *info = &variant->shader->info;
+   struct lp_jit_linear_context jit;
+   lp_jit_linear_llvm_func jit_func = variant->jit_linear_llvm;
+
+   struct lp_linear_sampler samp[LP_MAX_LINEAR_TEXTURES];
+   struct lp_linear_interp interp[LP_MAX_LINEAR_INPUTS];
+   uint8_t constants[LP_MAX_LINEAR_CONSTANTS][4];
+
+   const float w0 = a0[0][3];
+   float oow = 1.0f/w0;
+
+   unsigned input_mask = variant->linear_input_mask;
+   int nr_consts = info->base.file_max[TGSI_FILE_CONSTANT]+1;
+   int nr_tex = info->num_texs;
+   int i, j;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   /* Require constant w in these rectangles:
+    */
+   if (dadx[0][3] != 0.0f ||
+       dady[0][3] != 0.0f) {
+      if (LP_DEBUG & DEBUG_LINEAR2)
+         debug_printf("  -- w not constant\n");
+      goto fail;
+   }
+
+   /* XXX: Per statechange:
+    */
+   for (i = 0; i < nr_consts; i++) {
+      for (j = 0; j < 4; j++) {
+         float val = state->jit_context.constants[0][i*4+j];
+         if (val < 0.0f || val > 1.0f) {
+            if (LP_DEBUG & DEBUG_LINEAR2)
+               debug_printf("  -- const[%d] out of range\n", i);
+            goto fail;
+         }
+         constants[i][j] = (uint8_t)(val * 255.0f);
+      }
+   }
+   jit.constants = (const uint8_t (*)[4])constants;
+
+   /* We assume BGRA ordering */
+   assert(variant->key.cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM ||
+          variant->key.cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM);
+
+   jit.blend_color =
+         state->jit_context.u8_blend_color[32] +
+         (state->jit_context.u8_blend_color[16] << 8) +
+         (state->jit_context.u8_blend_color[0] << 16) +
+         (state->jit_context.u8_blend_color[48] << 24);
+
+   jit.alpha_ref_value = float_to_ubyte(state->jit_context.alpha_ref_value);
+
+   /* XXX: Per primitive:
+    */
+   while (input_mask) {
+      int i = u_bit_scan(&input_mask);
+      unsigned usage_mask = info->base.input_usage_mask[i];
+      boolean perspective =
+            info->base.input_interpolate[i] == TGSI_INTERPOLATE_PERSPECTIVE ||
+            (info->base.input_interpolate[i] == TGSI_INTERPOLATE_COLOR &&
+             !variant->key.flatshade);
+
+      if (!lp_linear_init_interp(&interp[i],
+                                 x, y, width, height,
+                                 usage_mask,
+                                 perspective,
+                                 oow,
+                                 a0[i+1],
+                                 dadx[i+1],
+                                 dady[i+1])) {
+         if (LP_DEBUG & DEBUG_LINEAR2)
+            debug_printf("  -- init_interp(%d) failed\n", i);
+         goto fail;
+      }
+
+      jit.inputs[i] = &interp[i].base;
+   }
+
+
+   /* XXX: Per primitive: Initialize linear or nearest samplers:
+    */
+   for (i = 0; i < nr_tex; i++) {
+      const struct lp_tgsi_texture_info *tex_info = &info->tex[i];
+      unsigned unit = tex_info->sampler_unit;
+
+      /* XXX: some texture coordinates are linear!
+       */
+      //boolean perspective = (info->base.input_interpolate[i] ==
+      //                       TGSI_INTERPOLATE_PERSPECTIVE);
+
+      if (!lp_linear_init_sampler(&samp[i],
+                                  tex_info,
+                                  &variant->key.samplers[unit],
+                                  &state->jit_context.textures[unit],
+                                  x, y, width, height,
+                                  a0, dadx, dady)) {
+         if (LP_DEBUG & DEBUG_LINEAR2)
+            debug_printf("  -- init_sampler(%d) failed\n", i);
+         goto fail;
+      }
+
+      jit.tex[i] = &samp[i].base;
+   }
+
+   /* JIT function already does blending */
+   jit.color0 = color + x * 4 + y * stride;
+   for (y = 0; y < height; y++) {
+      jit_func(&jit, 0, 0, width);
+      jit.color0 += stride;
+   }
+
+   return TRUE;
+
+fail:
+   /* Visually distinguish this from other fallbacks:
+    */
+   if (LP_DEBUG & DEBUG_LINEAR) {
+      return linear_fallback(state, x, y, width, height, color, stride);
+   }
+
+   return FALSE;
+}
+
+
+static void
+check_linear_interp_mask_a(struct lp_fragment_shader_variant *variant)
+{
+   const struct lp_tgsi_info *info = &variant->shader->info;
+   struct lp_jit_linear_context jit;
+
+   struct lp_linear_sampler samp[LP_MAX_LINEAR_TEXTURES];
+   struct lp_linear_interp interp[LP_MAX_LINEAR_INPUTS];
+   uint8_t constants[LP_MAX_LINEAR_CONSTANTS][4];
+   PIPE_ALIGN_VAR(16) uint8_t color0[TILE_SIZE*4];
+
+   int nr_inputs = info->base.file_max[TGSI_FILE_INPUT]+1;
+   int nr_tex = info->num_texs;
+   int i;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   jit.constants = (const uint8_t (*)[4])constants;
+
+   for (i = 0; i < nr_tex; i++) {
+      lp_linear_init_noop_sampler(&samp[i]);
+      jit.tex[i] = &samp[i].base;
+   }
+
+   for (i = 0; i < nr_inputs; i++) {
+      lp_linear_init_noop_interp(&interp[i]);
+      jit.inputs[i] = &interp[i].base;
+   }
+
+   jit.color0 = color0;
+
+   (void)variant->jit_linear_llvm(&jit, 0, 0, 0);
+
+   /* Find out which interpolators were called, and store this as a
+    * mask:
+    */
+   for (i = 0; i < nr_inputs; i++)
+      variant->linear_input_mask |= (interp[i].row[0] << i);
+}
+
+
+/* Until the above is working, look at texture information and guess
+ * that any input used as a texture coordinate is not used for
+ * anything else.
+ */
+static void
+check_linear_interp_mask_b(struct lp_fragment_shader_variant *variant)
+{
+   const struct lp_tgsi_info *info = &variant->shader->info;
+   int nr_inputs = info->base.file_max[TGSI_FILE_INPUT]+1;
+   int nr_tex = info->num_texs;
+   unsigned tex_mask = 0;
+   int i;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   for (i = 0; i < nr_tex; i++) {
+      const struct lp_tgsi_texture_info *tex_info = &info->tex[i];
+      const struct lp_tgsi_channel_info *schan = &tex_info->coord[0];
+      const struct lp_tgsi_channel_info *tchan = &tex_info->coord[1];
+      tex_mask |= 1 << schan->u.index;
+      tex_mask |= 1 << tchan->u.index;
+   }
+
+   variant->linear_input_mask = ((1 << nr_inputs) - 1) & ~tex_mask;
+}
+
+
+void
+lp_linear_check_variant(struct lp_fragment_shader_variant *variant)
+{
+   const struct lp_fragment_shader_variant_key *key = &variant->key;
+   const struct lp_fragment_shader *shader = variant->shader;
+   const struct lp_tgsi_info *info = &shader->info;
+   int i;
+
+   if (info->base.file_max[TGSI_FILE_CONSTANT] >= LP_MAX_LINEAR_CONSTANTS ||
+       info->base.file_max[TGSI_FILE_INPUT] >= LP_MAX_LINEAR_INPUTS) {
+      if (LP_DEBUG & DEBUG_LINEAR)
+         debug_printf("  -- too many inputs/constants\n");
+      goto fail;
+   }
+
+   /* If we have a fastpath which implements the entire varient, use
+    * that.
+    */
+   if (lp_linear_check_fastpath(variant)) {
+      return;
+   }
+
+   /* Otherwise, can we build up a spanline-based linear path for this
+    * variant?
+    */
+
+   /* Check static sampler state.
+    */
+   for (i = 0; i < info->num_texs; i++) {
+      const struct lp_tgsi_texture_info *tex_info = &info->tex[i];
+      unsigned unit = tex_info->sampler_unit;
+
+      /* XXX: Relax this once setup premultiplies by oow:
+       */
+      if (info->base.input_interpolate[unit] != TGSI_INTERPOLATE_PERSPECTIVE) {
+         if (LP_DEBUG & DEBUG_LINEAR)
+            debug_printf(" -- samp[%d]: texcoord not perspective\n", i);
+         goto fail;
+      }
+
+      if (!lp_linear_check_sampler(&key->samplers[unit], tex_info)) {
+         if (LP_DEBUG & DEBUG_LINEAR)
+            debug_printf(" -- samp[%d]: check_sampler failed\n", i);
+         goto fail;
+      }
+   }
+
+   /* Check shader.  May not have been jitted.
+    */
+   if (variant->linear_function == NULL) {
+      if (LP_DEBUG & DEBUG_LINEAR)
+         debug_printf("  -- no linear shader\n");
+      goto fail;
+   }
+
+   /* Hook in the catchall shader runner:
+    */
+   variant->jit_linear = lp_fs_linear_run;
+
+   /* Figure out which inputs we don't need to interpolate (because
+    * they are only used as texture coordinates).  This is important
+    * as we can cope with texture coordinates which exceed 1.0, but
+    * cannot do so for regular inputs.
+    */
+   if (1)
+      check_linear_interp_mask_a(variant);
+   else
+      check_linear_interp_mask_b(variant);
+
+
+   if (0) {
+      lp_debug_fs_variant(variant);
+      debug_printf("linear input mask: 0x%x\n", variant->linear_input_mask);
+   }
+
+   return;
+
+fail:
+   if (LP_DEBUG & DEBUG_LINEAR) {
+      lp_debug_fs_variant(variant);
+      debug_printf("    ----> no linear path for this variant\n");
+   }
+}
+
+
+#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_linear_fastpath.c b/src/gallium/drivers/llvmpipe/lp_linear_fastpath.c
new file mode 100644
index 00000000000..be2802976a0
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_linear_fastpath.c
@@ -0,0 +1,237 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "pipe/p_config.h"
+
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_pack_color.h"
+#include "util/u_surface.h"
+#include "util/u_sse.h"
+
+#include "lp_jit.h"
+#include "lp_rast.h"
+#include "lp_debug.h"
+#include "lp_state_fs.h"
+#include "lp_linear_priv.h"
+
+
+#if defined(PIPE_ARCH_SSE)
+
+
+/* This file contains various special-case fastpaths which implement
+ * the entire linear pipeline in a single funciton.
+ *
+ * These include simple blits and some debug code.
+ *
+ * These functions fully implement the linear path and do not need to
+ * be combined with blending, interpolation or sampling routines.
+ */
+
+/* Linear shader which implements the BLIT_RGBA shader with the
+ * additional constraints imposed by lp_setup_is_blit().
+ */
+static boolean
+lp_linear_blit_rgba_blit(const struct lp_rast_state *state,
+               unsigned x, unsigned y,
+               unsigned width, unsigned height,
+               const float (*a0)[4],
+               const float (*dadx)[4],
+               const float (*dady)[4],
+               uint8_t *color,
+               unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   const struct lp_jit_texture *texture = &context->textures[0];
+   const uint8_t *src;
+   unsigned src_stride;
+   int src_x, src_y;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   /* Require w==1.0:
+    */
+   if (a0[0][3] != 1.0 ||
+       dadx[0][3] != 0.0 ||
+       dady[0][3] != 0.0)
+      return FALSE;
+
+   src_x = x + util_iround(a0[1][0]*texture->width - 0.5f);
+   src_y = y + util_iround(a0[1][1]*texture->height - 0.5f);
+
+   src = texture->base;
+   src_stride = texture->row_stride[0];
+
+   /* Fall back to blit_rgba() if clamping required:
+    */
+   if (src_x < 0 ||
+       src_y < 0 ||
+       src_x + width > texture->width ||
+       src_y + height > texture->height)
+      return FALSE;
+
+   util_copy_rect(color, PIPE_FORMAT_B8G8R8A8_UNORM, stride,
+                  x, y,
+                  width, height,
+                  src, src_stride,
+                  src_x, src_y);
+
+   return TRUE;
+}
+
+
+/* Linear shader which implements the BLIT_RGB1 shader, with the
+ * additional constraints imposed by lp_setup_is_blit().
+ */
+static boolean
+lp_linear_blit_rgb1_blit(const struct lp_rast_state *state,
+               unsigned x, unsigned y,
+               unsigned width, unsigned height,
+               const float (*a0)[4],
+               const float (*dadx)[4],
+               const float (*dady)[4],
+               uint8_t *color,
+               unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   const struct lp_jit_texture *texture = &context->textures[0];
+   const uint8_t *src;
+   unsigned src_stride;
+   int src_x, src_y;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   /* Require w==1.0:
+    */
+   if (a0[0][3] != 1.0 ||
+       dadx[0][3] != 0.0 ||
+       dady[0][3] != 0.0)
+      return FALSE;
+
+   color += x * 4 + y * stride;
+
+   src_x = x + util_iround(a0[1][0]*texture->width - 0.5f);
+   src_y = y + util_iround(a0[1][1]*texture->height - 0.5f);
+
+   src = texture->base;
+   src_stride = texture->row_stride[0];
+   src += src_x * 4;
+   src += src_y * src_stride;
+
+   if (src_x < 0 ||
+       src_y < 0 ||
+       src_x + width > texture->width ||
+       src_y + height > texture->height)
+      return FALSE;
+
+   for (y = 0; y < height; y++) {
+      const uint32_t *src_row = (const uint32_t *)src;
+      uint32_t *dst_row = (uint32_t *)color;
+
+      for (x = 0; x < width; x++) {
+         *dst_row++ = *src_row++ | 0xff000000;
+      }
+
+      color += stride;
+      src += src_stride;
+   }
+
+   return TRUE;
+}
+
+/* Linear shader which always emits purple.  Used for debugging.
+ */
+static boolean
+lp_linear_purple(const struct lp_rast_state *state,
+              unsigned x, unsigned y,
+              unsigned width, unsigned height,
+              const float (*a0)[4],
+              const float (*dadx)[4],
+              const float (*dady)[4],
+              uint8_t *color,
+              unsigned stride)
+{
+   union util_color uc;
+
+   util_pack_color_ub(0xff, 0, 0xff, 0xff,
+                      PIPE_FORMAT_B8G8R8A8_UNORM, &uc);
+
+   util_fill_rect(color,
+                  PIPE_FORMAT_B8G8R8A8_UNORM,
+                  stride,
+                  x,
+                  y,
+                  width,
+                  height,
+                  &uc);
+
+   return TRUE;
+}
+
+/* Examine the fragment shader varient and determine whether we can
+ * substitute a fastpath linear shader implementation.
+ */
+boolean
+lp_linear_check_fastpath(struct lp_fragment_shader_variant *variant)
+{
+   enum pipe_format tex_format = variant->key.samplers[0].texture_state.format;
+
+   if (variant->shader->kind == LP_FS_KIND_BLIT_RGBA &&
+       tex_format == PIPE_FORMAT_B8G8R8A8_UNORM &&
+       is_nearest_clamp_sampler(&variant->key.samplers[0]) &&
+       variant->opaque) {
+      variant->jit_linear_blit             = lp_linear_blit_rgba_blit;
+   }
+
+   if (variant->shader->kind == LP_FS_KIND_BLIT_RGB1 &&
+       variant->opaque &&
+       (tex_format == PIPE_FORMAT_B8G8R8A8_UNORM ||
+        tex_format == PIPE_FORMAT_B8G8R8X8_UNORM) &&
+       is_nearest_clamp_sampler(&variant->key.samplers[0])) {
+      variant->jit_linear_blit             = lp_linear_blit_rgb1_blit;
+   }
+
+   if (0) {
+      variant->jit_linear                   = lp_linear_purple;
+   }
+
+
+   /* Stop now if jit_linear has been initialized.  Otherwise keep
+    * searching - even if jit_linear_blit has been instantiated.
+    */
+   return variant->jit_linear != NULL;
+}
+#else
+void
+boolean
+lp_linear_check_fastpath(struct lp_fragment_shader_variant *variant)
+{
+   return FALSE;
+}
+#endif
+
diff --git a/src/gallium/drivers/llvmpipe/lp_linear_interp.c b/src/gallium/drivers/llvmpipe/lp_linear_interp.c
new file mode 100644
index 00000000000..1006e40518c
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_linear_interp.c
@@ -0,0 +1,243 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "pipe/p_config.h"
+
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_pack_color.h"
+#include "util/u_rect.h"
+#include "util/u_sse.h"
+
+#include "lp_jit.h"
+#include "lp_rast.h"
+#include "lp_debug.h"
+#include "lp_state_fs.h"
+#include "lp_linear_priv.h"
+
+
+#if defined(PIPE_ARCH_SSE)
+
+#define FIXED15_ONE 0x7fff
+
+/* Translate floating point value to 1.15 unsigned fixed-point.
+ */
+static inline ushort
+float_to_ufixed_1_15(float f)
+{
+   return CLAMP((unsigned)(f * (float)FIXED15_ONE), 0, FIXED15_ONE);
+}
+
+
+/* Translate floating point value to 1.15 signed fixed-point.
+ */
+static inline int16_t
+float_to_sfixed_1_15(float f)
+{
+   return CLAMP((signed)(f * (float)FIXED15_ONE), -FIXED15_ONE, FIXED15_ONE);
+}
+
+
+/* Interpolate in 1.15 space, but produce a packed row of 0.8 values.
+ */
+static const uint32_t *
+interp_0_8(struct lp_linear_elem *elem)
+{
+   struct lp_linear_interp *interp = (struct lp_linear_interp *)elem;
+   uint32_t *row = interp->row;
+   __m128i a0 = interp->a0;
+   __m128i dadx = interp->dadx;
+   int width = (interp->width + 3) & ~3;
+   int i;
+
+   for (i = 0; i < width; i += 4) {
+      __m128i l, h;
+
+      l = _mm_srai_epi16(a0, 7);
+      a0 = _mm_add_epi16(a0, dadx);
+
+      h = _mm_srai_epi16(a0, 7);
+      a0 = _mm_add_epi16(a0, dadx);
+
+      *(__m128i *)&row[i] =  _mm_packus_epi16(l, h);
+   }
+
+   interp->a0 = _mm_add_epi16(interp->a0, interp->dady);
+   return interp->row;
+}
+
+static const uint32_t *
+interp_noop(struct lp_linear_elem *elem)
+{
+   struct lp_linear_interp *interp = (struct lp_linear_interp *)elem;
+   return interp->row;
+}
+
+
+static const uint32_t *
+interp_check(struct lp_linear_elem *elem)
+{
+   struct lp_linear_interp *interp = (struct lp_linear_interp *)elem;
+   interp->row[0] = 1;
+   return interp->row;
+}
+
+/* Not quite a noop - we use row[0] to track whether this gets called
+ * or not, so we can optimize which interpolants we care about.
+ */
+void
+lp_linear_init_noop_interp(struct lp_linear_interp *interp)
+{
+   interp->row[0] = 0;
+   interp->base.fetch = interp_check;
+}
+
+boolean
+lp_linear_init_interp(struct lp_linear_interp *interp,
+                      int x, int y, int width, int height,
+                      unsigned usage_mask,
+                      boolean perspective,
+                      float oow,
+                      const float *a0,
+                      const float *dadx,
+                      const float *dady)
+{
+   float s0[4];
+   float dsdx[4];
+   float dsdy[4];
+   int16_t s0_fp[8];
+   int16_t dsdx_fp[4];
+   int16_t dsdy_fp[4];
+   int j;
+
+   /* Zero coefficients to avoid using uninitialised values */
+   memset(s0, 0, sizeof(s0));
+   memset(dsdx, 0, sizeof(dsdx));
+   memset(dsdy, 0, sizeof(dsdy));
+   memset(s0_fp, 0, sizeof(s0_fp));
+   memset(dsdx_fp, 0, sizeof(dsdx_fp));
+   memset(dsdy_fp, 0, sizeof(dsdy_fp));
+
+   if (perspective) {
+      for (j = 0; j < 4; j++) {
+         if (usage_mask & (1<<j)) {
+            s0[j]   =   a0[j] * oow;
+            dsdx[j] = dadx[j] * oow;
+            dsdy[j] = dady[j] * oow;
+         }
+      }
+   } else {
+      for (j = 0; j < 4; j++) {
+         if (usage_mask & (1<<j)) {
+            s0[j]   =   a0[j];
+            dsdx[j] = dadx[j];
+            dsdy[j] = dady[j];
+         }
+      }
+   }
+
+   s0[0] += x * dsdx[0] + y * dsdy[0];
+   s0[1] += x * dsdx[1] + y * dsdy[1];
+   s0[2] += x * dsdx[2] + y * dsdy[2];
+   s0[3] += x * dsdx[3] + y * dsdy[3];
+
+   /* XXX: lift all of this into the rectangle setup code.
+    *
+    * For rectangles with linear shaders, at setup time:
+    *    - if w is constant (else mark as non-fastpath)
+    *        - premultiply perspective interpolants by w
+    *        - set w = 1 in position
+    *   - check all interpolants for min/max 0..1 (else mark as
+    *          non-fastpath)
+    */
+   for (j = 0; j < 4; j++) {
+      if (usage_mask & (1<<j)) {
+         float a = s0[j];
+         float b = s0[j] + (width  - 1) * dsdx[j];
+         float c = s0[j] + (height - 1) * dsdy[j];
+         float d = s0[j] + (height - 1) * dsdy[j] + (width - 1) * dsdx[j];
+
+         if (MIN4(a,b,c,d) < 0.0)
+            FAIL("min < 0.0");
+
+         if (MAX4(a,b,c,d) > 1.0)
+            FAIL("max > 1.0");
+
+         dsdx_fp[j]   = float_to_sfixed_1_15(dsdx[j]);
+         dsdy_fp[j]   = float_to_sfixed_1_15(dsdy[j]);
+
+         s0_fp[j]     = float_to_ufixed_1_15(s0[j]);
+         s0_fp[j + 4] = s0_fp[j] + dsdx_fp[j];
+
+         dsdx_fp[j] *= 2;
+      }
+   }
+
+   interp->width = align(width, 4);
+
+   interp->a0    = _mm_setr_epi16(s0_fp[2], s0_fp[1], s0_fp[0], s0_fp[3],
+                                  s0_fp[6], s0_fp[5], s0_fp[4], s0_fp[7]);
+
+   interp->dadx  = _mm_setr_epi16(dsdx_fp[2], dsdx_fp[1], dsdx_fp[0], dsdx_fp[3],
+                                  dsdx_fp[2], dsdx_fp[1], dsdx_fp[0], dsdx_fp[3]);
+
+   interp->dady  = _mm_setr_epi16(dsdy_fp[2], dsdy_fp[1], dsdy_fp[0], dsdy_fp[3],
+                                  dsdy_fp[2], dsdy_fp[1], dsdy_fp[0], dsdy_fp[3]);
+
+   /* If the value is y-invariant, eagerly calculate it here and then
+    * always return the precalculated value.
+    */
+   if (dsdy[0] == 0 &&
+       dsdy[1] == 0 &&
+       dsdy[2] == 0 &&
+       dsdy[3] == 0)
+   {
+      interp_0_8(&interp->base);
+      interp->base.fetch = interp_noop;
+   }
+   else {
+      interp->base.fetch = interp_0_8;
+   }
+
+   return TRUE;
+}
+
+#else
+boolean
+lp_linear_init_interp(struct lp_linear_interp *interp,
+                      int x, int y, int width, int height,
+                      unsigned usage_mask,
+                      boolean perspective,
+                      float oow,
+                      const float *a0,
+                      const float *dadx,
+                      const float *dady)
+{
+   return FALSE;
+}
+#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_linear_priv.h b/src/gallium/drivers/llvmpipe/lp_linear_priv.h
new file mode 100644
index 00000000000..756e2398e57
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_linear_priv.h
@@ -0,0 +1,169 @@
+
+#ifndef LP_LINEAR_PRIV_H
+#define LP_LINEAR_PRIV_H
+
+struct lp_linear_elem;
+
+typedef const uint32_t *(*lp_linear_func)(struct lp_linear_elem *base);
+
+
+struct lp_linear_elem {
+   lp_linear_func fetch;
+};
+
+
+/* "Linear" refers to the fact we're on the linear (non-swizzled)
+ * rasterization path.  Filtering mode may be either linear or
+ * nearest.
+ */
+struct lp_linear_sampler {
+   struct lp_linear_elem base;
+
+   const struct lp_jit_texture *texture;
+   int s;                       /* 16.16, biased by .5 */
+   int t;                       /* 16.16, biased by .5 */
+   int dsdx;                    /* 16.16 */
+   int dsdy;                    /* 16.16 */
+   int dtdx;                    /* 16.16 */
+   int dtdy;                    /* 16.16 */
+   int width;
+   boolean axis_aligned;
+
+   PIPE_ALIGN_VAR(16) uint32_t row[64];
+   PIPE_ALIGN_VAR(16) uint32_t stretched_row[2][64];
+
+   /**
+    * y coordinate of the rows stored in the stretched_row.
+    *
+    * Negative number means no stretched row is cached.
+    */
+   int stretched_row_y[2];
+
+   /**
+    * The index of stretched_row to receive the next stretched row.
+    */
+   int stretched_row_index;
+};
+
+/* "Linear" refers to the fact we're on the linear (non-swizzled)
+ * rasterization path.  Interpolation mode may be either constant,
+ * linear or perspective.
+ */
+struct lp_linear_interp {
+   struct lp_linear_elem base;
+
+   __m128i a0;
+   __m128i dadx;
+   __m128i dady;
+
+   int width;                   /* rounded up to multiple of 4 */
+
+   PIPE_ALIGN_VAR(16) uint32_t row[64];
+};
+
+
+/* Check for a sampler variant which matches our fetch_row
+ * implementation - normalized texcoords, single mipmap with
+ * nearest filtering.
+ */
+static inline boolean
+is_nearest_sampler(const struct lp_sampler_static_state *sampler)
+{
+   return
+      sampler->texture_state.target == PIPE_TEXTURE_2D &&
+      sampler->sampler_state.min_img_filter == PIPE_TEX_FILTER_NEAREST &&
+      sampler->sampler_state.mag_img_filter == PIPE_TEX_FILTER_NEAREST &&
+      (sampler->texture_state.level_zero_only ||
+       sampler->sampler_state.min_mip_filter == PIPE_TEX_MIPFILTER_NONE) &&
+      sampler->sampler_state.compare_mode == 0 &&
+      sampler->sampler_state.normalized_coords == 1;
+}
+
+
+/* Check for a sampler variant which matches our fetch_row
+ * implementation - normalized texcoords, single mipmap with
+ * linear filtering.
+ */
+static inline boolean
+is_linear_sampler(const struct lp_sampler_static_state *sampler)
+{
+   return
+      sampler->texture_state.target == PIPE_TEXTURE_2D &&
+      sampler->sampler_state.min_img_filter == PIPE_TEX_FILTER_LINEAR &&
+      sampler->sampler_state.mag_img_filter == PIPE_TEX_FILTER_LINEAR &&
+      (sampler->texture_state.level_zero_only ||
+       sampler->sampler_state.min_mip_filter == PIPE_TEX_MIPFILTER_NONE) &&
+      sampler->sampler_state.compare_mode == 0 &&
+      sampler->sampler_state.normalized_coords == 1;
+}
+
+
+/* Check for a sampler variant which matches is_nearest_sampler
+ * but has the additional constraints of using clamp wrapping
+ */
+static inline boolean
+is_nearest_clamp_sampler(const struct lp_sampler_static_state *sampler)
+{
+   return
+      is_nearest_sampler(sampler) &&
+      sampler->sampler_state.wrap_s == PIPE_TEX_WRAP_CLAMP_TO_EDGE &&
+      sampler->sampler_state.wrap_t == PIPE_TEX_WRAP_CLAMP_TO_EDGE;
+}
+
+
+/* Check for a sampler variant which matches is_linear_sampler
+ * but has the additional constraints of using clamp wrapping
+ */
+static inline boolean
+is_linear_clamp_sampler(const struct lp_sampler_static_state *sampler)
+{
+   return
+      is_linear_sampler(sampler) &&
+      sampler->sampler_state.wrap_s == PIPE_TEX_WRAP_CLAMP_TO_EDGE &&
+      sampler->sampler_state.wrap_t == PIPE_TEX_WRAP_CLAMP_TO_EDGE;
+}
+
+
+boolean
+lp_linear_init_interp(struct lp_linear_interp *interp,
+                      int x, int y, int width, int height,
+                      unsigned usage_mask,
+                      boolean perspective,
+                      float oow,
+                      const float *a0,
+                      const float *dadx,
+                      const float *dady);
+
+boolean
+lp_linear_init_sampler(struct lp_linear_sampler *samp,
+                       const struct lp_tgsi_texture_info *info,
+                       const struct lp_sampler_static_state *sampler_state,
+                       const struct lp_jit_texture *texture,
+                       int x0, int y0, int width, int height,
+                       const float (*a0)[4],
+                       const float (*dadx)[4],
+                       const float (*dady)[4]);
+
+
+boolean
+lp_linear_check_fastpath(struct lp_fragment_shader_variant *variant);
+
+boolean
+lp_linear_check_sampler(const struct lp_sampler_static_state *sampler,
+                        const struct lp_tgsi_texture_info *tex);
+
+
+void
+lp_linear_init_noop_interp(struct lp_linear_interp *interp);
+
+void
+lp_linear_init_noop_sampler(struct lp_linear_sampler *samp);
+
+
+#define FAIL(s) do {                                    \
+      if (LP_DEBUG & DEBUG_LINEAR)                      \
+         debug_printf("%s: %s\n", __FUNCTION__, s);     \
+      return FALSE;                                     \
+} while (0)
+
+#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_linear_sampler.c b/src/gallium/drivers/llvmpipe/lp_linear_sampler.c
new file mode 100644
index 00000000000..4f11703bfc2
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_linear_sampler.c
@@ -0,0 +1,1009 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "pipe/p_config.h"
+
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_pack_color.h"
+#include "util/u_rect.h"
+#include "util/u_sse.h"
+
+#include "lp_jit.h"
+#include "lp_debug.h"
+#include "lp_state_fs.h"
+#include "lp_linear_priv.h"
+
+#if defined(PIPE_ARCH_SSE)
+
+#define FIXED16_SHIFT  16
+#define FIXED16_ONE    (1<<16)
+#define FIXED16_HALF   (1<<15)
+
+/*
+ * Color tolerance.  Allow 1 bit of error in 8 bit unorm colors.
+ */
+#define FIXED16_TOL (FIXED16_ONE >> 7)
+
+/*
+ * Tolerance for texture coordinate derivatives when doing linear filtering.
+ *
+ * (Note that extra care needs to be taken when doing linear filtering as
+ * coordinates may snap up to neighbour texels inside the tile).
+ */
+#define FIXED16_TOL_DERIV (FIXED16_TOL / TILE_SIZE)
+
+static inline int
+float_to_fixed16(float f)
+{
+   return f * (float)FIXED16_ONE;
+}
+
+static inline int
+fixed16_frac(int x)
+{
+   return x & (FIXED16_ONE - 1);
+}
+
+static inline int
+fixed16_approx(int x, int y, int tol)
+{
+   return y - tol <= x && x <= y + tol;
+}
+
+
+/*
+ * Unstretched blit of a bgra texture.
+ */
+static const uint32_t *
+fetch_bgra_memcpy(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint32_t *src_row =
+      (const uint32_t *)((const uint8_t *)texture->base +
+                         (samp->t >> FIXED16_SHIFT) * texture->row_stride[0]);
+   const int s     = samp->s;
+   const int width = samp->width;
+   const uint32_t *row;
+
+   src_row = &src_row[s >> FIXED16_SHIFT];
+
+   if (((uintptr_t)src_row & 0xf) == 0) {
+      /* The source texels are already aligned. Return them */
+      row = src_row;
+   } else {
+      memcpy(samp->row, src_row, width * sizeof *row);
+      row = samp->row;
+   }
+
+   samp->t += samp->dtdy;
+   return row;
+}
+
+
+/*
+ * Unstretched blit of a bgrx texture.
+ */
+static const uint32_t *
+fetch_bgrx_memcpy(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint32_t *src_row =
+      (const uint32_t *)((const uint8_t *)texture->base +
+                         (samp->t >> FIXED16_SHIFT) * texture->row_stride[0]);
+   const int s     = samp->s;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int i;
+
+   src_row = &src_row[s >> FIXED16_SHIFT];
+
+   for (i = 0; i < width; i++) {
+      row[i] = src_row[i] | 0xff000000;
+   }
+
+   samp->t += samp->dtdy;
+   return row;
+}
+
+
+/*
+ * Perform nearest filtered lookup of a row of texels.  Texture lookup
+ * is assumed to be axis aligned but with arbitrary scaling.
+ *
+ * Texture coordinate interpolation is performed in 16.16 fixed point,
+ * not to be confused with the 1.15 format used by the interpolants.
+ *
+ * After 64 pixels (ie. in the next tile), the starting point will be
+ * recalculated with floating point arithmetic.
+ */
+static const uint32_t *
+fetch_bgra_axis_aligned(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint32_t *src_row =
+      (const uint32_t *)((const uint8_t *)texture->base +
+                         (samp->t >> FIXED16_SHIFT) * texture->row_stride[0]);
+   const int dsdx  = samp->dsdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      row[i] = src_row[s>>FIXED16_SHIFT];
+      s += dsdx;
+   }
+
+   samp->t += samp->dtdy;
+   return row;
+}
+
+static const uint32_t *
+fetch_bgrx_axis_aligned(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint32_t *src_row =
+      (const uint32_t *)((const uint8_t *)texture->base +
+                         (samp->t >> FIXED16_SHIFT) * texture->row_stride[0]);
+   const int dsdx  = samp->dsdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      row[i] = src_row[s>>FIXED16_SHIFT] | 0xff000000;
+      s += dsdx;
+   }
+
+   samp->t += samp->dtdy;
+   return row;
+}
+
+/* Non-axis aligned, but no clamping or wrapping required
+ */
+static const uint32_t *
+fetch_bgra(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint8_t *src = texture->base;
+   const int stride = texture->row_stride[0];
+   const int dsdx  = samp->dsdx;
+   const int dtdx  = samp->dtdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int t = samp->t;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      const uint8_t *texel = (src +
+                              (t>>FIXED16_SHIFT) * stride +
+                              (s>>FIXED16_SHIFT) * 4);
+
+      row[i] = *(const uint32_t *)texel;
+
+      s += dsdx;
+      t += dtdx;
+   }
+
+   samp->s += samp->dsdy;
+   samp->t += samp->dtdy;
+   return row;
+}
+
+
+static const uint32_t *
+fetch_bgrx(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint8_t *src = texture->base;
+   const int stride = texture->row_stride[0];
+   const int dsdx  = samp->dsdx;
+   const int dtdx  = samp->dtdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int t = samp->t;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      const uint8_t *texel = (src +
+                              (t>>FIXED16_SHIFT) * stride +
+                              (s>>FIXED16_SHIFT) * 4);
+
+      row[i] = (*(const uint32_t *)texel) | 0xff000000;
+
+      s += dsdx;
+      t += dtdx;
+   }
+
+   samp->s += samp->dsdy;
+   samp->t += samp->dtdy;
+   return row;
+}
+
+/* Non-axis aligned, clamped.
+ */
+static const uint32_t *
+fetch_bgra_clamp(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint8_t *src   = texture->base;
+   const int stride     = texture->row_stride[0];
+   const int tex_height = texture->height - 1;
+   const int tex_width  = texture->width - 1;
+   const int dsdx  = samp->dsdx;
+   const int dtdx  = samp->dtdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int t = samp->t;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      int ct = CLAMP(t>>FIXED16_SHIFT, 0, tex_height);
+      int cs = CLAMP(s>>FIXED16_SHIFT, 0, tex_width);
+
+      const uint8_t *texel = (src +
+                              ct * stride +
+                              cs * 4);
+
+      row[i] = *(const uint32_t *)texel;
+
+      s += dsdx;
+      t += dtdx;
+   }
+
+   samp->s += samp->dsdy;
+   samp->t += samp->dtdy;
+   return row;
+}
+
+static const uint32_t *
+fetch_bgrx_clamp(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint8_t *src   = texture->base;
+   const int stride     = texture->row_stride[0];
+   const int tex_height = texture->height - 1;
+   const int tex_width  = texture->width - 1;
+   const int dsdx  = samp->dsdx;
+   const int dtdx  = samp->dtdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int t = samp->t;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      int ct = CLAMP(t>>FIXED16_SHIFT, 0, tex_height);
+      int cs = CLAMP(s>>FIXED16_SHIFT, 0, tex_width);
+
+      const uint8_t *texel = (src +
+                              ct * stride +
+                              cs * 4);
+
+      row[i] = (*(const uint32_t *)texel) | 0xff000000;
+
+      s += dsdx;
+      t += dtdx;
+   }
+
+   samp->s += samp->dsdy;
+   samp->t += samp->dtdy;
+   return row;
+}
+
+/**
+ * Fetch and stretch one row.
+ */
+static inline const uint32_t *
+fetch_and_stretch_bgra_row(struct lp_linear_sampler *samp,
+                           int y)
+{
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint32_t *data = (const uint32_t *)texture->base;
+   const int stride = texture->row_stride[0] / sizeof(uint32_t);
+   const uint32_t * restrict src_row;
+   uint32_t * restrict dst_row;
+   const int width = samp->width;
+
+   /*
+    * Search the stretched row cache first.
+    */
+
+   if (y == samp->stretched_row_y[0]) {
+      samp->stretched_row_index = 1;
+      return samp->stretched_row[0];
+   }
+
+   if (y == samp->stretched_row_y[1]) {
+      samp->stretched_row_index = 0;
+      return samp->stretched_row[1];
+   }
+
+   /*
+    * Replace one entry.
+    */
+
+   src_row = data + y * stride;
+
+   dst_row = samp->stretched_row[samp->stretched_row_index];
+
+   if (fixed16_frac(samp->s) == 0 &&
+       samp->dsdx == FIXED16_ONE) { // TODO: could be relaxed
+      /*
+       * 1:1 blit on the x direction.
+       */
+
+      unsigned i;
+
+      src_row += samp->s >> FIXED16_SHIFT;
+
+      if (((uintptr_t)src_row & 0xf) == 0) {
+         /* The source texture is already aligned. Return it */
+         return src_row;
+      }
+
+      /* Copy the source texture */
+      for (i = 0; i < width; i += 4) {
+         __m128i src = _mm_loadu_si128((const __m128i *)&src_row[i]);
+         *(__m128i *)&dst_row[i] = src;
+      }
+   }
+   else {
+      util_sse2_stretch_row_8unorm((__m128i *)dst_row,
+                                   align(width, 4),
+                                   src_row, samp->s, samp->dsdx);
+   }
+
+   samp->stretched_row_y[samp->stretched_row_index] = y;
+   samp->stretched_row_index ^= 1;
+
+   return dst_row;
+}
+
+/* Maximise only as we fetch unscaled pixels linearly into a size-64
+ * temporary.  For minimise, we will want to either have a bigger
+ * temporary or fetch sparsely.
+ */
+static const uint32_t *
+fetch_bgra_axis_aligned_linear(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const int width = samp->width;
+   const uint32_t * restrict src_row0;
+   const uint32_t * restrict src_row1;
+   uint32_t * restrict row = samp->row;
+   int y = samp->t >> FIXED16_SHIFT;
+   int w = (samp->t >> 8) & 0xff;
+   int i;
+   __m128i wt;
+
+   samp->t += samp->dtdy;
+
+   src_row0 = fetch_and_stretch_bgra_row(samp, y);
+
+   if (w == 0) {
+      return src_row0;
+   }
+
+   src_row1 = fetch_and_stretch_bgra_row(samp, y + 1);
+
+   wt = _mm_set1_epi16(w);
+
+   /* Combine the two rows using a constant weight.
+    */
+   for (i = 0; i < width; i += 4) {
+      __m128i srca = _mm_load_si128((const __m128i *)&src_row0[i]);
+      __m128i srcb = _mm_load_si128((const __m128i *)&src_row1[i]);
+
+      *(__m128i *)&row[i] = util_sse2_lerp_epi8_fixed88(srca, srcb, &wt, &wt);
+   }
+
+   return row;
+}
+
+/* Non-axis-aligned version.  Don't try to take advantage of
+ * maximize.
+ */
+static const uint32_t *
+fetch_bgra_linear(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const int stride     = texture->row_stride[0] / sizeof(uint32_t);
+   const uint32_t *data  = (const uint32_t *)texture->base;
+   const int dsdx  = samp->dsdx;
+   const int dtdx  = samp->dtdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int t = samp->t;
+   int i, j;
+
+   for (i = 0; i < width; i += 4) {
+      union m128i si0, si1, si2, si3, ws, wt;
+      __m128i si02, si13;
+
+      for (j = 0; j < 4; j++) {
+         const uint32_t *src = data + (t >> 16) * stride + (s>>16);
+
+         si0.ui[j] = src[0];
+         si1.ui[j] = src[1];
+         si2.ui[j] = src[stride + 0];
+         si3.ui[j] = src[stride + 1];
+
+         ws.ui[j] = (s>>8) & 0xff;
+         wt.ui[j] = (t>>8) & 0xff;
+
+         s += dsdx;
+         t += dtdx;
+      }
+
+      ws.m = _mm_or_si128(ws.m, _mm_slli_epi32(ws.m, 16));
+      ws.m = _mm_or_si128(ws.m, _mm_slli_epi32(ws.m, 8));
+
+      wt.m = _mm_or_si128(wt.m, _mm_slli_epi32(wt.m, 16));
+      wt.m = _mm_or_si128(wt.m, _mm_slli_epi32(wt.m, 8));
+
+      si02 = util_sse2_lerp_epi8_fixed08(si0.m, si2.m, wt.m);
+      si13 = util_sse2_lerp_epi8_fixed08(si1.m, si3.m, wt.m);
+
+      *(__m128i *)&row[i] = util_sse2_lerp_epi8_fixed08(si02, si13, ws.m);
+   }
+
+   samp->s += samp->dsdy;
+   samp->t += samp->dtdy;
+   return row;
+}
+
+
+/* Clamped, non-axis-aligned version.  Don't try to take advantage of
+ * maximize.
+ */
+static const uint32_t *
+fetch_bgra_clamp_linear(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const struct lp_jit_texture *texture = samp->texture;
+   const uint32_t *data  = (const uint32_t *)texture->base;
+   const int stride     = texture->row_stride[0] / sizeof(uint32_t);
+   const int tex_height = texture->height - 1;
+   const int tex_width  = texture->width - 1;
+   const int dsdx  = samp->dsdx;
+   const int dtdx  = samp->dtdx;
+   const int width = samp->width;
+   uint32_t *row   = samp->row;
+   int s = samp->s;
+   int t = samp->t;
+   int i, j;
+   /* width, height, stride (in pixels) must be smaller than 32768 */
+   __m128i dsdx4, dtdx4, s4, t4, stride4, w4, h4, zero, one;
+   s4 = _mm_set1_epi32(s);
+   t4 = _mm_set1_epi32(t);
+   s4 = _mm_add_epi32(s4, _mm_set_epi32(3*dsdx, 2*dsdx, dsdx, 0));
+   t4 =  _mm_add_epi32(t4, _mm_set_epi32(3*dtdx, 2*dtdx, dtdx, 0));
+   dsdx4 = _mm_set1_epi32(4*dsdx);
+   dtdx4 = _mm_set1_epi32(4*dtdx);
+   stride4 = _mm_set1_epi32(stride);
+   w4 = _mm_set1_epi32(tex_width);
+   h4 = _mm_set1_epi32(tex_height);
+   zero = _mm_setzero_si128();
+   one = _mm_set1_epi32(1);
+
+   for (i = 0; i < width; i += 4) {
+      union m128i addr[4];
+      __m128i ws, wt, wsl, wsh, wtl, wth;
+      __m128i s4s, t4s, cs0, cs1, ct0, ct1, tmp, si[4];
+
+      s4s = _mm_srli_epi32(s4, 16);
+      t4s = _mm_srli_epi32(t4, 16);
+      cs0 = _mm_min_epi16(_mm_max_epi16(s4s, zero), w4);
+      cs1 = _mm_add_epi16(s4s, one);
+      cs1 = _mm_min_epi16(_mm_max_epi16(cs1, zero), w4);
+      ct0 = _mm_min_epi16(_mm_max_epi16(t4s, zero), h4);
+      ct1 = _mm_add_epi16(t4s, one);
+      ct1 = _mm_min_epi16(_mm_max_epi16(ct1, zero), h4);
+      tmp = _mm_madd_epi16(ct0, stride4);
+      addr[0].m = _mm_add_epi32(tmp, cs0);
+      addr[1].m = _mm_add_epi32(tmp, cs1);
+      tmp = _mm_madd_epi16(ct1, stride4);
+      addr[2].m = _mm_add_epi32(tmp, cs0);
+      addr[3].m = _mm_add_epi32(tmp, cs1);
+
+      for (j = 0; j < 4; j++) {
+         __m128i ld1, ld2, ld3;
+         si[j] = _mm_cvtsi32_si128(data[addr[j].ui[0]]);
+         ld1 = _mm_cvtsi32_si128(data[addr[j].ui[1]]);
+         si[j] = _mm_unpacklo_epi32(si[j], ld1);
+         ld2 = _mm_cvtsi32_si128(data[addr[j].ui[2]]);
+         ld3 = _mm_cvtsi32_si128(data[addr[j].ui[3]]);
+         ld2 = _mm_unpacklo_epi32(ld2, ld3);
+         si[j] =  _mm_unpacklo_epi64(si[j], ld2);
+      }
+
+      ws = _mm_srli_epi32(s4, 8);
+      ws = _mm_and_si128(ws, _mm_set1_epi32(0xFF));
+      wt = _mm_srli_epi32(t4, 8);
+      wt = _mm_and_si128(wt, _mm_set1_epi32(0xFF));
+
+      s4 = _mm_add_epi32(s4, dsdx4);
+      t4 = _mm_add_epi32(t4, dtdx4);
+
+#if 0
+/* scalar code for reference */
+      for (j = 0; j < 4; j++) {
+         int s0 = s >> FIXED16_SHIFT;
+         int t0 = t >> FIXED16_SHIFT;
+         int cs0 = CLAMP(s0    , 0, tex_width);
+         int cs1 = CLAMP(s0 + 1, 0, tex_width);
+         int ct0 = CLAMP(t0    , 0, tex_height);
+         int ct1 = CLAMP(t0 + 1, 0, tex_height);
+
+         si0.ui[j] = data[ct0 * stride + cs0];
+         si1.ui[j] = data[ct0 * stride + cs1];
+         si2.ui[j] = data[ct1 * stride + cs0];
+         si3.ui[j] = data[ct1 * stride + cs1];
+
+         ws.ui[j] = (s>>8) & 0xff;
+         wt.ui[j] = (t>>8) & 0xff;
+
+         s += dsdx;
+         t += dtdx;
+      }
+#endif
+
+      ws = _mm_or_si128(ws, _mm_slli_epi32(ws, 16));
+      wsl = _mm_shuffle_epi32(ws, _MM_SHUFFLE(1,1,0,0));
+      wsh = _mm_shuffle_epi32(ws, _MM_SHUFFLE(3,3,2,2));
+
+      wt = _mm_or_si128(wt, _mm_slli_epi32(wt, 16));
+      wtl = _mm_shuffle_epi32(wt, _MM_SHUFFLE(1,1,0,0));
+      wth = _mm_shuffle_epi32(wt, _MM_SHUFFLE(3,3,2,2));
+
+      *(__m128i *)&row[i] = util_sse2_lerp_2d_epi8_fixed88(si[0], si[2],
+                                                           &si[1], &si[3],
+                                                           &wtl, &wth,
+                                                           &wsl, &wsh);
+   }
+
+   samp->s += samp->dsdy;
+   samp->t += samp->dtdy;
+   return row;
+}
+
+static const uint32_t *
+fetch_bgrx_axis_aligned_linear(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const __m128i mask = _mm_set1_epi32(0xff000000);
+   uint32_t *dst_row = samp->row;
+   const uint32_t *src_row;
+   int width = samp->width;
+   int i;
+
+   src_row = fetch_bgra_axis_aligned_linear(&samp->base);
+
+   for (i = 0; i < width; i += 4) {
+      __m128i bgra = *(__m128i *)&src_row[i];
+      __m128i bgrx = _mm_or_si128(bgra, mask);
+      *(__m128i *)&dst_row[i] = bgrx;
+   }
+
+   return dst_row;
+}
+
+
+static const uint32_t *
+fetch_bgrx_clamp_linear(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const __m128i mask = _mm_set1_epi32(0xff000000);
+   uint32_t *row   = samp->row;
+   int width = samp->width;
+   int i;
+
+   fetch_bgra_clamp_linear(&samp->base);
+
+   for (i = 0; i < width; i += 4) {
+      __m128i bgra = *(__m128i *)&row[i];
+      __m128i bgrx = _mm_or_si128(bgra, mask);
+      *(__m128i *)&row[i] = bgrx;
+   }
+
+   return row;
+}
+
+
+static const uint32_t *
+fetch_bgrx_linear(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   const __m128i mask = _mm_set1_epi32(0xff000000);
+   uint32_t *row   = samp->row;
+   int width = samp->width;
+   int i;
+
+   fetch_bgra_linear(&samp->base);
+
+   for (i = 0; i < width; i += 4) {
+      __m128i bgra = *(__m128i *)&row[i];
+      __m128i bgrx = _mm_or_si128(bgra, mask);
+      *(__m128i *)&row[i] = bgrx;
+   }
+
+   return row;
+}
+
+
+static boolean
+sampler_is_nearest(const struct lp_linear_sampler *samp,
+                   const struct lp_sampler_static_state *sampler_state,
+                   boolean minify)
+{
+   unsigned img_filter;
+
+   if (minify)
+      img_filter = sampler_state->sampler_state.min_img_filter;
+   else
+      img_filter = sampler_state->sampler_state.mag_img_filter;
+
+   /* Is it obviously nearest?
+    */
+   if (img_filter == PIPE_TEX_FILTER_NEAREST)
+      return TRUE;
+
+   /* Otherwise look for linear samplers which devolve to nearest.
+    */
+
+   /* Needs to be axis aligned.
+    */
+   if (!samp->axis_aligned)
+      return FALSE;
+
+   if (0) {
+      /* For maximizing shaders, revert to nearest
+       */
+      if (samp->dsdx < -FIXED16_HALF && samp->dsdx < FIXED16_HALF &&
+          samp->dtdy < -FIXED16_HALF && samp->dtdy < FIXED16_HALF)
+         return TRUE;
+
+      /* For severely minimising shaders, revert to nearest:
+       */
+      if ((samp->dsdx < 2 * FIXED16_ONE || samp->dsdx > 2 * FIXED16_ONE) &&
+          (samp->dtdy < 2 * FIXED16_ONE || samp->dtdy > 2 * FIXED16_ONE))
+         return TRUE;
+   }
+
+   /*
+    * Must be near a pixel center:
+    */
+   if (!fixed16_approx(fixed16_frac(samp->s), FIXED16_HALF, FIXED16_TOL) ||
+       !fixed16_approx(fixed16_frac(samp->t), FIXED16_HALF, FIXED16_TOL))
+      return FALSE;
+
+   /*
+    * Must make a full step between pixels:
+    */
+   if (!fixed16_approx(samp->dsdx, FIXED16_ONE, FIXED16_TOL_DERIV) ||
+       !fixed16_approx(samp->dtdy, FIXED16_ONE, FIXED16_TOL_DERIV))
+      return FALSE;
+
+   /* Treat it as nearest!
+    */
+   return TRUE;
+}
+
+/* XXX: Lots of static-state parameters being passed in here but very
+ * little info is extracted from each one.  Consolidate it all down to
+ * something succinct in the prepare phase?
+ */
+boolean
+lp_linear_init_sampler(struct lp_linear_sampler *samp,
+                       const struct lp_tgsi_texture_info *info,
+                       const struct lp_sampler_static_state *sampler_state,
+                       const struct lp_jit_texture *texture,
+                       int x0, int y0, int width, int height,
+                       const float (*a0)[4],
+                       const float (*dadx)[4],
+                       const float (*dady)[4])
+{
+   const struct lp_tgsi_channel_info *schan = &info->coord[0];
+   const struct lp_tgsi_channel_info *tchan = &info->coord[1];
+
+   float w0   =   a0[0][3];
+
+   float s0   =   a0[schan->u.index+1][schan->swizzle];
+   float dsdx = dadx[schan->u.index+1][schan->swizzle];
+   float dsdy = dady[schan->u.index+1][schan->swizzle];
+
+   float t0   =   a0[tchan->u.index+1][tchan->swizzle];
+   float dtdx = dadx[tchan->u.index+1][tchan->swizzle];
+   float dtdy = dady[tchan->u.index+1][tchan->swizzle];
+
+   int mins, mint, maxs, maxt;
+   float oow = 1.0f / w0;
+   float width_oow = texture->width * oow;
+   float height_oow = texture->height * oow;
+   float fdsdx = dsdx * width_oow;
+   float fdsdy = dsdy * width_oow;
+   float fdtdx = dtdx * height_oow;
+   float fdtdy = dtdy * height_oow;
+   int fetch_width;
+   int fetch_height;
+   boolean minify;
+   boolean need_wrap;
+   boolean is_nearest;
+
+   samp->texture = texture;
+   samp->width = width;
+
+   samp->s = float_to_fixed16(fdsdx * x0 +
+                              fdsdy * y0 +
+                              s0 * width_oow);
+
+   samp->t = float_to_fixed16(fdtdx * x0 +
+                              fdtdy * y0 +
+                              t0 * height_oow);
+
+   samp->dsdx = float_to_fixed16(fdsdx);
+   samp->dsdy = float_to_fixed16(fdsdy);
+   samp->dtdx = float_to_fixed16(fdtdx);
+   samp->dtdy = float_to_fixed16(fdtdy);
+
+
+   samp->axis_aligned = (samp->dsdy == 0 &&
+                         samp->dtdx == 0); // TODO: could be relaxed
+
+   {
+      int dsdx = samp->dsdx >= 0 ? samp->dsdx : -samp->dsdx;
+      int dsdy = samp->dsdy >= 0 ? samp->dsdy : -samp->dsdy;
+      int dtdx = samp->dtdx >= 0 ? samp->dtdx : -samp->dtdx;
+      int dtdy = samp->dtdy >= 0 ? samp->dtdy : -samp->dtdy;
+      int rho = MAX4(dsdx, dsdy, dtdx, dtdy);
+
+      minify = (rho > FIXED16_ONE);
+   }
+
+   is_nearest = sampler_is_nearest(samp, sampler_state, minify);
+
+   if (!is_nearest) {
+      samp->s -= FIXED16_HALF;
+      samp->t -= FIXED16_HALF;
+   }
+
+   /* Check for clamping.  This rarely happens as we're rejecting interpolants
+    * which fall outside the 0..1 range.
+    */
+
+   if (is_nearest) {
+      /* Nearest fetch routines don't employ SSE and always operate one pixel
+       * at a time.
+       */
+      fetch_width = width - 1;
+   }
+   else {
+      /* Linear fetch routines employ SSE, and always fetch groups of four
+       * texels.
+       */
+      fetch_width = align(width, 4) - 1;
+   }
+   fetch_height = height - 1;
+
+   if (samp->axis_aligned) {
+      int s0 = samp->s;
+      int s1 = samp->s + fetch_width  * samp->dsdx;
+      int t0 = samp->t;
+      int t1 = samp->t + fetch_height * samp->dtdy;
+
+      mins = MIN2(s0, s1);
+      mint = MIN2(t0, t1);
+      maxs = MAX2(s0, s1);
+      maxt = MAX2(t0, t1);
+   }
+   else {
+      int s0 = samp->s;
+      int s1 = samp->s + fetch_width  * samp->dsdx;
+      int s2 = samp->s + fetch_height * samp->dsdy;
+      int s3 = samp->s + fetch_width  * samp->dsdx + fetch_height * samp->dsdy;
+      int t0 = samp->t;
+      int t1 = samp->t + fetch_width  * samp->dtdx;
+      int t2 = samp->t + fetch_height * samp->dtdy;
+      int t3 = samp->t + fetch_width  * samp->dtdx + fetch_height * samp->dtdy;
+
+      mins = MIN4(s0, s1, s2, s3);
+      mint = MIN4(t0, t1, t2, t3);
+      maxs = MAX4(s0, s1, s2, s3);
+      maxt = MAX4(t0, t1, t2, t3);
+   }
+
+   if (is_nearest) {
+      need_wrap = (mins < 0 ||
+                   mint < 0 ||
+                   maxs >= (texture->width  << FIXED16_SHIFT) ||
+                   maxt >= (texture->height << FIXED16_SHIFT));
+   } else {
+      need_wrap = (mins < 0 ||
+                   mint < 0 ||
+                   maxs + FIXED16_ONE >= (texture->width  << FIXED16_SHIFT) ||
+                   maxt + FIXED16_ONE >= (texture->height << FIXED16_SHIFT));
+   }
+
+   if (0 && need_wrap) {
+      debug_printf("%u x %u %s\n",
+                   texture->width, texture->height,
+                   is_nearest ? "nearest" : "linear");
+      debug_printf("mins = %f\n", mins*1.0f/FIXED16_ONE);
+      debug_printf("mint = %f\n", mint*1.0f/FIXED16_ONE);
+      debug_printf("maxs = %f\n", maxs*1.0f/FIXED16_ONE);
+      debug_printf("maxt = %f\n", maxt*1.0f/FIXED16_ONE);
+      debug_printf("\n");
+   }
+
+   /* We accept any mode below, but we only implement clamping.
+    */
+   if (need_wrap &&
+       (sampler_state->sampler_state.wrap_s != PIPE_TEX_WRAP_CLAMP_TO_EDGE ||
+        sampler_state->sampler_state.wrap_t != PIPE_TEX_WRAP_CLAMP_TO_EDGE)) {
+       return FALSE;
+   }
+
+   if (is_nearest) {
+      switch (sampler_state->texture_state.format) {
+      case PIPE_FORMAT_B8G8R8A8_UNORM:
+         if (need_wrap)
+            samp->base.fetch = fetch_bgra_clamp;
+         else if (!samp->axis_aligned)
+            samp->base.fetch = fetch_bgra;
+         else if (samp->dsdx != FIXED16_ONE) // TODO: could be relaxed
+            samp->base.fetch = fetch_bgra_axis_aligned;
+         else
+            samp->base.fetch = fetch_bgra_memcpy;
+
+         return TRUE;
+
+      case PIPE_FORMAT_B8G8R8X8_UNORM:
+         if (need_wrap)
+            samp->base.fetch = fetch_bgrx_clamp;
+         else if (!samp->axis_aligned)
+            samp->base.fetch = fetch_bgrx;
+         else if (samp->dsdx != FIXED16_ONE) // TODO: could be relaxed
+            samp->base.fetch = fetch_bgrx_axis_aligned;
+         else
+            samp->base.fetch = fetch_bgrx_memcpy;
+
+         return TRUE;
+
+      default:
+         break;
+      }
+
+      FAIL("unknown format for nearest");
+   }
+   else {
+      samp->stretched_row_y[0] = -1;
+      samp->stretched_row_y[1] = -1;
+      samp->stretched_row_index = 0;
+
+      switch (sampler_state->texture_state.format) {
+      case PIPE_FORMAT_B8G8R8A8_UNORM:
+         if (need_wrap)
+            samp->base.fetch = fetch_bgra_clamp_linear;
+         else if (!samp->axis_aligned)
+            samp->base.fetch = fetch_bgra_linear;
+         else
+            samp->base.fetch = fetch_bgra_axis_aligned_linear;
+
+         return TRUE;
+
+      case PIPE_FORMAT_B8G8R8X8_UNORM:
+         if (need_wrap)
+            samp->base.fetch = fetch_bgrx_clamp_linear;
+         else if (!samp->axis_aligned)
+            samp->base.fetch = fetch_bgrx_linear;
+         else
+            samp->base.fetch = fetch_bgrx_axis_aligned_linear;
+         return TRUE;
+
+      default:
+         break;
+      }
+
+      FAIL("unknown format");
+   }
+}
+
+
+static const uint32_t *
+fetch_noop(struct lp_linear_elem *elem)
+{
+   struct lp_linear_sampler *samp = (struct lp_linear_sampler *)elem;
+   return samp->row;
+}
+
+
+void
+lp_linear_init_noop_sampler(struct lp_linear_sampler *samp)
+{
+   samp->base.fetch = fetch_noop;
+}
+
+/* Check the variant for linear path compatibility.
+ */
+boolean
+lp_linear_check_sampler(const struct lp_sampler_static_state *sampler,
+                        const struct lp_tgsi_texture_info *tex)
+{
+   if (tex->modifier != LP_BLD_TEX_MODIFIER_NONE)
+      return FALSE;
+
+   if (tex->target != TGSI_TEXTURE_2D)
+      return FALSE;
+
+   if (tex->coord[0].file != TGSI_FILE_INPUT ||
+       tex->coord[1].file != TGSI_FILE_INPUT)
+      return FALSE;
+
+   /* These are the only sampling modes we support at the moment.
+    *
+    * Actually we'll accept any mode as we're failing on any
+    * interpolant which exceeds 0..1.  Clamping is applied only to
+    * avoid invalid reads.
+    */
+   if (!is_nearest_sampler(sampler) &&
+       !is_linear_sampler(sampler))
+      return FALSE;
+
+   /* These are the only texture formats we support at the moment
+    */
+   if (sampler->texture_state.format != PIPE_FORMAT_B8G8R8A8_UNORM &&
+       sampler->texture_state.format != PIPE_FORMAT_B8G8R8X8_UNORM)
+      return FALSE;
+
+   return TRUE;
+}
+
+#else
+boolean
+lp_linear_check_sampler(const struct lp_sampler_static_state *sampler,
+                        const struct lp_tgsi_texture_info *tex)
+{
+   return FALSE;
+}
+#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_perf.c b/src/gallium/drivers/llvmpipe/lp_perf.c
index a4548bccf1c..6c9fafaad82 100644
--- a/src/gallium/drivers/llvmpipe/lp_perf.c
+++ b/src/gallium/drivers/llvmpipe/lp_perf.c
@@ -50,6 +50,8 @@ lp_print_counters(void)
 
       debug_printf("llvmpipe: nr_triangles:                 %9u\n", lp_count.nr_tris);
       debug_printf("llvmpipe: nr_culled_triangles:          %9u\n", lp_count.nr_culled_tris);
+      debug_printf("llvmpipe: nr_rectangles:                %9u\n", lp_count.nr_rects);
+      debug_printf("llvmpipe: nr_culled_rectangles:         %9u\n", lp_count.nr_culled_rects);
 
       total_64 = (lp_count.nr_empty_64 + 
                   lp_count.nr_fully_covered_64 +
@@ -58,11 +60,14 @@ lp_print_counters(void)
       p1 = 100.0 * (float) lp_count.nr_empty_64 / (float) total_64;
       p2 = 100.0 * (float) lp_count.nr_fully_covered_64 / (float) total_64;
       p3 = 100.0 * (float) lp_count.nr_partially_covered_64 / (float) total_64;
+      p4 = 100.0 * (float) lp_count.nr_blit_64 / (float) total_64;
       p5 = 100.0 * (float) lp_count.nr_shade_opaque_64 / (float) total_64;
       p6 = 100.0 * (float) lp_count.nr_shade_64 / (float) total_64;
 
       debug_printf("llvmpipe: nr_64x64:                     %9u\n", total_64);
       debug_printf("llvmpipe:   nr_fully_covered_64x64:     %9u (%3.0f%% of %u)\n", lp_count.nr_fully_covered_64, p2, total_64);
+      debug_printf("llvmpipe:     nr_blit_64x64:            %9u (%3.0f%% of %u)\n", lp_count.nr_blit_64, p4, total_64);
+      debug_printf("llvmpipe:        nr_pure_blit_64x64:    %9u (%3.0f%% of %u)\n", lp_count.nr_pure_blit_64, 0.0, lp_count.nr_blit_64);
       debug_printf("llvmpipe:     nr_shade_opaque_64x64:    %9u (%3.0f%% of %u)\n", lp_count.nr_shade_opaque_64, p5, total_64);
       debug_printf("llvmpipe:        nr_pure_shade_opaque:  %9u (%3.0f%% of %u)\n", lp_count.nr_pure_shade_opaque_64, 0.0, lp_count.nr_shade_opaque_64);
       debug_printf("llvmpipe:     nr_shade_64x64:           %9u (%3.0f%% of %u)\n", lp_count.nr_shade_64, p6, total_64);
@@ -98,6 +103,17 @@ lp_print_counters(void)
       debug_printf("llvmpipe:   nr_empty_4x4:               %9u (%3.0f%% of %u)\n", lp_count.nr_empty_4, p1, total_4);
       debug_printf("llvmpipe:   nr_non_empty_4x4:           %9u (%3.0f%% of %u)\n", lp_count.nr_non_empty_4, p4, total_4);
 
+      total_4 = (lp_count.nr_rect_partially_covered_4 +
+                 lp_count.nr_rect_fully_covered_4);
+
+      p1 = 100.0 * (float) lp_count.nr_rect_partially_covered_4 / (float) total_4;
+      p2 = 100.0 * (float) lp_count.nr_rect_fully_covered_4 / (float) total_4;
+
+      debug_printf("llvmpipe: nr_rect_4x4:                  %9u\n", total_4);
+      debug_printf("llvmpipe:   nr_rect_full_4x4:           %9u (%3.0f%% of %u)\n", lp_count.nr_rect_fully_covered_4, p1, total_4);
+      debug_printf("llvmpipe:   nr_rect_part_4x4:           %9u (%3.0f%% of %u)\n", lp_count.nr_rect_partially_covered_4, p2, total_4);
+
+
       debug_printf("llvmpipe: nr_color_tile_clear:          %9u\n", lp_count.nr_color_tile_clear);
       debug_printf("llvmpipe: nr_color_tile_load:           %9u\n", lp_count.nr_color_tile_load);
       debug_printf("llvmpipe: nr_color_tile_store:          %9u\n", lp_count.nr_color_tile_store);
diff --git a/src/gallium/drivers/llvmpipe/lp_perf.h b/src/gallium/drivers/llvmpipe/lp_perf.h
index ace85c7fef4..61e3fa8ddbc 100644
--- a/src/gallium/drivers/llvmpipe/lp_perf.h
+++ b/src/gallium/drivers/llvmpipe/lp_perf.h
@@ -42,9 +42,13 @@ struct lp_counters
 {
    unsigned nr_tris;
    unsigned nr_culled_tris;
+   unsigned nr_rects;
+   unsigned nr_culled_rects;
    unsigned nr_empty_64;
    unsigned nr_fully_covered_64;
    unsigned nr_partially_covered_64;
+   unsigned nr_blit_64;
+   unsigned nr_pure_blit_64;
    unsigned nr_pure_shade_opaque_64;
    unsigned nr_pure_shade_64;
    unsigned nr_shade_64;
@@ -55,6 +59,8 @@ struct lp_counters
    unsigned nr_empty_4;
    unsigned nr_fully_covered_4;
    unsigned nr_partially_covered_4;
+   unsigned nr_rect_fully_covered_4;
+   unsigned nr_rect_partially_covered_4;
    unsigned nr_non_empty_4;
    unsigned nr_llvm_compiles;
    int64_t llvm_compile_time;  /**< total, in microseconds */
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.c b/src/gallium/drivers/llvmpipe/lp_rast.c
index 6b5160667bf..87dc68b992e 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast.c
@@ -508,6 +508,126 @@ lp_rast_shade_quads_mask(struct lp_rasterizer_task *task,
    lp_rast_shade_quads_mask_sample(task, inputs, x, y, new_mask);
 }
 
+/**
+ * Directly copy pixels from a texture to the destination color buffer.
+ * This is a bin command called during bin processing.
+ */
+static void
+lp_rast_blit_tile_to_dest(struct lp_rasterizer_task *task,
+                          const union lp_rast_cmd_arg arg)
+{
+   const struct lp_scene *scene = task->scene;
+   const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
+   const struct lp_rast_state *state = task->state;
+   struct lp_fragment_shader_variant *variant = state->variant;
+   const struct lp_jit_texture *texture = &state->jit_context.textures[0];
+   const uint8_t *src;
+   uint8_t *dst;
+   unsigned src_stride;
+   unsigned dst_stride;
+   struct pipe_surface *cbuf = scene->fb.cbufs[0];
+   const unsigned face_slice = cbuf->u.tex.first_layer;
+   const unsigned level = cbuf->u.tex.level;
+   struct llvmpipe_resource *lpt = llvmpipe_resource(cbuf->texture);
+   int src_x, src_y;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   if (inputs->disable) {
+      /* This command was partially binned and has been disabled */
+      return;
+   }
+
+   dst = llvmpipe_get_texture_image_address(lpt, face_slice, level);
+
+   if (!dst)
+      return;
+
+   dst_stride = lpt->row_stride[level];
+
+   src = texture->base;
+   src_stride = texture->row_stride[0];
+
+   src_x = util_iround(GET_A0(inputs)[1][0]*texture->width - 0.5f);
+   src_y = util_iround(GET_A0(inputs)[1][1]*texture->height - 0.5f);
+
+   src_x = src_x + task->x;
+   src_y = src_y + task->y;
+
+   if (0) {
+      union util_color uc;
+      uc.ui[0] = 0xff0000ff;
+      util_fill_rect(dst,
+                     cbuf->format,
+                     dst_stride,
+                     task->x,
+                     task->y,
+                     task->width,
+                     task->height,
+                     &uc);
+      return;
+   }
+
+   if (src_x >= 0 &&
+       src_y >= 0 &&
+       src_x + task->width <= texture->width &&
+       src_y + task->height <= texture->height) {
+
+      if (variant->shader->kind == LP_FS_KIND_BLIT_RGBA ||
+          (variant->shader->kind == LP_FS_KIND_BLIT_RGB1 &&
+           cbuf->format == PIPE_FORMAT_B8G8R8X8_UNORM)) {
+         util_copy_rect(dst,
+                        cbuf->format,
+                        dst_stride,
+                        task->x, task->y,
+                        task->width, task->height,
+                        src, src_stride,
+                        src_x, src_y);
+         return;
+      }
+
+      if (variant->shader->kind == LP_FS_KIND_BLIT_RGB1) {
+         if (cbuf->format == PIPE_FORMAT_B8G8R8A8_UNORM) {
+            int x, y;
+
+            dst += task->x * 4;
+            src += src_x * 4;
+            dst += task->y * dst_stride;
+            src += src_y * src_stride;
+
+            for (y = 0; y < task->height; ++y) {
+               const uint32_t *src_row = (const uint32_t *)src;
+               uint32_t *dst_row = (uint32_t *)dst;
+
+               for (x = 0; x < task->width; ++x) {
+                  *dst_row++ = *src_row++ | 0xff000000;
+               }
+               dst += dst_stride;
+               src += src_stride;
+            }
+
+            return;
+         }
+      }
+
+   }
+
+   /*
+    * Fall back to the jit shaders.
+    */
+
+   lp_rast_shade_tile_opaque(task, arg);
+}
+
+static void
+lp_rast_blit_tile(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg)
+{
+   /* This kindof just works, but isn't efficient:
+    */
+   lp_rast_blit_tile_to_dest(task, arg);
+}
+
 /**
  * Begin a new occlusion query.
  * This is a bin command put in all bins.
@@ -597,8 +717,123 @@ lp_rast_tile_end(struct lp_rasterizer_task *task)
    task->bin = NULL;
 }
 
-static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
-{
+
+
+
+ 
+
+/* Currently have two rendering paths only - the general case triangle
+ * path and the super-specialized blit/clear path.
+ */
+#define TRI   ((LP_RAST_FLAGS_TRI <<1)-1)        /* general case */
+#define RECT  ((LP_RAST_FLAGS_RECT<<1)-1)        /* direct rectangle rasterizer */
+#define BLIT  ((LP_RAST_FLAGS_BLIT<<1)-1)        /* write direct-to-dest */
+
+static const unsigned
+rast_flags[] = {
+   BLIT,                        /* clear color */
+   TRI,                         /* clear zstencil */
+   TRI,                         /* triangle_1 */
+   TRI,                         /* triangle_2 */
+   TRI,                         /* triangle_3 */
+   TRI,                         /* triangle_4 */
+   TRI,                         /* triangle_5 */
+   TRI,                         /* triangle_6 */
+   TRI,                         /* triangle_7 */
+   TRI,                         /* triangle_8 */
+   TRI,                         /* triangle_3_4 */
+   TRI,                         /* triangle_3_16 */
+   TRI,                         /* triangle_4_16 */
+   RECT,                        /* shade_tile */
+   RECT,                        /* shade_tile_opaque */
+   TRI,                         /* begin_query */
+   TRI,                         /* end_query */
+   BLIT,                        /* set_state, */
+   TRI,                         /* lp_rast_triangle_32_1 */
+   TRI,                         /* lp_rast_triangle_32_2 */
+   TRI,                         /* lp_rast_triangle_32_3 */
+   TRI,                         /* lp_rast_triangle_32_4 */
+   TRI,                         /* lp_rast_triangle_32_5 */
+   TRI,                         /* lp_rast_triangle_32_6 */
+   TRI,                         /* lp_rast_triangle_32_7 */
+   TRI,                         /* lp_rast_triangle_32_8 */
+   TRI,                         /* lp_rast_triangle_32_3_4 */
+   TRI,                         /* lp_rast_triangle_32_3_16 */
+   TRI,                         /* lp_rast_triangle_32_4_16 */
+   TRI,                         /* lp_rast_triangle_ms_1 */
+   TRI,                         /* lp_rast_triangle_ms_2 */
+   TRI,                         /* lp_rast_triangle_ms_3 */
+   TRI,                         /* lp_rast_triangle_ms_4 */
+   TRI,                         /* lp_rast_triangle_ms_5 */
+   TRI,                         /* lp_rast_triangle_ms_6 */
+   TRI,                         /* lp_rast_triangle_ms_7 */
+   TRI,                         /* lp_rast_triangle_ms_8 */
+   TRI,                         /* lp_rast_triangle_ms_3_4 */
+   TRI,                         /* lp_rast_triangle_ms_3_16 */
+   TRI,                         /* lp_rast_triangle_ms_4_16 */
+
+   RECT,                        /* rectangle */
+   BLIT,                        /* blit */
+};
+
+/* 
+ */
+static const lp_rast_cmd_func
+dispatch_blit[] = {
+   lp_rast_clear_color,
+   NULL,                        /* clear_zstencil */
+   NULL,                        /* triangle_1 */
+   NULL,                        /* triangle_2 */
+   NULL,                        /* triangle_3 */
+   NULL,                        /* triangle_4 */
+   NULL,                        /* triangle_5 */
+   NULL,                        /* triangle_6 */
+   NULL,                        /* triangle_7 */
+   NULL,                        /* triangle_8 */
+   NULL,                        /* triangle_3_4 */
+   NULL,                        /* triangle_3_16 */
+   NULL,                        /* triangle_4_16 */
+   NULL,                        /* shade_tile */
+   NULL,                        /* shade_tile_opaque */
+   NULL,                        /* begin_query */
+   NULL,                        /* end_query */
+   lp_rast_set_state,           /* set_state */
+   NULL,                        /* lp_rast_triangle_32_1 */
+   NULL,                        /* lp_rast_triangle_32_2 */
+   NULL,                        /* lp_rast_triangle_32_3 */
+   NULL,                        /* lp_rast_triangle_32_4 */
+   NULL,                        /* lp_rast_triangle_32_5 */
+   NULL,                        /* lp_rast_triangle_32_6 */
+   NULL,                        /* lp_rast_triangle_32_7 */
+   NULL,                        /* lp_rast_triangle_32_8 */
+   NULL,                        /* lp_rast_triangle_32_3_4 */
+   NULL,                        /* lp_rast_triangle_32_3_16 */
+   NULL,                        /* lp_rast_triangle_32_4_16 */
+   NULL,                        /* lp_rast_triangle_ms_1 */
+   NULL,                        /* lp_rast_triangle_ms_2 */
+   NULL,                        /* lp_rast_triangle_ms_3 */
+   NULL,                        /* lp_rast_triangle_ms_4 */
+   NULL,                        /* lp_rast_triangle_ms_5 */
+   NULL,                        /* lp_rast_triangle_ms_6 */
+   NULL,                        /* lp_rast_triangle_ms_7 */
+   NULL,                        /* lp_rast_triangle_ms_8 */
+   NULL,                        /* lp_rast_triangle_ms_3_4 */
+   NULL,                        /* lp_rast_triangle_ms_3_16 */
+   NULL,                        /* lp_rast_triangle_ms_4_16 */
+
+   NULL,                        /* rectangle */
+   lp_rast_blit_tile_to_dest,
+};
+
+
+
+/* Triangle and general case rasterization: Use the SOA llvm shdaers,
+ * an active swizzled tile for each color buf, etc.  Don't blit/clear
+ * directly to destination surface as we know there are swizzled
+ * operations coming.
+ */
+static const lp_rast_cmd_func
+dispatch_tri[] = {
    lp_rast_clear_color,
    lp_rast_clear_zstencil,
    lp_rast_triangle_1,
@@ -639,27 +874,133 @@ static lp_rast_cmd_func dispatch[LP_RAST_OP_MAX] =
    lp_rast_triangle_ms_3_4,
    lp_rast_triangle_ms_3_16,
    lp_rast_triangle_ms_4_16,
+   lp_rast_rectangle,
+   lp_rast_blit_tile,
 };
 
 
+/* Debug rasterization with most fastpaths disabled.
+ */
+static const lp_rast_cmd_func
+dispatch_tri_debug[] =
+{
+   lp_rast_clear_color,
+   lp_rast_clear_zstencil,
+   lp_rast_triangle_1,
+   lp_rast_triangle_2,
+   lp_rast_triangle_3,
+   lp_rast_triangle_4,
+   lp_rast_triangle_5,
+   lp_rast_triangle_6,
+   lp_rast_triangle_7,
+   lp_rast_triangle_8,
+   lp_rast_triangle_3_4,
+   lp_rast_triangle_3_16,
+   lp_rast_triangle_4_16,
+   lp_rast_shade_tile,
+   lp_rast_shade_tile,
+   lp_rast_begin_query,
+   lp_rast_end_query,
+   lp_rast_set_state,
+   lp_rast_triangle_32_1,
+   lp_rast_triangle_32_2,
+   lp_rast_triangle_32_3,
+   lp_rast_triangle_32_4,
+   lp_rast_triangle_32_5,
+   lp_rast_triangle_32_6,
+   lp_rast_triangle_32_7,
+   lp_rast_triangle_32_8,
+   lp_rast_triangle_32_3_4,
+   lp_rast_triangle_32_3_16,
+   lp_rast_triangle_32_4_16,
+   lp_rast_triangle_ms_1,
+   lp_rast_triangle_ms_2,
+   lp_rast_triangle_ms_3,
+   lp_rast_triangle_ms_4,
+   lp_rast_triangle_ms_5,
+   lp_rast_triangle_ms_6,
+   lp_rast_triangle_ms_7,
+   lp_rast_triangle_ms_8,
+   lp_rast_triangle_ms_3_4,
+   lp_rast_triangle_ms_3_16,
+   lp_rast_triangle_ms_4_16,
+
+   lp_rast_rectangle,
+   lp_rast_shade_tile,
+};
+
+struct lp_bin_info
+lp_characterize_bin(const struct cmd_bin *bin)
+{
+   struct cmd_block *block;
+   struct lp_bin_info info;
+   unsigned andflags = ~0;
+   unsigned k, j = 0;
+
+   STATIC_ASSERT(ARRAY_SIZE(rast_flags) == LP_RAST_OP_MAX);
+
+   for (block = bin->head; block; block = block->next) {
+      for (k = 0; k < block->count; k++, j++) {
+         andflags &= rast_flags[block->cmd[k]];
+      }
+   }
+   
+   info.type = andflags;
+   info.count = j;
+
+   return info;
+}
+
+
 static void
-do_rasterize_bin(struct lp_rasterizer_task *task,
-                 const struct cmd_bin *bin,
-                 int x, int y)
+blit_rasterize_bin(struct lp_rasterizer_task *task,
+                   const struct cmd_bin *bin)
 {
    const struct cmd_block *block;
    unsigned k;
 
-   if (0)
-      lp_debug_bin(bin, x, y);
+   STATIC_ASSERT(ARRAY_SIZE(dispatch_blit) == LP_RAST_OP_MAX);
 
+   if (0) debug_printf("%s\n", __FUNCTION__);
    for (block = bin->head; block; block = block->next) {
       for (k = 0; k < block->count; k++) {
-         dispatch[block->cmd[k]]( task, block->arg[k] );
+         dispatch_blit[block->cmd[k]]( task, block->arg[k] );
       }
    }
 }
 
+static void
+tri_rasterize_bin(struct lp_rasterizer_task *task,
+                  const struct cmd_bin *bin,
+                  int x, int y)
+{
+   const struct cmd_block *block;
+   unsigned k;
+
+   STATIC_ASSERT(ARRAY_SIZE(dispatch_tri) == LP_RAST_OP_MAX);
+
+   for (block = bin->head; block; block = block->next) {
+      for (k = 0; k < block->count; k++) {
+         dispatch_tri[block->cmd[k]]( task, block->arg[k] );
+      }
+   }
+}
+
+static void
+debug_rasterize_bin(struct lp_rasterizer_task *task,
+                  const struct cmd_bin *bin)
+{
+   const struct cmd_block *block;
+   unsigned k;
+
+   STATIC_ASSERT(ARRAY_SIZE(dispatch_tri_debug) == LP_RAST_OP_MAX);
+
+   for (block = bin->head; block; block = block->next) {
+      for (k = 0; k < block->count; k++) {
+         dispatch_tri_debug[block->cmd[k]]( task, block->arg[k] );
+      }
+   }
+}
 
 
 /**
@@ -672,9 +1013,20 @@ static void
 rasterize_bin(struct lp_rasterizer_task *task,
               const struct cmd_bin *bin, int x, int y )
 {
+   struct lp_bin_info info = lp_characterize_bin(bin);
+
    lp_rast_tile_begin( task, bin, x, y );
 
-   do_rasterize_bin(task, bin, x, y);
+   if (LP_DEBUG & DEBUG_NO_FASTPATH)
+      debug_rasterize_bin(task, bin);
+   else if (info.type & LP_RAST_FLAGS_BLIT)
+      blit_rasterize_bin(task, bin);
+   else if (task->scene->permit_linear_rasterizer &&
+            !(LP_PERF & PERF_NO_RAST_LINEAR) &&
+            (info.type & LP_RAST_FLAGS_RECT))
+      lp_linear_rasterize_bin(task, bin);
+   else
+      tri_rasterize_bin(task, bin, x, y);
 
    lp_rast_tile_end(task);
 
@@ -682,7 +1034,9 @@ rasterize_bin(struct lp_rasterizer_task *task,
    /* Debug/Perf flags:
     */
    if (bin->head->count == 1) {
-      if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
+      if (bin->head->cmd[0] == LP_RAST_OP_BLIT)
+         LP_COUNT(nr_pure_blit_64);
+      else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE_OPAQUE)
          LP_COUNT(nr_pure_shade_opaque_64);
       else if (bin->head->cmd[0] == LP_RAST_OP_SHADE_TILE)
          LP_COUNT(nr_pure_shade_64);
diff --git a/src/gallium/drivers/llvmpipe/lp_rast.h b/src/gallium/drivers/llvmpipe/lp_rast.h
index 2a2fdceb77f..6e7a16ed649 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast.h
@@ -39,6 +39,7 @@
 
 #include "pipe/p_compiler.h"
 #include "util/u_pack_color.h"
+#include "util/u_rect.h"
 #include "lp_jit.h"
 
 
@@ -92,6 +93,15 @@ struct lp_rast_state {
 };
 
 
+/**
+ * Texture blit offsets.
+ */
+struct lp_rast_blit {
+   int16_t x0;
+   int16_t y0;
+};
+
+
 /**
  * Coefficients necessary to run the shader at a given location.
  * First coefficient is position.
@@ -101,7 +111,8 @@ struct lp_rast_shader_inputs {
    unsigned frontfacing:1;      /** True for front-facing */
    unsigned disable:1;          /** Partially binned, disable this command */
    unsigned opaque:1;           /** Is opaque */
-   unsigned pad0:13;            /* wasted space */
+   unsigned is_blit:1;          /* blit */
+   unsigned pad0:12;            /* wasted space */
    unsigned view_index:16;
    unsigned stride;             /* how much to advance data between a0, dadx, dady */
    unsigned layer;              /* the layer to render to (from gs, already clamped) */
@@ -146,6 +157,31 @@ struct lp_rast_triangle {
 };
 
 
+#define RECT_PLANE_LEFT   0x1
+#define RECT_PLANE_RIGHT  0x2
+#define RECT_PLANE_TOP    0x4
+#define RECT_PLANE_BOTTOM 0x8
+
+/**
+ * Rasterization information for a screen-aligned rectangle known to
+ * be in this bin, plus inputs to run the shader:
+ * These fields are tile- and bin-independent.
+ * Objects of this type are put into the lp_setup_context::data buffer.
+ */
+struct lp_rast_rectangle {
+#ifdef DEBUG
+   float v[2][2]; /**< diagonal corners */
+#endif
+
+   /* Rectangle boundaries in integer pixels:
+    */
+   struct u_rect box;
+
+   /* inputs for the shader */
+   struct lp_rast_shader_inputs inputs;
+};
+
+
 struct lp_rast_clear_rb {
    union util_color color_val;
    unsigned cbuf;
@@ -179,6 +215,7 @@ union lp_rast_cmd_arg {
       const struct lp_rast_triangle *tri;
       unsigned plane_mask;
    } triangle;
+   const struct lp_rast_rectangle *rectangle;
    const struct lp_rast_state *set_state;
    const struct lp_rast_clear_rb *clear_rb;
    struct {
@@ -227,6 +264,14 @@ lp_rast_arg_triangle_contained( const struct lp_rast_triangle *triangle,
    return arg;
 }
 
+static inline union lp_rast_cmd_arg
+lp_rast_arg_rectangle( const struct lp_rast_rectangle *rectangle )
+{
+   union lp_rast_cmd_arg arg;
+   arg.rectangle = rectangle;
+   return arg;
+}
+
 static inline union lp_rast_cmd_arg
 lp_rast_arg_state( const struct lp_rast_state *state )
 {
@@ -317,9 +362,27 @@ lp_rast_arg_null( void )
 #define LP_RAST_OP_MS_TRIANGLE_3_4   0x25
 #define LP_RAST_OP_MS_TRIANGLE_3_16  0x26
 #define LP_RAST_OP_MS_TRIANGLE_4_16  0x27
-#define LP_RAST_OP_MAX               0x28
+#define LP_RAST_OP_RECTANGLE         0x28  /* Keep at end */
+#define LP_RAST_OP_BLIT              0x29  /* Keep at end */
+
+#define LP_RAST_OP_MAX               0x2a
 #define LP_RAST_OP_MASK              0xff
 
+/* Returned by characterize_bin:
+ */
+#define LP_RAST_FLAGS_TRI            (0x1)
+#define LP_RAST_FLAGS_RECT           (0x2)
+#define LP_RAST_FLAGS_TILE           (0x4)
+#define LP_RAST_FLAGS_BLIT           (0x8)
+
+struct lp_bin_info {
+   unsigned type:8;
+   unsigned count:24;
+};
+
+struct lp_bin_info
+lp_characterize_bin(const struct cmd_bin *bin);
+
 void
 lp_debug_bins( struct lp_scene *scene );
 void
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_debug.c b/src/gallium/drivers/llvmpipe/lp_rast_debug.c
index e36ade0e307..dc46bfd1046 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_debug.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_debug.c
@@ -23,7 +23,7 @@ static char get_label( int i )
 
 
 
-static const char *cmd_names[LP_RAST_OP_MAX] = 
+static const char *cmd_names[] =
 {
    "clear_color",
    "clear_zstencil",
@@ -54,10 +54,24 @@ static const char *cmd_names[LP_RAST_OP_MAX] =
    "triangle_32_3_4",
    "triangle_32_3_16",
    "triangle_32_4_16",
+   "lp_rast_triangle_ms_1",
+   "lp_rast_triangle_ms_2",
+   "lp_rast_triangle_ms_3",
+   "lp_rast_triangle_ms_4",
+   "lp_rast_triangle_ms_5",
+   "lp_rast_triangle_ms_6",
+   "lp_rast_triangle_ms_7",
+   "lp_rast_triangle_ms_8",
+   "lp_rast_triangle_ms_3_4",
+   "lp_rast_triangle_ms_3_16",
+   "lp_rast_triangle_ms_4_16",
+   "rectangle",
+   "blit_tile",
 };
 
 static const char *cmd_name(unsigned cmd)
 {
+   STATIC_ASSERT(ARRAY_SIZE(cmd_names) == LP_RAST_OP_MAX);
    assert(ARRAY_SIZE(cmd_names) > cmd);
    return cmd_names[cmd];
 }
@@ -78,7 +92,9 @@ get_variant( const struct lp_rast_state *state,
        block->cmd[k] == LP_RAST_OP_TRIANGLE_4 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_5 ||
        block->cmd[k] == LP_RAST_OP_TRIANGLE_6 ||
-       block->cmd[k] == LP_RAST_OP_TRIANGLE_7)
+       block->cmd[k] == LP_RAST_OP_TRIANGLE_7 ||
+       block->cmd[k] == LP_RAST_OP_RECTANGLE ||
+       block->cmd[k] == LP_RAST_OP_BLIT)
       return state->variant;
 
    return NULL;
@@ -98,6 +114,39 @@ is_blend( const struct lp_rast_state *state,
    return FALSE;
 }
 
+static boolean
+is_linear( const struct lp_rast_state *state,
+           const struct cmd_block *block,
+           int k )
+{
+   if (block->cmd[k] == LP_RAST_OP_BLIT)
+      return state->variant->jit_linear_blit != NULL;
+
+   if (block->cmd[k] == LP_RAST_OP_SHADE_TILE ||
+       block->cmd[k] == LP_RAST_OP_SHADE_TILE_OPAQUE)
+      return state->variant->jit_linear != NULL;
+
+   if (block->cmd[k] == LP_RAST_OP_RECTANGLE)
+      return state->variant->jit_linear != NULL;
+
+   return FALSE;
+}
+
+
+static const char *
+get_fs_kind( const struct lp_rast_state *state,
+             const struct cmd_block *block,
+             int k )
+{
+   const struct lp_fragment_shader_variant *variant = get_variant(state, block, k);
+
+   if (variant)
+      return lp_debug_fs_kind(variant->shader->kind);
+
+   return "";
+}
+
+
 
 
 static void
@@ -105,9 +154,25 @@ debug_bin( const struct cmd_bin *bin, int x, int y )
 {
    const struct lp_rast_state *state = NULL;
    const struct cmd_block *head = bin->head;
+   const char *type;
+   struct lp_bin_info info;
    int i, j = 0;
 
-   debug_printf("bin %d,%d:\n", x, y);
+   info = lp_characterize_bin(bin);
+   
+   if (info.type & LP_RAST_FLAGS_BLIT)
+      type = "blit";
+   else if (info.type & LP_RAST_FLAGS_TILE)
+      type = "tile";
+   else if (info.type & LP_RAST_FLAGS_RECT)
+      type = "rect";
+   else if (info.type & LP_RAST_FLAGS_TRI)
+      type = "tri";
+   else
+      type = "unknown";
+         
+   
+   debug_printf("bin %d,%d: type %s\n", x, y, type);
                 
    while (head) {
       for (i = 0; i < head->count; i++, j++) {
@@ -138,7 +203,64 @@ static void plot(struct tile *tile,
 
 
 
+/**
+ * Scan the tile in chunks and figure out which pixels to rasterize
+ * for this rectangle.
+ */
+static int
+debug_rectangle(int x, int y,
+                const union lp_rast_cmd_arg arg,
+                struct tile *tile,
+                char val)
+{
+   const struct lp_rast_rectangle *rect = arg.rectangle;
+   boolean blend = tile->state->variant->key.blend.rt[0].blend_enable;
+   unsigned i,j, count = 0;
 
+   /* Check for "disabled" rectangles generated in out-of-memory
+    * conditions.
+    */
+   if (rect->inputs.disable) {
+      /* This command was partially binned and has been disabled */
+      return 0;
+   }
+
+   for (i = 0; i < TILE_SIZE; i++)
+   {
+      for (j = 0; j < TILE_SIZE; j++)
+      {
+         if (rect->box.x0 <= x + i &&
+             rect->box.x1 >= x + i &&
+             rect->box.y0 <= y + j &&
+             rect->box.y1 >= y + j)
+         {
+            plot(tile, i, j, val, blend);
+            count++;
+         }
+      }
+   }
+   return count;
+}
+
+
+static int
+debug_blit_tile(int x, int y,
+                const union lp_rast_cmd_arg arg,
+                struct tile *tile,
+                char val)
+{
+   const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
+   unsigned i,j;
+
+   if (inputs->disable)
+      return 0;
+
+   for (i = 0; i < TILE_SIZE; i++)
+      for (j = 0; j < TILE_SIZE; j++)
+         plot(tile, i, j, val, FALSE);
+   
+   return TILE_SIZE * TILE_SIZE;
+}
 
 
 static int
@@ -259,6 +381,8 @@ do_debug_bin( struct tile *tile,
    for (block = bin->head; block; block = block->next) {
       for (k = 0; k < block->count; k++, j++) {
          boolean blend = is_blend(tile->state, block, k);
+         boolean linear = is_linear(tile->state, block, k);
+         const char *fskind = get_fs_kind(tile->state, block, k);
          char val = get_label(j);
          int count = 0;
             
@@ -272,6 +396,9 @@ do_debug_bin( struct tile *tile,
              block->cmd[k] == LP_RAST_OP_CLEAR_ZSTENCIL)
             count = debug_clear_tile(tx, ty, block->arg[k], tile, val);
 
+         if (block->cmd[k] == LP_RAST_OP_BLIT)
+            count = debug_blit_tile(tx, ty, block->arg[k], tile, val);
+
          if (block->cmd[k] == LP_RAST_OP_SHADE_TILE ||
              block->cmd[k] == LP_RAST_OP_SHADE_TILE_OPAQUE)
             count = debug_shade_tile(tx, ty, block->arg[k], tile, val);
@@ -285,12 +412,20 @@ do_debug_bin( struct tile *tile,
              block->cmd[k] == LP_RAST_OP_TRIANGLE_7)
             count = debug_triangle(tx, ty, block->arg[k], tile, val);
 
+         if (block->cmd[k] == LP_RAST_OP_RECTANGLE)
+            count = debug_rectangle(tx, ty, block->arg[k], tile, val);
+
          if (print_cmds) {
             debug_printf(" % 5d", count);
 
+            debug_printf(" %20s", fskind);
+
             if (blend)
                debug_printf(" blended");
             
+            if (linear)
+               debug_printf(" linear");
+
             debug_printf("\n");
          }
       }
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_linear.c b/src/gallium/drivers/llvmpipe/lp_rast_linear.c
new file mode 100644
index 00000000000..3da1d5f84d1
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_rast_linear.c
@@ -0,0 +1,263 @@
+/**************************************************************************
+ *
+ * Copyright 2009-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+#include <limits.h>
+#include "util/u_memory.h"
+#include "util/u_math.h"
+#include "util/u_rect.h"
+#include "util/u_surface.h"
+#include "util/u_pack_color.h"
+
+#include "lp_scene_queue.h"
+#include "lp_debug.h"
+#include "lp_fence.h"
+#include "lp_perf.h"
+#include "lp_query.h"
+#include "lp_rast.h"
+#include "lp_rast_priv.h"
+#include "lp_scene.h"
+
+
+static void
+lp_rast_linear_clear(struct lp_rasterizer_task *task,
+                     const union lp_rast_cmd_arg arg)
+{
+   const struct lp_scene *scene = task->scene;
+   union util_color uc;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   uc = arg.clear_rb->color_val;
+
+   util_fill_rect(scene->cbufs[0].map,
+                  PIPE_FORMAT_B8G8R8A8_UNORM,
+                  scene->cbufs[0].stride,
+                  task->x,
+                  task->y,
+                  task->width,
+                  task->height,
+                  &uc);
+}
+
+/* Run the scanline version of the shader across the whole tile.
+ */
+static void
+lp_rast_linear_tile(struct lp_rasterizer_task *task,
+                   const union lp_rast_cmd_arg arg)
+{
+   const struct lp_rast_shader_inputs *inputs = arg.shade_tile;
+   const struct lp_rast_state *state;
+   struct lp_fragment_shader_variant *variant;
+   const struct lp_scene *scene = task->scene;
+
+   if (inputs->disable)
+      return;
+
+   state = task->state;
+   assert(state);
+   if (!state) {
+      return;
+   }
+   variant = state->variant;
+
+   if (variant->jit_linear_blit &&
+       inputs->is_blit)
+   {
+      if (variant->jit_linear_blit(state,
+                                   task->x,
+                                   task->y,
+                                   task->width,
+                                   task->height,
+                                   (const float (*)[4])GET_A0(inputs),
+                                   (const float (*)[4])GET_DADX(inputs),
+                                   (const float (*)[4])GET_DADY(inputs),
+                                   scene->cbufs[0].map,
+                                   scene->cbufs[0].stride))
+         return;
+   }
+
+
+   if (variant->jit_linear) {
+      if (variant->jit_linear(state,
+                              task->x,
+                              task->y,
+                              task->width,
+                              task->height,
+                              (const float (*)[4])GET_A0(inputs),
+                              (const float (*)[4])GET_DADX(inputs),
+                              (const float (*)[4])GET_DADY(inputs),
+                              scene->cbufs[0].map,
+                              scene->cbufs[0].stride))
+         return;
+   }
+
+   {
+      struct u_rect box;
+      box.x0 = task->x;
+      box.x1 = task->x + task->width - 1;
+      box.y0 = task->y;
+      box.y1 = task->y + task->height - 1;
+      lp_rast_linear_rect_fallback(task, inputs, &box);
+   }
+}
+
+
+/* Run the scanline version of the shader on a rectangle within the
+ * tile.
+ */
+static void
+lp_rast_linear_rect(struct lp_rasterizer_task *task,
+                    const union lp_rast_cmd_arg arg)
+{
+   const struct lp_scene *scene = task->scene;
+   const struct lp_rast_rectangle *rect = arg.rectangle;
+   const struct lp_rast_shader_inputs *inputs = &rect->inputs;
+   const struct lp_rast_state *state = task->state;
+   struct lp_fragment_shader_variant *variant = state->variant;
+   struct u_rect box;
+   int width, height;
+
+   if (inputs->disable)
+      return;
+
+   box.x0 = task->x;
+   box.y0 = task->y;
+   box.x1 = task->x + task->width - 1;
+   box.y1 = task->y + task->height - 1;
+
+   u_rect_find_intersection(&rect->box, &box);
+
+   width  = box.x1 - box.x0 + 1;
+   height = box.y1 - box.y0 + 1;
+
+   /* Note that blit primitives can end up in the non-full-tile path,
+    * the binner currently doesn't try to classify sub-tile
+    * primitives.  Can detect them here though.
+    */
+   if (variant->jit_linear_blit &&
+       inputs->is_blit)
+   {
+      if (variant->jit_linear_blit(state,
+                                   box.x0, box.y0,
+                                   width, height,
+                                   (const float (*)[4])GET_A0(inputs),
+                                   (const float (*)[4])GET_DADX(inputs),
+                                   (const float (*)[4])GET_DADY(inputs),
+                                   scene->cbufs[0].map,
+                                   scene->cbufs[0].stride))
+         return;
+   }
+
+   if (variant->jit_linear)
+   {
+      if (variant->jit_linear(state,
+                              box.x0, box.y0,
+                              width, height,
+                              (const float (*)[4])GET_A0(inputs),
+                              (const float (*)[4])GET_DADX(inputs),
+                              (const float (*)[4])GET_DADY(inputs),
+                              scene->cbufs[0].map,
+                              scene->cbufs[0].stride))
+         return;
+   }
+
+   lp_rast_linear_rect_fallback(task, inputs, &box);
+}
+
+
+static const lp_rast_cmd_func
+dispatch_linear[] = {
+   lp_rast_linear_clear,        /* clear_color */
+   NULL,                        /* clear_zstencil */
+   NULL,                        /* triangle_1 */
+   NULL,                        /* triangle_2 */
+   NULL,                        /* triangle_3 */
+   NULL,                        /* triangle_4 */
+   NULL,                        /* triangle_5 */
+   NULL,                        /* triangle_6 */
+   NULL,                        /* triangle_7 */
+   NULL,                        /* triangle_8 */
+   NULL,                        /* triangle_3_4 */
+   NULL,                        /* triangle_3_16 */
+   NULL,                        /* triangle_4_16 */
+   lp_rast_linear_tile,         /* shade_tile */
+   lp_rast_linear_tile,         /* shade_tile_opaque */
+   NULL,                        /* begin_query */
+   NULL,                        /* end_query */
+   lp_rast_set_state,           /* set_state */
+   NULL,                        /* lp_rast_triangle_32_1 */
+   NULL,                        /* lp_rast_triangle_32_2 */
+   NULL,                        /* lp_rast_triangle_32_3 */
+   NULL,                        /* lp_rast_triangle_32_4 */
+   NULL,                        /* lp_rast_triangle_32_5 */
+   NULL,                        /* lp_rast_triangle_32_6 */
+   NULL,                        /* lp_rast_triangle_32_7 */
+   NULL,                        /* lp_rast_triangle_32_8 */
+   NULL,                        /* lp_rast_triangle_32_3_4 */
+   NULL,                        /* lp_rast_triangle_32_3_16 */
+   NULL,                        /* lp_rast_triangle_32_4_16 */
+
+   NULL,                        /* lp_rast_triangle_ms_1 */
+   NULL,                        /* lp_rast_triangle_ms_2 */
+   NULL,                        /* lp_rast_triangle_ms_3 */
+   NULL,                        /* lp_rast_triangle_ms_4 */
+   NULL,                        /* lp_rast_triangle_ms_5 */
+   NULL,                        /* lp_rast_triangle_ms_6 */
+   NULL,                        /* lp_rast_triangle_ms_7 */
+   NULL,                        /* lp_rast_triangle_ms_8 */
+   NULL,                        /* lp_rast_triangle_ms_3_4 */
+   NULL,                        /* lp_rast_triangle_ms_3_16 */
+   NULL,                        /* lp_rast_triangle_ms_4_16 */
+
+   lp_rast_linear_rect,         /* rect */
+   lp_rast_linear_tile,         /* blit */
+};
+
+/* Assumptions for this path:
+ *   - Single color buffer, PIPE_FORMAT_B8G8R8A8_UNORM
+ *   - No depth buffer
+ *   - All primitives in bins are rect, tile, blit or clear.
+ *   - All shaders have a linear variant.
+ */
+void
+lp_linear_rasterize_bin(struct lp_rasterizer_task *task,
+                        const struct cmd_bin *bin)
+{
+   const struct cmd_block *block;
+   unsigned k;
+
+   STATIC_ASSERT(ARRAY_SIZE(dispatch_linear) == LP_RAST_OP_MAX);
+
+   if (0) debug_printf("%s\n", __FUNCTION__);
+
+   for (block = bin->head; block; block = block->next) {
+      for (k = 0; k < block->count; k++) {
+         assert(dispatch_linear[block->cmd[k]]);
+         dispatch_linear[block->cmd[k]]( task, block->arg[k] );
+      }
+   }
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_linear_fallback.c b/src/gallium/drivers/llvmpipe/lp_rast_linear_fallback.c
new file mode 100644
index 00000000000..ede03993475
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_rast_linear_fallback.c
@@ -0,0 +1,303 @@
+/**************************************************************************
+ *
+ * Copyright 2007-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+/*
+ * Rasterization for binned rectangles within a tile
+ */
+
+#include <limits.h>
+#include "util/u_math.h"
+#include "lp_debug.h"
+#include "lp_perf.h"
+#include "lp_rast_priv.h"
+
+
+/* Our 16-pixel stamps are layed out as:
+ *
+ *    0  1  2  3
+ *    4  5  6  7
+ *    8  9  10 11
+ *    12 13 14 15
+ *
+ * Define bitmasks for each row and column in this layout:
+ */
+#define COLUMN0 ((1<<0)|(1<<4)|(1<<8) |(1<<12))
+#define COLUMN1 ((1<<1)|(1<<5)|(1<<9) |(1<<13))
+#define COLUMN2 ((1<<2)|(1<<6)|(1<<10)|(1<<14))
+#define COLUMN3 ((1<<3)|(1<<7)|(1<<11)|(1<<15))
+
+#define ROW0 ((1<<0) |(1<<1) |(1<<2) |(1<<3))
+#define ROW1 ((1<<4) |(1<<5) |(1<<6) |(1<<7))
+#define ROW2 ((1<<8) |(1<<9) |(1<<10)|(1<<11))
+#define ROW3 ((1<<12)|(1<<13)|(1<<14)|(1<<15))
+
+#define STAMP_SIZE 4
+
+static unsigned left_mask_tab[STAMP_SIZE] = {
+   COLUMN0 | COLUMN1 | COLUMN2 | COLUMN3,
+   COLUMN1 | COLUMN2 | COLUMN3,
+   COLUMN2 | COLUMN3,
+   COLUMN3,
+};
+
+static unsigned right_mask_tab[STAMP_SIZE] = {
+   COLUMN0,
+   COLUMN0 | COLUMN1,
+   COLUMN0 | COLUMN1 | COLUMN2,
+   COLUMN0 | COLUMN1 | COLUMN2 | COLUMN3,
+};
+
+static unsigned top_mask_tab[STAMP_SIZE] = {
+   ROW0 | ROW1 | ROW2 | ROW3,
+   ROW1 | ROW2 | ROW3,
+   ROW2 | ROW3,
+   ROW3,
+};
+
+static unsigned bottom_mask_tab[STAMP_SIZE] = {
+   ROW0,
+   ROW0 | ROW1,
+   ROW0 | ROW1 | ROW2,
+   ROW0 | ROW1 | ROW2 | ROW3,
+};
+
+
+/**
+ * Shade all pixels in a 4x4 block.  The fragment code omits the
+ * triangle in/out tests.
+ * \param x, y location of 4x4 block in window coords
+ */
+static void
+shade_quads_all( struct lp_rasterizer_task *task,
+                 const struct lp_rast_shader_inputs *inputs,
+                 unsigned x, unsigned y )
+{
+   const struct lp_scene *scene = task->scene;
+   const struct lp_rast_state *state = task->state;
+   struct lp_fragment_shader_variant *variant = state->variant;
+   uint8_t *color = scene->cbufs[0].map;
+   unsigned stride = scene->cbufs[0].stride;
+   uint8_t *cbufs[1];
+   unsigned strides[1];
+
+   color += x * 4;
+   color += y * stride;
+   cbufs[0] = color;
+   strides[0] = stride;
+
+   assert(!variant->key.depth.enabled);
+
+   /* run shader on 4x4 block */
+   BEGIN_JIT_CALL(state, task);
+   variant->jit_function[RAST_WHOLE]( &state->jit_context,
+                                      x, y,
+                                      1,
+                                      (const float (*)[4])GET_A0(inputs),
+                                      (const float (*)[4])GET_DADX(inputs),
+                                      (const float (*)[4])GET_DADY(inputs),
+                                      cbufs,
+                                      NULL,
+                                      0xffff,
+                                      &task->thread_data,
+                                      strides, 0, 0, 0 );
+   END_JIT_CALL();
+}
+
+static void
+shade_quads_mask(struct lp_rasterizer_task *task,
+                 const struct lp_rast_shader_inputs *inputs,
+                 unsigned x, unsigned y,
+                 unsigned mask)
+{
+   const struct lp_rast_state *state = task->state;
+   struct lp_fragment_shader_variant *variant = state->variant;
+   const struct lp_scene *scene = task->scene;
+   uint8_t *color = scene->cbufs[0].map;
+   unsigned stride = scene->cbufs[0].stride;
+   uint8_t *cbufs[1];
+   unsigned strides[1];
+
+   color += x * 4;
+   color += y * stride;
+   cbufs[0] = color;
+   strides[0] = stride;
+
+   assert(!variant->key.depth.enabled);
+
+   /* Propagate non-interpolated raster state */
+   task->thread_data.raster_state.viewport_index = inputs->viewport_index;
+
+   /* run shader on 4x4 block */
+   BEGIN_JIT_CALL(state, task);
+   variant->jit_function[RAST_EDGE_TEST](&state->jit_context,
+                                         x, y,
+                                         1,
+                                         (const float (*)[4])GET_A0(inputs),
+                                         (const float (*)[4])GET_DADX(inputs),
+                                         (const float (*)[4])GET_DADY(inputs),
+                                         cbufs,
+                                         NULL,
+                                         mask,
+                                         &task->thread_data,
+                                         strides, 0, 0, 0);
+   END_JIT_CALL();
+}
+
+/* Shade a 4x4 stamp completely within the rectangle.
+ */
+static inline void
+full(struct lp_rasterizer_task *task,
+     const struct lp_rast_shader_inputs *inputs,
+     unsigned ix, unsigned iy)
+{
+   shade_quads_all(task,
+                   inputs,
+                   ix * STAMP_SIZE,
+                   iy * STAMP_SIZE);
+}
+
+/* Shade a 4x4 stamp which may be partially outside the rectangle,
+ * according to the mask parameter.
+ */
+static inline void
+partial(struct lp_rasterizer_task *task,
+        const struct lp_rast_shader_inputs *inputs,
+        unsigned ix, unsigned iy,
+        unsigned mask)
+{
+   /* Unfortunately we can end up generating full blocks on this path,
+    * need to catch them.
+    */
+   if (mask == 0xffff)
+      full(task, inputs, ix, iy);
+   else {
+      assert(mask);
+      shade_quads_mask(task,
+                       inputs,
+                       ix * STAMP_SIZE,
+                       iy * STAMP_SIZE,
+                       mask);
+   }
+}
+
+
+/**
+ * Run the full SoA shader.
+ */
+void
+lp_rast_linear_rect_fallback(struct lp_rasterizer_task *task,
+                             const struct lp_rast_shader_inputs *inputs,
+                             const struct u_rect *box)
+{
+   unsigned ix0, ix1, iy0, iy1;
+   unsigned left_mask;
+   unsigned right_mask;
+   unsigned top_mask;
+   unsigned bottom_mask;
+   unsigned i,j;
+
+   /* The interior of the rectangle (if there is one) will be
+    * rasterized as full 4x4 stamps.
+    *
+    * At each edge of the rectangle, however, there will be a fringe
+    * of partial blocks where the edge lands somewhere in the middle
+    * of a 4x4-pixel stamp.
+    *
+    * For each edge, precalculate a mask of the pixels inside that
+    * edge for the first 4x4-pixel stamp.
+    *
+    * Note that at the corners, and for narrow rectangles, an
+    * individual stamp may have two or more edges active.  We'll deal
+    * with that below by combining these masks as appropriate.
+    */
+   left_mask   = left_mask_tab   [box->x0 & (STAMP_SIZE - 1)];
+   right_mask  = right_mask_tab  [box->x1 & (STAMP_SIZE - 1)];
+   top_mask    = top_mask_tab    [box->y0 & (STAMP_SIZE - 1)];
+   bottom_mask = bottom_mask_tab [box->y1 & (STAMP_SIZE - 1)];
+
+   ix0 = box->x0 / STAMP_SIZE;
+   ix1 = box->x1 / STAMP_SIZE;
+   iy0 = box->y0 / STAMP_SIZE;
+   iy1 = box->y1 / STAMP_SIZE;
+
+   /* Various special cases.
+    */
+   if (ix0 == ix1 && iy0 == iy1) {
+      /* Rectangle is contained within a single 4x4-pixel stamp:
+       */
+      partial(task, inputs, ix0, iy0,
+              (left_mask & right_mask &
+               top_mask & bottom_mask));
+   }
+   else if (ix0 == ix1) {
+      /* Left and right edges fall on the same 4-pixel-wide column:
+       */
+      unsigned mask = left_mask & right_mask;
+      partial(task, inputs, ix0, iy0, mask & top_mask);
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(task, inputs, ix0, i, mask);
+      partial(task, inputs, ix0, iy1, mask & bottom_mask);
+   }
+   else if (iy0 == iy1) {
+      /* Top and bottom edges fall on the same 4-pixel-wide row:
+       */
+      unsigned mask = top_mask & bottom_mask;
+      partial(task, inputs, ix0, iy0, mask & left_mask);
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(task, inputs, i, iy0, mask);
+      partial(task, inputs, ix1, iy0, mask & right_mask);
+   }
+   else {
+      /* Each pair of edges falls in a separate 4-pixel-wide
+       * row/column.
+       */
+      partial(task, inputs, ix0, iy0, left_mask  & top_mask);
+      partial(task, inputs, ix0, iy1, left_mask  & bottom_mask);
+      partial(task, inputs, ix1, iy0, right_mask & top_mask);
+      partial(task, inputs, ix1, iy1, right_mask & bottom_mask);
+
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(task, inputs, i, iy0, top_mask);
+
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(task, inputs, i, iy1, bottom_mask);
+
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(task, inputs, ix0, i, left_mask);
+
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(task, inputs, ix1, i, right_mask);
+
+      /* Full interior blocks
+       */
+      for (j = iy0 + 1; j < iy1; j++) {
+         for (i = ix0 + 1; i < ix1; i++) {
+            full(task, inputs, i, j);
+         }
+      }
+   }
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_priv.h b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
index c8154348ef9..c4da9cca2ff 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_priv.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_priv.h
@@ -176,6 +176,7 @@ lp_rast_get_color_block_pointer(struct lp_rasterizer_task *task,
    color = task->color_tiles[buf] + pixel_offset;
 
    if (layer) {
+      assert(layer <= task->scene->fb_max_layer);
       color += layer * task->scene->cbufs[buf].layer_stride;
    }
 
@@ -347,6 +348,10 @@ void lp_rast_triangle_32_3_16( struct lp_rasterizer_task *,
 void lp_rast_triangle_32_4_16( struct lp_rasterizer_task *, 
                             const union lp_rast_cmd_arg );
 
+
+void lp_rast_rectangle( struct lp_rasterizer_task *, 
+                        const union lp_rast_cmd_arg );
+
 void lp_rast_triangle_ms_1( struct lp_rasterizer_task *,
                          const union lp_rast_cmd_arg );
 void lp_rast_triangle_ms_2( struct lp_rasterizer_task *,
@@ -406,4 +411,13 @@ lp_rast_set_state(struct lp_rasterizer_task *task,
 void
 lp_debug_bin( const struct cmd_bin *bin, int x, int y );
 
+void
+lp_linear_rasterize_bin(struct lp_rasterizer_task *task,
+                        const struct cmd_bin *bin);
+
+void
+lp_rast_linear_rect_fallback(struct lp_rasterizer_task *task,
+                             const struct lp_rast_shader_inputs *inputs,
+                             const struct u_rect *box);
+
 #endif
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_rect.c b/src/gallium/drivers/llvmpipe/lp_rast_rect.c
new file mode 100644
index 00000000000..77b122bddfa
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_rast_rect.c
@@ -0,0 +1,255 @@
+/**************************************************************************
+ *
+ * Copyright 2007-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+/*
+ * Rasterization for binned rectangles within a tile
+ */
+
+#include <limits.h>
+#include "util/u_math.h"
+#include "lp_debug.h"
+#include "lp_perf.h"
+#include "lp_rast_priv.h"
+
+/* Our 16-pixel stamps are layed out as:
+ *
+ *    0  1  2  3
+ *    4  5  6  7
+ *    8  9  10 11
+ *    12 13 14 15
+ *
+ * Define bitmasks for each row and column in this layout:
+ */
+#define COLUMN0 ((1<<0)|(1<<4)|(1<<8) |(1<<12))
+#define COLUMN1 ((1<<1)|(1<<5)|(1<<9) |(1<<13))
+#define COLUMN2 ((1<<2)|(1<<6)|(1<<10)|(1<<14))
+#define COLUMN3 ((1<<3)|(1<<7)|(1<<11)|(1<<15))
+
+#define ROW0 ((1<<0) |(1<<1) |(1<<2) |(1<<3))
+#define ROW1 ((1<<4) |(1<<5) |(1<<6) |(1<<7))
+#define ROW2 ((1<<8) |(1<<9) |(1<<10)|(1<<11))
+#define ROW3 ((1<<12)|(1<<13)|(1<<14)|(1<<15))
+
+#define STAMP_SIZE 4
+
+static unsigned left_mask_tab[STAMP_SIZE] = {
+   COLUMN0 | COLUMN1 | COLUMN2 | COLUMN3,
+   COLUMN1 | COLUMN2 | COLUMN3,
+   COLUMN2 | COLUMN3,
+   COLUMN3,
+};
+
+static unsigned right_mask_tab[STAMP_SIZE] = {
+   COLUMN0,
+   COLUMN0 | COLUMN1,
+   COLUMN0 | COLUMN1 | COLUMN2,
+   COLUMN0 | COLUMN1 | COLUMN2 | COLUMN3,
+};
+
+static unsigned top_mask_tab[STAMP_SIZE] = {
+   ROW0 | ROW1 | ROW2 | ROW3,
+   ROW1 | ROW2 | ROW3,
+   ROW2 | ROW3,
+   ROW3,
+};
+
+static unsigned bottom_mask_tab[STAMP_SIZE] = {
+   ROW0,
+   ROW0 | ROW1,
+   ROW0 | ROW1 | ROW2,
+   ROW0 | ROW1 | ROW2 | ROW3,
+};
+
+static inline void
+full(struct lp_rasterizer_task *task,
+     const struct lp_rast_rectangle *rect,
+     unsigned ix, unsigned iy)
+{
+   LP_COUNT(nr_rect_fully_covered_4);
+   lp_rast_shade_quads_all(task,
+                           &rect->inputs,
+                           task->x + ix * STAMP_SIZE,
+                           task->y + iy * STAMP_SIZE);
+}
+
+static inline void
+partial(struct lp_rasterizer_task *task,
+        const struct lp_rast_rectangle *rect,
+        unsigned ix, unsigned iy,
+        unsigned mask)
+{
+   /* Unfortunately we can end up generating full blocks on this path,
+    * need to catch them.
+    */
+   if (mask == 0xffff)
+      full(task, rect, ix, iy);
+   else {
+      assert(mask);
+      LP_COUNT(nr_rect_partially_covered_4);
+      lp_rast_shade_quads_mask(task,
+                               &rect->inputs,
+                               task->x + ix * STAMP_SIZE,
+                               task->y + iy * STAMP_SIZE,
+                               mask);
+   }
+}
+
+
+static inline void
+intersect_rect_and_tile(struct lp_rasterizer_task *task,
+                        const struct lp_rast_rectangle *rect,
+                        struct u_rect *box)
+{
+   box->x0 = task->x;
+   box->y0 = task->y;
+   box->x1 = task->x + TILE_SIZE - 1;
+   box->y1 = task->y + TILE_SIZE - 1;
+
+   assert(u_rect_test_intersection(&rect->box, box));
+
+   u_rect_find_intersection(&rect->box, box);
+
+   box->x0 -= task->x;
+   box->x1 -= task->x;
+   box->y0 -= task->y;
+   box->y1 -= task->y;
+}
+
+
+/**
+ * Scan the tile in chunks and figure out which pixels to rasterize
+ * for this rectangle.
+ */
+void
+lp_rast_rectangle(struct lp_rasterizer_task *task,
+                  const union lp_rast_cmd_arg arg)
+{
+   const struct lp_rast_rectangle *rect = arg.rectangle;
+
+   struct u_rect box;
+   unsigned ix0, ix1, iy0, iy1;
+   unsigned left_mask;
+   unsigned right_mask;
+   unsigned top_mask;
+   unsigned bottom_mask;
+   unsigned i,j;
+
+   /* Check for "disabled" rectangles generated in out-of-memory
+    * conditions.
+    */
+   if (rect->inputs.disable) {
+      /* This command was partially binned and has been disabled */
+      return;
+   }
+
+   /* Intersect the rectangle with this tile.
+    */
+   intersect_rect_and_tile(task, rect, &box);
+
+   /* The interior of the rectangle (if there is one) will be
+    * rasterized as full 4x4 stamps.
+    *
+    * At each edge of the rectangle, however, there will be a fringe
+    * of partial blocks where the edge lands somewhere in the middle
+    * of a 4-pixel stamp.
+    *
+    * For each edge, precalculate a mask of the pixels inside that
+    * edge for the first 4-pixel stamp.
+    *
+    * Note that at the corners, and for narrow rectangles, an
+    * individual stamp may have two or more edges active.  We'll deal
+    * with that below by combining these masks as appropriate.
+    */
+   left_mask   = left_mask_tab   [box.x0 & (STAMP_SIZE - 1)];
+   right_mask  = right_mask_tab  [box.x1 & (STAMP_SIZE - 1)];
+   top_mask    = top_mask_tab    [box.y0 & (STAMP_SIZE - 1)];
+   bottom_mask = bottom_mask_tab [box.y1 & (STAMP_SIZE - 1)];
+
+   ix0 = box.x0 / STAMP_SIZE;
+   ix1 = box.x1 / STAMP_SIZE;
+   iy0 = box.y0 / STAMP_SIZE;
+   iy1 = box.y1 / STAMP_SIZE;
+
+   /* Various special cases.
+    */
+   if (ix0 == ix1 && iy0 == iy1) {
+      /* Rectangle is contained within a single 4x4 stamp:
+       */
+      partial(task, rect, ix0, iy0,
+              (left_mask & right_mask &
+               top_mask & bottom_mask));
+   }
+   else if (ix0 == ix1) {
+      /* Left and right edges fall on the same 4-pixel-wide column:
+       */
+      unsigned mask = left_mask & right_mask;
+      partial(task, rect, ix0, iy0, mask & top_mask);
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(task, rect, ix0, i, mask);
+      partial(task, rect, ix0, iy1, mask & bottom_mask);
+   }
+   else if (iy0 == iy1) {
+      /* Top and bottom edges fall on the same 4-pixel-wide row:
+       */
+      unsigned mask = top_mask & bottom_mask;
+      partial(task, rect, ix0, iy0, mask & left_mask);
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(task, rect, i, iy0, mask);
+      partial(task, rect, ix1, iy0, mask & right_mask);
+   }
+   else {
+      /* Each pair of edges falls in a separate 4-pixel-wide
+       * row/column.
+       */
+      partial(task, rect, ix0, iy0, left_mask  & top_mask);
+      partial(task, rect, ix0, iy1, left_mask  & bottom_mask);
+      partial(task, rect, ix1, iy0, right_mask & top_mask);
+      partial(task, rect, ix1, iy1, right_mask & bottom_mask);
+
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(task, rect, i, iy0, top_mask);
+
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(task, rect, i, iy1, bottom_mask);
+
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(task, rect, ix0, i, left_mask);
+
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(task, rect, ix1, i, right_mask);
+
+      /* Full interior blocks
+       */
+      for (j = iy0 + 1; j < iy1; j++) {
+         for (i = ix0 + 1; i < ix1; i++) {
+            full(task, rect, i, j);
+         }
+      }
+   }
+}
+
+
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri.c b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
index 01fbd2d9d74..5874331d463 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri.c
@@ -261,6 +261,31 @@ sign_bits4(const __m128i *cstep, int cdiff)
    return _mm_movemask_epi8(result);
 }
 
+#define COLUMN0 ((1<<0)|(1<<4)|(1<<8) |(1<<12))
+#define COLUMN1 ((1<<1)|(1<<5)|(1<<9) |(1<<13))
+#define COLUMN2 ((1<<2)|(1<<6)|(1<<10)|(1<<14))
+#define COLUMN3 ((1<<3)|(1<<7)|(1<<11)|(1<<15))
+
+#define ROW0 ((1<<0) |(1<<1) |(1<<2) |(1<<3))
+#define ROW1 ((1<<4) |(1<<5) |(1<<6) |(1<<7))
+#define ROW2 ((1<<8) |(1<<9) |(1<<10)|(1<<11))
+#define ROW3 ((1<<12)|(1<<13)|(1<<14)|(1<<15))
+
+#define STAMP_SIZE 4
+static unsigned bottom_mask_tab[STAMP_SIZE] = { 
+   ROW3,
+   ROW3 | ROW2,
+   ROW3 | ROW2 | ROW1,
+   ROW3 | ROW2 | ROW1 | ROW0,
+};
+
+static unsigned right_mask_tab[STAMP_SIZE] = { 
+   COLUMN3,
+   COLUMN3 | COLUMN2,
+   COLUMN3 | COLUMN2 | COLUMN1,
+   COLUMN3 | COLUMN2 | COLUMN1 | COLUMN0,
+};
+
 
 #define NR_PLANES 3
 
diff --git a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
index 47280e591a6..bced19c2252 100644
--- a/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
+++ b/src/gallium/drivers/llvmpipe/lp_rast_tri_tmp.h
@@ -363,6 +363,16 @@ TRI_16(struct lp_rasterizer_task *task,
 
    outmask = 0;                 /* outside one or more trivial reject planes */
    
+   if (x + 12 >= 64) {
+      int i = ((x + 12) - 64) / 4;
+      outmask |= right_mask_tab[i];
+   }
+
+   if (y + 12 >= 64) {
+      int i = ((y + 12) - 64) / 4;
+      outmask |= bottom_mask_tab[i];
+   }
+
    x += task->x;
    y += task->y;
 
diff --git a/src/gallium/drivers/llvmpipe/lp_scene.c b/src/gallium/drivers/llvmpipe/lp_scene.c
index 0888341e049..49db1832ead 100644
--- a/src/gallium/drivers/llvmpipe/lp_scene.c
+++ b/src/gallium/drivers/llvmpipe/lp_scene.c
@@ -68,9 +68,7 @@ lp_scene_create( struct pipe_context *pipe )
       return NULL;
 
    scene->pipe = pipe;
-
-   scene->data.head =
-      CALLOC_STRUCT(data_block);
+   scene->data.head = &scene->data.first;
 
    (void) mtx_init(&scene->mutex, mtx_plain);
 
@@ -101,8 +99,7 @@ lp_scene_destroy(struct lp_scene *scene)
 {
    lp_fence_reference(&scene->fence, NULL);
    mtx_destroy(&scene->mutex);
-   assert(scene->data.head->next == NULL);
-   FREE(scene->data.head);
+   assert(scene->data.head == &scene->data.first);
    FREE(scene);
 }
 
@@ -129,8 +126,8 @@ lp_scene_is_empty(struct lp_scene *scene )
 
 
 /* Returns true if there has ever been a failed allocation attempt in
- * this scene.  Used in triangle emit to avoid having to check success
- * at each bin.
+ * this scene.  Used in triangle/rectangle emit to avoid having to
+ * check success at each bin.
  */
 boolean
 lp_scene_is_oom(struct lp_scene *scene)
@@ -222,7 +219,7 @@ lp_scene_begin_rasterization(struct lp_scene *scene)
 void
 lp_scene_end_rasterization(struct lp_scene *scene )
 {
-   int i, j;
+   int i;
 
    /* Unmap color buffers */
    for (i = 0; i < scene->fb.nr_cbufs; i++) {
@@ -248,19 +245,7 @@ lp_scene_end_rasterization(struct lp_scene *scene )
 
    /* Reset all command lists:
     */
-   for (i = 0; i < scene->tiles_x; i++) {
-      for (j = 0; j < scene->tiles_y; j++) {
-         struct cmd_bin *bin = lp_scene_get_bin(scene, i, j);
-         bin->head = NULL;
-         bin->tail = NULL;
-         bin->last_state = NULL;
-      }
-   }
-
-   /* If there are any bins which weren't cleared by the loop above,
-    * they will be caught (on debug builds at least) by this assert:
-    */
-   assert(lp_scene_is_empty(scene));
+   memset(scene->tile, 0, sizeof scene->tile);
 
    /* Decrement texture ref counts
     */
@@ -310,13 +295,14 @@ lp_scene_end_rasterization(struct lp_scene *scene )
       struct data_block_list *list = &scene->data;
       struct data_block *block, *tmp;
 
-      for (block = list->head->next; block; block = tmp) {
+      for (block = list->head; block; block = tmp) {
          tmp = block->next;
-	 FREE(block);
+         if (block != &list->first)
+            FREE(block);
       }
 
+      list->head = &list->first;
       list->head->next = NULL;
-      list->head->used = 0;
    }
 
    lp_fence_reference(&scene->fence, NULL);
diff --git a/src/gallium/drivers/llvmpipe/lp_scene.h b/src/gallium/drivers/llvmpipe/lp_scene.h
index 4ee2e5a8bc3..a089e6a4983 100644
--- a/src/gallium/drivers/llvmpipe/lp_scene.h
+++ b/src/gallium/drivers/llvmpipe/lp_scene.h
@@ -54,7 +54,8 @@ struct lp_rast_state;
  */
 #define CMD_BLOCK_MAX 29
 
-/* Bytes per data block.
+/* Bytes per data block.  This effectively limits the maximum constant buffer
+ * size.
  */
 #define DATA_BLOCK_SIZE (64 * 1024)
 
@@ -181,6 +182,8 @@ struct lp_scene {
    unsigned resource_reference_size;
 
    boolean alloc_failed;
+   boolean permit_linear_rasterizer;
+
    /**
     * Number of active tiles in each dimension.
     * This basically the framebuffer size divided by tile size
@@ -236,7 +239,7 @@ lp_scene_alloc( struct lp_scene *scene, unsigned size)
 
    if (LP_DEBUG & DEBUG_MEM)
       debug_printf("alloc %u block %u/%u tot %u/%u\n",
-		   size, block->used, DATA_BLOCK_SIZE,
+		   size, block->used, (unsigned)DATA_BLOCK_SIZE,
 		   scene->scene_size, LP_SCENE_MAX_SIZE);
 
    if (block->used + size > DATA_BLOCK_SIZE) {
@@ -270,7 +273,7 @@ lp_scene_alloc_aligned( struct lp_scene *scene, unsigned size,
    if (LP_DEBUG & DEBUG_MEM)
       debug_printf("alloc %u block %u/%u tot %u/%u\n",
 		   size + alignment - 1,
-		   block->used, DATA_BLOCK_SIZE,
+		   block->used, (unsigned)DATA_BLOCK_SIZE,
 		   scene->scene_size, LP_SCENE_MAX_SIZE);
        
    if (block->used + size + alignment - 1 > DATA_BLOCK_SIZE) {
@@ -288,17 +291,6 @@ lp_scene_alloc_aligned( struct lp_scene *scene, unsigned size,
 }
 
 
-/* Put back data if we decide not to use it, eg. culled triangles.
- */
-static inline void
-lp_scene_putback_data( struct lp_scene *scene, unsigned size)
-{
-   struct data_block_list *list = &scene->data;
-   assert(list->head && list->head->used >= size);
-   list->head->used -= size;
-}
-
-
 /** Return pointer to a particular tile's bin. */
 static inline struct cmd_bin *
 lp_scene_get_bin(struct lp_scene *scene, unsigned x, unsigned y)
diff --git a/src/gallium/drivers/llvmpipe/lp_screen.c b/src/gallium/drivers/llvmpipe/lp_screen.c
index e7b5f2e50c6..f57e3f95461 100644
--- a/src/gallium/drivers/llvmpipe/lp_screen.c
+++ b/src/gallium/drivers/llvmpipe/lp_screen.c
@@ -70,11 +70,15 @@ static const struct debug_named_value lp_debug_flags[] = {
    { "counters", DEBUG_COUNTERS, NULL },
    { "scene", DEBUG_SCENE, NULL },
    { "fence", DEBUG_FENCE, NULL },
+   { "no_fastpath", DEBUG_NO_FASTPATH, NULL },
+   { "linear", DEBUG_LINEAR, NULL },
+   { "linear2", DEBUG_LINEAR2, NULL },
    { "mem", DEBUG_MEM, NULL },
    { "fs", DEBUG_FS, NULL },
    { "cs", DEBUG_CS, NULL },
    { "tgsi_ir", DEBUG_TGSI_IR, NULL },
    { "cache_stats", DEBUG_CACHE_STATS, NULL },
+   { "accurate_a0", DEBUG_ACCURATE_A0 },
    DEBUG_NAMED_VALUE_END
 };
 #endif
@@ -89,6 +93,8 @@ static const struct debug_named_value lp_perf_flags[] = {
    { "no_blend",       PERF_NO_BLEND, NULL },
    { "no_depth",       PERF_NO_DEPTH, NULL },
    { "no_alphatest",   PERF_NO_ALPHATEST, NULL },
+   { "no_rast_linear", PERF_NO_RAST_LINEAR, NULL },
+   { "no_shade",       PERF_NO_SHADE, NULL },
    DEBUG_NAMED_VALUE_END
 };
 
@@ -1026,7 +1032,7 @@ llvmpipe_create_screen(struct sw_winsys *winsys)
    screen->use_tgsi = (LP_DEBUG & DEBUG_TGSI_IR);
    screen->num_threads = util_get_cpu_caps()->nr_cpus > 1 ? util_get_cpu_caps()->nr_cpus : 0;
 #ifdef EMBEDDED_DEVICE
-   screen->num_threads = 0;
+   screen->num_threads = MIN2(screen->num_threads, 2);
 #endif
    screen->num_threads = debug_get_num_option("LP_NUM_THREADS", screen->num_threads);
    screen->num_threads = MIN2(screen->num_threads, LP_MAX_THREADS);
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.c b/src/gallium/drivers/llvmpipe/lp_setup.c
index 8bd0893465c..4c0c0ad50e9 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup.c
@@ -39,6 +39,7 @@
 #include "util/u_inlines.h"
 #include "util/u_memory.h"
 #include "util/u_pack_color.h"
+#include "util/u_cpu_detect.h"
 #include "util/u_viewport.h"
 #include "draw/draw_pipe.h"
 #include "util/os_time.h"
@@ -53,6 +54,7 @@
 #include "lp_setup_context.h"
 #include "lp_screen.h"
 #include "lp_state.h"
+#include "lp_jit.h"
 #include "frontend/sw_winsys.h"
 
 #include "draw/draw_context.h"
@@ -84,6 +86,7 @@ lp_setup_get_empty_scene(struct lp_setup_context *setup)
 
    lp_scene_begin_binning(setup->scene, &setup->fb);
 
+   setup->scene->permit_linear_rasterizer = setup->permit_linear_rasterizer;
 }
 
 
@@ -98,6 +101,20 @@ first_triangle( struct lp_setup_context *setup,
    setup->triangle( setup, v0, v1, v2 );
 }
 
+static boolean
+first_rectangle( struct lp_setup_context *setup,
+                 const float (*v0)[4],
+                 const float (*v1)[4],
+                 const float (*v2)[4],
+                 const float (*v3)[4],
+                 const float (*v4)[4],
+                 const float (*v5)[4])
+{
+   assert(setup->state == SETUP_ACTIVE);
+   lp_setup_choose_rect( setup );
+   return setup->rect( setup, v0, v1, v2, v3, v4, v5 );
+}
+
 static void
 first_line( struct lp_setup_context *setup,
 	    const float (*v0)[4],
@@ -117,7 +134,8 @@ first_point( struct lp_setup_context *setup,
    setup->point( setup, v0 );
 }
 
-void lp_setup_reset( struct lp_setup_context *setup )
+void
+lp_setup_reset( struct lp_setup_context *setup )
 {
    unsigned i;
 
@@ -145,6 +163,7 @@ void lp_setup_reset( struct lp_setup_context *setup )
    setup->line = first_line;
    setup->point = first_point;
    setup->triangle = first_triangle;
+   setup->rect = first_rectangle;
 }
 
 
@@ -576,6 +595,7 @@ lp_setup_set_triangle_state( struct lp_setup_context *setup,
    setup->ccw_is_frontface = ccw_is_frontface;
    setup->cullmode = cull_mode;
    setup->triangle = first_triangle;
+   setup->rect = first_rectangle;
    setup->multisample = multisample;
    setup->pixel_offset = half_pixel_center ? 0.5f : 0.0f;
    setup->bottom_edge_rule = bottom_edge_rule;
@@ -600,6 +620,7 @@ lp_setup_set_line_state( struct lp_setup_context *setup,
 void 
 lp_setup_set_point_state( struct lp_setup_context *setup,
                           float point_size,
+                          boolean point_tri_clip,
                           boolean point_size_per_vertex,
                           uint sprite_coord_enable,
                           uint sprite_coord_origin,
@@ -610,6 +631,7 @@ lp_setup_set_point_state( struct lp_setup_context *setup,
    setup->point_size = point_size;
    setup->sprite_coord_enable = sprite_coord_enable;
    setup->sprite_coord_origin = sprite_coord_origin;
+   setup->point_tri_clip = point_tri_clip;
    setup->point_size_per_vertex = point_size_per_vertex;
    setup->legacy_points = !point_quad_rasterization;
 }
@@ -833,6 +855,7 @@ lp_setup_set_rasterizer_discard(struct lp_setup_context *setup,
       setup->line = first_line;
       setup->point = first_point;
       setup->triangle = first_triangle;
+      setup->rect = first_rectangle;
    }
 }
 
@@ -846,6 +869,24 @@ lp_setup_set_vertex_info(struct lp_setup_context *setup,
 }
 
 
+void 
+lp_setup_set_linear_mode( struct lp_setup_context *setup,
+                          boolean mode )
+{
+   /* The linear rasterizer requires sse2 both at compile and runtime,
+    * in particular for the code in lp_rast_linear_fallback.c.  This
+    * is more than ten-year-old technology, so it's a reasonable
+    * baseline.
+    */
+#if defined(PIPE_ARCH_SSE)
+   setup->permit_linear_rasterizer = (mode &&
+                                      util_get_cpu_caps()->has_sse2);
+#else
+   setup->permit_linear_rasterizer = FALSE;
+#endif
+}
+
+
 /**
  * Called during state validation when LP_NEW_VIEWPORT is set.
  */
@@ -855,6 +896,7 @@ lp_setup_set_viewports(struct lp_setup_context *setup,
                        const struct pipe_viewport_state *viewports)
 {
    struct llvmpipe_context *lp = llvmpipe_context(setup->pipe);
+   float half_height, x0, y0;
    unsigned i;
 
    LP_DBG(DEBUG_SETUP, "%s\n", __FUNCTION__);
@@ -862,6 +904,26 @@ lp_setup_set_viewports(struct lp_setup_context *setup,
    assert(num_viewports <= PIPE_MAX_VIEWPORTS);
    assert(viewports);
 
+   /*
+    * Linear rasterizer path for scissor/viewport intersection.
+    *
+    * Calculate "scissor" rect from the (first) viewport.
+    * Just like stored scissor rects need inclusive coords.
+    * For rounding, assume half pixel center (d3d9 should not end up
+    * with fractional viewports) - quite obviously for msaa we'd need
+    * fractional values here (and elsewhere for the point bounding box).
+    *
+    * See: lp_setup.c::try_update_scene_state
+    */
+   half_height = fabsf(viewports[0].scale[1]);
+   x0 = viewports[0].translate[0] - viewports[0].scale[0];
+   y0 = viewports[0].translate[1] - half_height;
+   setup->vpwh.x0 = (int)(x0 + 0.5f);
+   setup->vpwh.x1 = (int)(viewports[0].scale[0] * 2.0f + x0 - 0.5f);
+   setup->vpwh.y0 = (int)(y0 + 0.5f);
+   setup->vpwh.y1 = (int)(half_height * 2.0f + y0 - 0.5f);
+   setup->dirty |= LP_SETUP_NEW_SCISSOR;
+
    /*
     * For use in lp_state_fs.c, propagate the viewport values for all viewports.
     */
@@ -882,7 +944,7 @@ lp_setup_set_viewports(struct lp_setup_context *setup,
 
 
 /**
- * Called during state validation when LP_NEW_SAMPLER_VIEW is set.
+ * Called directly by llvmpipe_set_sampler_views
  */
 void
 lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
@@ -1032,7 +1094,6 @@ lp_setup_set_fragment_sampler_views(struct lp_setup_context *setup,
    setup->dirty |= LP_SETUP_NEW_FS;
 }
 
-
 /**
  * Called during state validation when LP_NEW_SAMPLER is set.
  */
@@ -1066,6 +1127,8 @@ lp_setup_set_fragment_sampler_state(struct lp_setup_context *setup,
 }
 
 
+
+
 /**
  * Is the given texture referenced by any scene?
  * Note: we have to check all scenes including any scenes currently
@@ -1320,6 +1383,8 @@ try_update_scene_state( struct lp_setup_context *setup )
 
    if (setup->dirty & LP_SETUP_NEW_SCISSOR) {
       unsigned i;
+      setup->scissor_or_vp_clip = setup->scissor_test;
+
       for (i = 0; i < PIPE_MAX_VIEWPORTS; ++i) {
          setup->draw_regions[i] = setup->framebuffer;
          if (setup->scissor_test) {
@@ -1327,6 +1392,36 @@ try_update_scene_state( struct lp_setup_context *setup )
                                          &setup->draw_regions[i]);
          }
       }
+      if (setup->permit_linear_rasterizer) {
+         /* NOTE: this only takes first vp into account. */
+         boolean need_vp_scissoring = !!memcmp(&setup->vpwh, &setup->framebuffer,
+                                               sizeof(setup->framebuffer));
+         assert(setup->viewport_index_slot < 0);
+         setup->scissor_or_vp_clip |= need_vp_scissoring;
+         if (need_vp_scissoring) {
+            u_rect_possible_intersection(&setup->vpwh,
+                                         &setup->draw_regions[0]);
+         }
+      }
+      else if (setup->point_tri_clip) {
+         /*
+          * for d3d-style point clipping, we're going to need
+          * the fake vp scissor too. Hence do the intersection with vp,
+          * but don't indicate this. As above this will only work for first vp
+          * which should be ok because we instruct draw to only skip point
+          * clipping when there's only one viewport (this works because d3d10
+          * points are always single pixel).
+          * (Also note that if we have permit_linear_rasterizer this will
+          * cause large points to always get vp scissored, regardless the
+          * point_tri_clip setting.)
+          */
+         boolean need_vp_scissoring = !!memcmp(&setup->vpwh, &setup->framebuffer,
+                                               sizeof(setup->framebuffer));
+         if (need_vp_scissoring) {
+            u_rect_possible_intersection(&setup->vpwh,
+                                         &setup->draw_regions[0]);
+         }
+      }
    }
 
    setup->dirty = 0;
diff --git a/src/gallium/drivers/llvmpipe/lp_setup.h b/src/gallium/drivers/llvmpipe/lp_setup.h
index a3456c281ea..a6e8e1f49ed 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup.h
+++ b/src/gallium/drivers/llvmpipe/lp_setup.h
@@ -88,7 +88,8 @@ lp_setup_set_line_state( struct lp_setup_context *setup,
 
 void 
 lp_setup_set_point_state( struct lp_setup_context *setup,
-                          float point_size,                          
+                          float point_size,
+                          boolean point_tri_clip,
                           boolean point_size_per_vertex,
                           uint sprite_coord_enable,
                           uint sprite_coord_origin,
@@ -168,6 +169,10 @@ void
 lp_setup_set_vertex_info( struct lp_setup_context *setup, 
                           struct vertex_info *info );
 
+void
+lp_setup_set_linear_mode( struct lp_setup_context *setup, 
+                          boolean permit_linear_rasterizer );
+
 void
 lp_setup_begin_query(struct lp_setup_context *setup,
                      struct llvmpipe_query *pq);
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_analysis.c b/src/gallium/drivers/llvmpipe/lp_setup_analysis.c
new file mode 100644
index 00000000000..68457e96d42
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_setup_analysis.c
@@ -0,0 +1,398 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMWare, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+/**
+ * Look for common topology patterns which can be converted into rectangles.
+ */
+
+
+#include "lp_setup_context.h"
+#include "draw/draw_vbuf.h"
+#include "draw/draw_vertex.h"
+#include "util/u_memory.h"
+#include "util/u_math.h"
+#include "lp_state_fs.h"
+#include "lp_state_setup.h"
+#include "lp_perf.h"
+
+
+/**
+ * Duplicated from lp_setup_vbuf.c.
+ */
+typedef const float (*const_float4_ptr)[4];
+
+
+static inline
+const_float4_ptr get_vert(const void *vertex_buffer, int index, int stride)
+{
+   return (const_float4_ptr)((char *)vertex_buffer + index * stride);
+}
+
+
+/* Aero sends these weird zero area triangles.  Test for them here.
+ */
+static boolean
+is_zero_area(const float (*v0)[4],
+             const float (*v1)[4],
+             const float (*v2)[4])
+{
+   /* Specialized test for v0.y == v1.y == v2.y.
+    */
+   return (v0[0][1] == v1[0][1] &&
+           v0[0][1] == v2[0][1]);
+}
+
+
+/**
+ * Assuming axis-aligned stretch blit (s a function of x alone, t a
+ * function of y alone), create a new vertex as in:
+ *
+ *   vx------+
+ *    |      |
+ *    |      |
+ *   out-----vy
+ */
+static void
+make_vert(const float (*vx)[4],
+          const float (*vy)[4],
+          float (*out)[4])
+{
+   out[0][0] = vx[0][0];
+   out[0][1] = vy[0][1];
+   out[0][2] = vx[0][2];
+   out[0][3] = vx[0][3];
+   out[1][0] = vx[1][0];
+   out[1][1] = vy[1][1];
+}
+
+
+/* Calculate axis-aligned interpolant for s as a function of x.
+ */
+static void
+calc_interps(float x0, float x1,
+             float s0, float s1,
+             float *a, float *b)
+{
+   assert(x0 != x1);
+   *a = (s0 - s1) / (x0 - x1);
+   *b = s0 - *a * x0;
+}
+
+
+/* Validate axis-aligned interpolant for s and t as functions of x and
+ * y respectively.
+ */
+static boolean
+test_interps(const_float4_ptr v,
+             float as, float bs,
+             float at, float bt)
+{
+   float s = as * v[0][0] + bs;
+   float t = at * v[0][1] + bt;
+   return (util_is_approx(s, v[1][0], 1/4096.0) &&
+           util_is_approx(t, v[1][1], 1/4096.0));
+}
+
+
+static void
+rect(struct lp_setup_context *setup,
+     const float (*v0)[4],
+     const float (*v1)[4],
+     const float (*v2)[4])
+{
+   ASSERTED int culled = LP_COUNT_GET(nr_culled_rects);
+
+   if (0) {
+      float as, bs, at, bt;
+      calc_interps(v0[0][0], v2[0][0], v0[1][0], v2[1][0], &as, &bs);
+      calc_interps(v0[0][1], v2[0][1], v0[1][1], v2[1][1], &at, &bt);
+      assert(test_interps(v1, as, bs, at, bt));
+   }
+
+   assert(v0[0][0] == v1[0][0]);
+   assert(v1[0][1] == v2[0][1]);
+
+   lp_rect_cw(setup, v0, v1, v2, TRUE);
+
+   assert(culled == LP_COUNT_GET(nr_culled_rects));
+}
+
+
+/**
+ * Check this is an axis-aligned rectangle as in:
+ *
+ *   v3------v0
+ *    |      |
+ *    |      |
+ *   v2------v1
+ */
+static boolean
+test_rect(const_float4_ptr v0,
+          const_float4_ptr v1,
+          const_float4_ptr v2,
+          const_float4_ptr v3)
+{
+   if (v0[0][0] != v1[0][0] ||
+       v1[0][1] != v2[0][1] ||
+       v2[0][0] != v3[0][0] ||
+       v3[0][1] != v0[0][1])
+      return FALSE;
+
+   if (v0[0][3] != 1.0 ||
+       v1[0][3] != 1.0 ||
+       v2[0][3] != 1.0 ||
+       v3[0][3] != 1.0)
+      return FALSE;
+
+   return TRUE;
+}
+
+
+/**
+ * Aero sends the following shape as
+ *
+ *  18                                                        12
+ *    +-------------------------------------------------/----+
+ *    |\                                      /---------    /|
+ *    | \                           /---------             / |
+ *    |  \                /---------                      /  |
+ *    |   \     /---------                               /   |
+ * vA +    +--------------------------------------------+    + vC
+ *    |    | 9                                        6 |    |
+ *    |   /|                                            |\   |
+ *    |   ||                                            ||   |
+ *    |  / |                                            | \  |
+ *    |  | |                                            | |  |
+ *    |  | | 3                                        0 |  \ |
+ * vB + /  +--------------------------------------------+  | + vD
+ *    | | /                               ---------/     \ | |
+ *    |/ /                      ---------/                \ \|
+ *    ||/             ---------/                           \||
+ *    |/    ---------/                                      \|
+ *    +----/-------------------------------------------------+
+ *   1                                                        2
+ *
+ * and in the following decomposition:
+ *   (0, 1, 2)
+ *   (3, 0, 1),
+ *   (6, 0, 2),
+ *   (9, 3, 1),
+ *   (12, 2, 6),
+ *   (12, 6, 9),
+ *   (18, 1, 9),
+ *   (18, 9, 12),
+ *
+ * There's no straight-forward way to interpret the existing vertices
+ * as rectangles.  Instead we convert this into four axis-aligned
+ * rectangles by introducing new vertices at vA, vB, vC and vD, and
+ * then drawing rectangles.
+ */
+static boolean
+check_elts24(struct lp_setup_context *setup, const void *vb, int stride)
+{
+   const int count = 24;
+   const int uniq[8] = { 0, 1, 2, 3, 6, 9, 12, 18 };
+   const int elts[24] = {
+      0, 1, 2,
+      3, 0, 1,
+      6, 0, 2,
+      9, 3, 1,
+      12, 2, 6,
+      12, 6, 9,
+      18, 1, 9,
+      18, 9, 12
+   };
+   const_float4_ptr v0  = get_vert(vb, stride, 0);
+   const_float4_ptr v1  = get_vert(vb, stride, 1);
+   const_float4_ptr v2  = get_vert(vb, stride, 2);
+   const_float4_ptr v3  = get_vert(vb, stride, 3);
+   const_float4_ptr v6  = get_vert(vb, stride, 6);
+   const_float4_ptr v9  = get_vert(vb, stride, 9);
+   const_float4_ptr v12 = get_vert(vb, stride, 12);
+   const_float4_ptr v18 = get_vert(vb, stride, 18);
+
+   /* Could just calculate a set of interpolants and bin rectangle
+    * commands for this triangle list directly.  Instead, introduce
+    * some new vertices and feed to the rectangle setup code:
+    */
+   PIPE_ALIGN_VAR(16) float vA[2][4];
+   PIPE_ALIGN_VAR(16) float vB[2][4];
+   PIPE_ALIGN_VAR(16) float vC[2][4];
+   PIPE_ALIGN_VAR(16) float vD[2][4];
+
+   float as, bs;
+   float at, bt;
+   int i;
+
+   if (stride != 32)
+      return FALSE;
+
+   /* Check the shape is two rectangles:
+    */
+   if (!test_rect(v12, v2, v1, v18))
+      return FALSE;
+
+   if (!test_rect(v6, v0, v3, v9))
+      return FALSE;
+
+   /* XXX: check one rect is inside the other?
+    */
+
+   /* Check our tesselation matches:
+    */
+   for (i = 0; i < count; i++) {
+      if (memcmp(get_vert(vb, i, stride),
+                 get_vert(vb, elts[i], stride),
+                 6 * sizeof(float)) != 0)
+         return FALSE;
+   }
+
+   /* Test that this is a stretch blit, meaning we should be able to
+    * introduce vertices at will.
+    */
+   calc_interps(v0[0][0], v2[0][0], v0[1][0], v2[1][0], &as, &bs);
+   calc_interps(v0[0][1], v2[0][1], v0[1][1], v2[1][1], &at, &bt);
+
+   for (i = 0; i < ARRAY_SIZE(uniq); i++) {
+      const_float4_ptr v = get_vert(vb, stride, i);
+      if (!test_interps(v, as, bs, at, bt))
+         return FALSE;
+   }
+
+   make_vert(v18, v9, vA);
+   make_vert(v18, v3, vB);
+   make_vert(v12, v9, vC);
+   make_vert(v12, v3, vD);
+
+   assert(test_interps((const_float4_ptr)vA, as, bs, at, bt));
+   assert(test_interps((const_float4_ptr)vB, as, bs, at, bt));
+   assert(test_interps((const_float4_ptr)vC, as, bs, at, bt));
+   assert(test_interps((const_float4_ptr)vD, as, bs, at, bt));
+
+   rect(setup,
+        (const_float4_ptr)v12,
+        (const_float4_ptr)vC,
+        (const_float4_ptr)vA);
+
+   rect(setup,
+        (const_float4_ptr)v9,
+        (const_float4_ptr)v3,
+        (const_float4_ptr)vB);
+
+   rect(setup,
+        (const_float4_ptr)vD,
+        (const_float4_ptr)v2,
+        (const_float4_ptr)v1);
+
+   rect(setup,
+        (const_float4_ptr)vC,
+        (const_float4_ptr)vD,
+        (const_float4_ptr)v0);
+
+   return TRUE;
+}
+
+boolean
+lp_setup_analyse_triangles(struct lp_setup_context *setup,
+                           const void *vb,
+                           int stride,
+                           int nr)
+{
+   int i;
+   const boolean variant_blit = setup->fs.current.variant->blit;
+
+   if (0) {
+      debug_printf("%s %d\n", __FUNCTION__, nr);
+
+      if (variant_blit) {
+         debug_printf("  - blit variant\n");
+      }
+
+      for (i = 0; i < nr; i += 3) {
+         const_float4_ptr v0 = get_vert(vb, i, stride);
+         const_float4_ptr v1 = get_vert(vb, i+1, stride);
+         const_float4_ptr v2 = get_vert(vb, i+2, stride);
+         lp_setup_print_triangle(setup, v0, v1, v2);
+      }
+   }
+
+   /* When drawing some window navigator bars, aero sends a mixed up
+    * rectangle:
+    *
+    *    - first triangle ccw
+    *    - second triangle cw
+    *    - third triangle zero area.
+    */
+   if (nr == 9 &&
+       is_zero_area(get_vert(vb, nr-1, stride),
+                    get_vert(vb, nr-2, stride),
+                    get_vert(vb, nr-3, stride)))
+   {
+      const float (*v0)[4] = get_vert(vb, 0, stride);
+      const float (*v1)[4] = get_vert(vb, 1, stride);
+      const float (*v2)[4] = get_vert(vb, 2, stride);
+      const float (*v3)[4] = get_vert(vb, 3, stride);
+      const float (*v4)[4] = get_vert(vb, 4, stride);
+      const float (*v5)[4] = get_vert(vb, 5, stride);
+
+      /*
+       * [   v0,       v1,       v2   ]  [   v3,       v4,       v5   ]
+       * [(X0, Y0), (X0, Y1), (X1, Y1)]  [(X1, Y0), (X1, Y1), (X0, Y0)]
+       */
+      if (v0[0][0] == v1[0][0] && v0[0][0] == v5[0][0] &&
+          v2[0][0] == v3[0][0] && v2[0][0] == v4[0][0] &&
+          v0[0][1] == v3[0][1] && v0[0][1] == v5[0][1] &&
+          v1[0][1] == v2[0][1] && v1[0][1] == v4[0][1]) {
+
+         lp_rect_cw(setup, v0, v1, v2, TRUE);
+      }
+      return TRUE;
+   }
+
+   /* When highlighting (?) windows, aero sends a window border
+    * comprised of non-rectangular triangles, but which as a whole can
+    * be decomposed into rectangles.
+    *
+    * Again, with a zero-area trailing triangle.
+    *
+    * This requires introducing a couple of new vertices, which are
+    * luckily easy to compute.
+    */
+   if (nr == 27 &&
+       variant_blit &&
+       setup->setup.variant->key.inputs[0].src_index == 1 &&
+       setup->setup.variant->key.inputs[0].usage_mask == 0x3 &&
+       is_zero_area(get_vert(vb, nr-1, stride),
+                    get_vert(vb, nr-2, stride),
+                    get_vert(vb, nr-3, stride)) &&
+       check_elts24(setup, vb, stride))
+   {
+      return TRUE;
+   }
+
+   return FALSE;
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_context.h b/src/gallium/drivers/llvmpipe/lp_setup_context.h
index e65dbeb06cd..fe645fc9516 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_context.h
+++ b/src/gallium/drivers/llvmpipe/lp_setup_context.h
@@ -96,15 +96,18 @@ struct lp_setup_context
    struct llvmpipe_query *active_queries[LP_MAX_ACTIVE_BINNED_QUERIES];
    unsigned active_binned_queries;
 
-   boolean flatshade_first;
-   boolean ccw_is_frontface;
-   boolean scissor_test;
-   boolean point_size_per_vertex;
-   boolean legacy_points;
-   boolean rasterizer_discard;
-   boolean multisample;
-   boolean rectangular_lines;
-   unsigned cullmode;
+   unsigned flatshade_first:1;
+   unsigned ccw_is_frontface:1;
+   unsigned scissor_test:1;
+   unsigned scissor_or_vp_clip:1;
+   unsigned point_tri_clip:1;
+   unsigned point_size_per_vertex:1;
+   unsigned legacy_points:1;
+   unsigned rasterizer_discard:1;
+   unsigned permit_linear_rasterizer:1;
+   unsigned multisample:1;
+   unsigned rectangular_lines:1;
+   unsigned cullmode:2; /**< PIPE_FACE_x */
    unsigned bottom_edge_rule;
    float pixel_offset;
    float line_width;
@@ -117,6 +120,7 @@ struct lp_setup_context
    struct pipe_framebuffer_state fb;
    struct u_rect framebuffer;
    struct u_rect scissors[PIPE_MAX_VIEWPORTS];
+   struct u_rect vpwh;
    struct u_rect draw_regions[PIPE_MAX_VIEWPORTS];   /* intersection of fb & scissor */
    struct lp_jit_viewport viewports[PIPE_MAX_VIEWPORTS];
 
@@ -179,6 +183,15 @@ struct lp_setup_context
                      const float (*v0)[4],
                      const float (*v1)[4],
                      const float (*v2)[4]);
+
+   boolean
+   (*rect)( struct lp_setup_context *,
+            const float (*v0)[4],
+            const float (*v1)[4],
+            const float (*v2)[4],
+            const float (*v3)[4],
+            const float (*v4)[4],
+            const float (*v5)[4]);
 };
 
 static inline void
@@ -199,6 +212,7 @@ scissor_planes_needed(boolean scis_planes[4], const struct u_rect *bbox,
 void lp_setup_choose_triangle( struct lp_setup_context *setup );
 void lp_setup_choose_line( struct lp_setup_context *setup );
 void lp_setup_choose_point( struct lp_setup_context *setup );
+void lp_setup_choose_rect( struct lp_setup_context *setup );
 
 void lp_setup_init_vbuf(struct lp_setup_context *setup);
 
@@ -209,6 +223,15 @@ void lp_setup_destroy( struct lp_setup_context *setup );
 
 boolean lp_setup_flush_and_restart(struct lp_setup_context *setup);
 
+boolean
+lp_setup_whole_tile(struct lp_setup_context *setup,
+                    const struct lp_rast_shader_inputs *inputs,
+                    int tx, int ty);
+
+boolean
+lp_setup_is_blit(const struct lp_setup_context *setup,
+                 const struct lp_rast_shader_inputs *inputs);
+
 void
 lp_setup_print_triangle(struct lp_setup_context *setup,
                         const float (*v0)[4],
@@ -220,6 +243,19 @@ lp_setup_print_vertex(struct lp_setup_context *setup,
                       const char *name,
                       const float (*v)[4]);
 
+void
+lp_rect_cw(struct lp_setup_context *setup,
+           const float (*v0)[4],
+           const float (*v1)[4],
+           const float (*v2)[4],
+           boolean frontfacing);
+
+void 
+lp_setup_triangle_ccw( struct lp_setup_context *setup,
+                       const float (*v0)[4],
+                       const float (*v1)[4],
+                       const float (*v2)[4],
+                       boolean front );
 
 struct lp_rast_triangle *
 lp_setup_alloc_triangle(struct lp_scene *scene,
@@ -227,6 +263,16 @@ lp_setup_alloc_triangle(struct lp_scene *scene,
                         unsigned nr_planes,
                         unsigned *tri_size);
 
+struct lp_rast_rectangle *
+lp_setup_alloc_rectangle(struct lp_scene *scene,
+                         unsigned nr_inputs);
+
+boolean
+lp_setup_analyse_triangles(struct lp_setup_context *setup,
+                           const void *vb,
+                           int stride,
+                           int nr);
+
 boolean
 lp_setup_bin_triangle(struct lp_setup_context *setup,
                       struct lp_rast_triangle *tri,
@@ -235,4 +281,9 @@ lp_setup_bin_triangle(struct lp_setup_context *setup,
                       int nr_planes,
                       unsigned scissor_index);
 
+boolean
+lp_setup_bin_rectangle(struct lp_setup_context *setup,
+                       struct lp_rast_rectangle *rect);
+
+
 #endif
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_line.c b/src/gallium/drivers/llvmpipe/lp_setup_line.c
index 38ff15f213e..2688c60ee70 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_line.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_line.c
@@ -586,15 +586,8 @@ try_setup_line( struct lp_setup_context *setup,
       bbox.y1--;
    }
 
-   if (bbox.x1 < bbox.x0 ||
-       bbox.y1 < bbox.y0) {
-      if (0) debug_printf("empty bounding box\n");
-      LP_COUNT(nr_culled_tris);
-      return TRUE;
-   }
-
    if (!u_rect_test_intersection(&setup->draw_regions[viewport_index], &bbox)) {
-      if (0) debug_printf("offscreen\n");
+      if (0) debug_printf("no intersection\n");
       LP_COUNT(nr_culled_tris);
       return TRUE;
    }
@@ -611,9 +604,11 @@ try_setup_line( struct lp_setup_context *setup,
     * Determine how many scissor planes we need, that is drop scissor
     * edges if the bounding box of the tri is fully inside that edge.
     */
-   scissor = &setup->draw_regions[viewport_index];
-   scissor_planes_needed(s_planes, &bboxpos, scissor);
-   nr_planes += s_planes[0] + s_planes[1] + s_planes[2] + s_planes[3];
+   if (setup->scissor_or_vp_clip) {
+      scissor = &setup->draw_regions[viewport_index];
+      scissor_planes_needed(s_planes, &bboxpos, scissor);
+      nr_planes += s_planes[0] + s_planes[1] + s_planes[2] + s_planes[3];
+   }
 
    line = lp_setup_alloc_triangle(scene,
                                   key->num_inputs,
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_point.c b/src/gallium/drivers/llvmpipe/lp_setup_point.c
index bca2d47d135..335b9d0ccfc 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_point.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_point.c
@@ -352,11 +352,8 @@ try_setup_point( struct lp_setup_context *setup,
    int adj = (setup->bottom_edge_rule != 0) ? 1 : 0;
    float pixel_offset = setup->multisample ? 0.0 : setup->pixel_offset;
    struct lp_scene *scene = setup->scene;
-   struct lp_rast_triangle *point;
-   unsigned bytes;
    struct u_rect bbox;
    int x[2], y[2];
-   unsigned nr_planes = 4;
    struct point_info info;
    unsigned viewport_index = 0;
    unsigned layer = 0;
@@ -461,56 +458,63 @@ try_setup_point( struct lp_setup_context *setup,
    }
 
    if (!u_rect_test_intersection(&setup->draw_regions[viewport_index], &bbox)) {
-      if (0) debug_printf("offscreen\n");
+      if (0) debug_printf("no intersection\n");
       LP_COUNT(nr_culled_tris);
       return TRUE;
    }
 
    u_rect_find_intersection(&setup->draw_regions[viewport_index], &bbox);
 
-   point = lp_setup_alloc_triangle(scene,
-                                   key->num_inputs,
-                                   nr_planes,
-                                   &bytes);
-   if (!point)
-      return FALSE;
-
+   /* We can't use rectangle reasterizer for non-legacy points for now. */
+   if (!setup->legacy_points || setup->multisample) {
+      struct lp_rast_triangle *point;
+      struct lp_rast_plane *plane;
+      unsigned bytes;
+      unsigned nr_planes = 4;
+      
+      point = lp_setup_alloc_triangle(scene,
+                                      key->num_inputs,
+                                      nr_planes,
+                                      &bytes);
+     if (!point)
+        return FALSE;
+        
 #ifdef DEBUG
-   point->v[0][0] = v0[0][0];
-   point->v[0][1] = v0[0][1];
+      point->v[0][0] = v0[0][0];
+      point->v[0][1] = v0[0][1];
 #endif
 
-   LP_COUNT(nr_tris);
+      LP_COUNT(nr_tris);
 
-   if (draw_will_inject_frontface(lp_context->draw) &&
-       setup->face_slot > 0) {
-      point->inputs.frontfacing = v0[setup->face_slot][0];
-   } else {
-      point->inputs.frontfacing = TRUE;
-   }
+      if (draw_will_inject_frontface(lp_context->draw) &&
+          setup->face_slot > 0) {
+         point->inputs.frontfacing = v0[setup->face_slot][0];
+      } else {
+         point->inputs.frontfacing = TRUE;
+      }
 
-   info.v0 = v0;
-   info.dx01 = 0;
-   info.dx12 = fixed_width;
-   info.dy01 = fixed_width;
-   info.dy12 = 0;
-   info.a0 = GET_A0(&point->inputs);
-   info.dadx = GET_DADX(&point->inputs);
-   info.dady = GET_DADY(&point->inputs);
-   info.frontfacing = point->inputs.frontfacing;
+      info.v0 = v0;
+      info.dx01 = 0;
+      info.dx12 = fixed_width;
+      info.dy01 = fixed_width;
+      info.dy12 = 0;
+      info.a0 = GET_A0(&point->inputs);
+      info.dadx = GET_DADX(&point->inputs);
+      info.dady = GET_DADY(&point->inputs);
+      info.frontfacing = point->inputs.frontfacing;
    
-   /* Setup parameter interpolants:
-    */
-   setup_point_coefficients(setup, &info);
+      /* Setup parameter interpolants:
+       */
+      setup_point_coefficients(setup, &info);
 
-   point->inputs.disable = FALSE;
-   point->inputs.opaque = FALSE;
-   point->inputs.layer = layer;
-   point->inputs.viewport_index = viewport_index;
-   point->inputs.view_index = setup->view_index;
+      point->inputs.disable = FALSE;
+      point->inputs.is_blit = FALSE;
+      point->inputs.opaque = setup->fs.current.variant->opaque;
+      point->inputs.layer = layer;
+      point->inputs.viewport_index = viewport_index;
+      point->inputs.view_index = setup->view_index;
 
-   {
-      struct lp_rast_plane *plane = GET_PLANES(point);
+      plane = GET_PLANES(point);
 
       plane[0].dcdx = ~0U << 8;
       plane[0].dcdy = 0;
@@ -540,9 +544,57 @@ try_setup_point( struct lp_setup_context *setup,
          else
             plane[3].c++; /* bottom-left */
       }
-   }
 
-   return lp_setup_bin_triangle(setup, point, &bbox, &bbox, nr_planes, viewport_index);
+      return lp_setup_bin_triangle(setup, point, &bbox, &bbox, nr_planes, viewport_index);
+
+   } else {
+      struct lp_rast_rectangle *point;
+      point = lp_setup_alloc_rectangle(scene,
+                                       key->num_inputs);
+      if (!point)
+         return FALSE;
+#ifdef DEBUG
+      point->v[0][0] = v0[0][0];
+      point->v[0][1] = v0[0][1];
+#endif
+
+      point->box.x0 = bbox.x0;
+      point->box.x1 = bbox.x1;
+      point->box.y0 = bbox.y0;
+      point->box.y1 = bbox.y1;
+
+      LP_COUNT(nr_tris);
+
+      if (draw_will_inject_frontface(lp_context->draw) &&
+          setup->face_slot > 0) {
+         point->inputs.frontfacing = v0[setup->face_slot][0];
+      } else {
+         point->inputs.frontfacing = TRUE;
+      }
+
+      info.v0 = v0;
+      info.dx01 = 0;
+      info.dx12 = fixed_width;
+      info.dy01 = fixed_width;
+      info.dy12 = 0;
+      info.a0 = GET_A0(&point->inputs);
+      info.dadx = GET_DADX(&point->inputs);
+      info.dady = GET_DADY(&point->inputs);
+      info.frontfacing = point->inputs.frontfacing;
+   
+      /* Setup parameter interpolants:
+       */
+      setup_point_coefficients(setup, &info);
+
+      point->inputs.disable = FALSE;
+      point->inputs.is_blit = FALSE;
+      point->inputs.opaque = setup->fs.current.variant->opaque;
+      point->inputs.layer = layer;
+      point->inputs.viewport_index = viewport_index;
+      point->inputs.view_index = setup->view_index;
+
+      return lp_setup_bin_rectangle(setup, point);
+   }
 }
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_rect.c b/src/gallium/drivers/llvmpipe/lp_setup_rect.c
new file mode 100644
index 00000000000..7c1234eafe7
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_setup_rect.c
@@ -0,0 +1,900 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL THE AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+/**
+ * Setup/binning code for screen-aligned quads.
+ */
+
+#include "util/u_math.h"
+#include "util/u_memory.h"
+#include "lp_perf.h"
+#include "lp_setup_context.h"
+#include "lp_rast.h"
+#include "lp_state_fs.h"
+#include "lp_state_setup.h"
+
+
+#define NUM_CHANNELS 4
+
+#define UNDETERMINED_BLIT  -1
+
+
+static inline int
+subpixel_snap(float a)
+{
+   return util_iround(FIXED_ONE * a);
+}
+
+
+static inline float
+fixed_to_float(int a)
+{
+   return a * (1.0f / FIXED_ONE);
+}
+
+
+/**
+ * Alloc space for a new rectangle plus the input.a0/dadx/dady arrays
+ * immediately after it.
+ * The memory is allocated from the per-scene pool, not per-tile.
+ * \param size  returns number of bytes allocated
+ * \param nr_inputs  number of fragment shader inputs
+ * \return pointer to rectangle space
+ */
+struct lp_rast_rectangle *
+lp_setup_alloc_rectangle(struct lp_scene *scene, unsigned nr_inputs)
+{
+   unsigned input_array_sz = NUM_CHANNELS * (nr_inputs + 1) * sizeof(float);
+   struct lp_rast_rectangle *rect;
+   unsigned bytes;
+
+   bytes = sizeof(*rect) + (3 * input_array_sz);
+
+   rect = lp_scene_alloc_aligned( scene, bytes, 16 );
+   if (rect == NULL)
+      return NULL;
+
+   rect->inputs.stride = input_array_sz;
+
+   return rect;
+}
+
+
+/**
+ * The rectangle covers the whole tile- shade whole tile.
+ * XXX no rectangle/triangle dependencies in this file - share it with
+ * the same code in lp_setup_tri.c
+ * \param tx, ty  the tile position in tiles, not pixels
+ */
+boolean
+lp_setup_whole_tile(struct lp_setup_context *setup,
+                    const struct lp_rast_shader_inputs *inputs,
+                    int tx, int ty)
+{
+   struct lp_scene *scene = setup->scene;
+
+   LP_COUNT(nr_fully_covered_64);
+
+   /* if variant is opaque and scissor doesn't effect the tile */
+   if (inputs->opaque) {
+      /* Several things prevent this optimization from working:
+       * - For layered rendering we can't determine if this covers the same layer
+       * as previous rendering (or in case of clears those actually always cover
+       * all layers so optimization is impossible). Need to use fb_max_layer and
+       * not setup->layer_slot to determine this since even if there's currently
+       * no slot assigned previous rendering could have used one.
+       * - If there were any Begin/End query commands in the scene then those
+       * would get removed which would be very wrong. Furthermore, if queries
+       * were just active we also can't do the optimization since to get
+       * accurate query results we unfortunately need to execute the rendering
+       * commands.
+       */
+      if (!scene->fb.zsbuf && scene->fb_max_layer == 0 && !scene->had_queries) {
+         /*
+          * All previous rendering will be overwritten so reset the bin.
+          */
+         lp_scene_bin_reset( scene, tx, ty );
+      }
+
+      if (inputs->is_blit) {
+         LP_COUNT(nr_blit_64);
+         return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                             setup->fs.stored,
+                                             LP_RAST_OP_BLIT,
+                                             lp_rast_arg_inputs(inputs) );
+      }
+      else {
+         LP_COUNT(nr_shade_opaque_64);
+         return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                             setup->fs.stored,
+                                             LP_RAST_OP_SHADE_TILE_OPAQUE,
+                                             lp_rast_arg_inputs(inputs) );
+      }
+   }
+   else {
+      LP_COUNT(nr_shade_64);
+      return lp_scene_bin_cmd_with_state( scene, tx, ty,
+                                          setup->fs.stored,
+                                          LP_RAST_OP_SHADE_TILE,
+                                          lp_rast_arg_inputs(inputs) );
+   }
+}
+
+
+boolean
+lp_setup_is_blit(const struct lp_setup_context *setup,
+                 const struct lp_rast_shader_inputs *inputs)
+{
+   const struct lp_fragment_shader_variant *variant =
+      setup->fs.current.variant;
+
+   if (variant->blit) {
+      /*
+       * Detect blits.
+       */
+      const struct lp_jit_texture *texture =
+         &setup->fs.current.jit_context.textures[0];
+      float dsdx, dsdy, dtdx, dtdy;
+
+      /* XXX: dadx vs dady confusion below?
+       */
+      dsdx = GET_DADX(inputs)[1][0]*texture->width;
+      dsdy = GET_DADX(inputs)[1][1]*texture->width;
+      dtdx = GET_DADY(inputs)[1][0]*texture->height;
+      dtdy = GET_DADY(inputs)[1][1]*texture->height;
+
+      /*
+       * We don't need to check s0/t0 tolerances
+       * as we establish as pre-condition that there is no
+       * texture filtering.
+       */
+
+      assert(variant->key.samplers[0].sampler_state.min_img_filter == PIPE_TEX_FILTER_NEAREST);
+      assert(variant->key.samplers[0].sampler_state.mag_img_filter == PIPE_TEX_FILTER_NEAREST);
+
+      /*
+       * Check for 1:1 match of texels to dest pixels
+       */
+
+      if (util_is_approx(dsdx, 1.0f, 1.0f/LP_MAX_WIDTH) &&
+          util_is_approx(dsdy, 0.0f, 1.0f/LP_MAX_HEIGHT) &&
+          util_is_approx(dtdx, 0.0f, 1.0f/LP_MAX_WIDTH) &&
+          util_is_approx(dtdy, 1.0f, 1.0f/LP_MAX_HEIGHT)) {
+         return true;
+      }
+      else {
+#if 0
+         debug_printf("dsdx = %f\n", dsdx);
+         debug_printf("dsdy = %f\n", dsdy);
+         debug_printf("dtdx = %f\n", dtdx);
+         debug_printf("dtdy = %f\n", dtdy);
+         debug_printf("\n");
+#endif
+         return FALSE;
+      }
+   }
+
+   return FALSE;
+}
+
+
+static inline void
+partial(struct lp_setup_context *setup,
+        const struct lp_rast_rectangle *rect,
+        unsigned ix, unsigned iy,
+        unsigned mask)
+{
+   if (mask == 0) {
+      assert(rect->box.x0 <= ix * TILE_SIZE);
+      assert(rect->box.y0 <= iy * TILE_SIZE);
+      assert(rect->box.x1 >= (ix+1) * TILE_SIZE - 1);
+      assert(rect->box.y1 >= (iy+1) * TILE_SIZE - 1);
+
+      lp_setup_whole_tile(setup, &rect->inputs, ix, iy);
+   }
+   else {
+      LP_COUNT(nr_partially_covered_64);
+      lp_scene_bin_cmd_with_state( setup->scene,
+                                   ix, iy,
+                                   setup->fs.stored,
+                                   LP_RAST_OP_RECTANGLE,
+                                   lp_rast_arg_rectangle(rect) );
+   }
+}
+
+
+/**
+ * Setup/bin a screen-aligned rect.
+ * We need three corner vertices in order to correctly setup
+ * interpolated parameters.  We *could* get away with just the
+ * diagonal vertices but it'd cause ugliness elsewhere.
+ *
+ *   + -------v0
+ *   |        |
+ *  v2 ------ v1
+ *
+ * By an unfortunate mixup between GL and D3D coordinate spaces, half
+ * of this file talks about clockwise rectangles (which were CCW in GL
+ * coordinate space), while the other half prefers to work with D3D
+ * CCW rectangles.
+ */
+static boolean
+try_rect_cw(struct lp_setup_context *setup,
+            const float (*v0)[4],
+            const float (*v1)[4],
+            const float (*v2)[4],
+            boolean frontfacing)
+{
+   const struct lp_fragment_shader_variant *variant =
+      setup->fs.current.variant;
+   const struct lp_setup_variant_key *key = &setup->setup.variant->key;
+   struct lp_scene *scene = setup->scene;
+   struct lp_rast_rectangle *rect;
+   boolean cw;
+   struct u_rect bbox;
+   unsigned viewport_index = 0;
+   unsigned layer = 0;
+   const float (*pv)[4];
+
+   /* x/y positions in fixed point */
+   int x0 = subpixel_snap(v0[0][0] - setup->pixel_offset);
+   int x1 = subpixel_snap(v1[0][0] - setup->pixel_offset);
+   int x2 = subpixel_snap(v2[0][0] - setup->pixel_offset);
+   int y0 = subpixel_snap(v0[0][1] - setup->pixel_offset);
+   int y1 = subpixel_snap(v1[0][1] - setup->pixel_offset);
+   int y2 = subpixel_snap(v2[0][1] - setup->pixel_offset);
+
+   LP_COUNT(nr_rects);
+
+   /* Cull clockwise rects without overflowing.
+    */
+   cw = (x2 < x1) ^ (y0 < y2);
+   if (cw) {
+      LP_COUNT(nr_culled_rects);
+      return TRUE;
+   }
+
+   if (setup->flatshade_first) {
+      pv = v0;
+   }
+   else {
+      pv = v2;
+   }
+   if (setup->viewport_index_slot > 0) {
+      unsigned *udata = (unsigned*)pv[setup->viewport_index_slot];
+      viewport_index = lp_clamp_viewport_idx(*udata);
+   }
+   if (setup->layer_slot > 0) {
+      layer = *(unsigned*)pv[setup->layer_slot];
+      layer = MIN2(layer, scene->fb_max_layer);
+   }
+
+   /* Bounding rectangle (in pixels) */
+   {
+      /* Yes this is necessary to accurately calculate bounding boxes
+       * with the two fill-conventions we support.  GL (normally) ends
+       * up needing a bottom-left fill convention, which requires
+       * slightly different rounding.
+       */
+      int adj = (setup->bottom_edge_rule != 0) ? 1 : 0;
+
+      bbox.x0 = (MIN3(x0, x1, x2) + (FIXED_ONE-1)) >> FIXED_ORDER;
+      bbox.x1 = (MAX3(x0, x1, x2) + (FIXED_ONE-1)) >> FIXED_ORDER;
+      bbox.y0 = (MIN3(y0, y1, y2) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
+      bbox.y1 = (MAX3(y0, y1, y2) + (FIXED_ONE-1) + adj) >> FIXED_ORDER;
+
+      /* Inclusive coordinates:
+       */
+      bbox.x1--;
+      bbox.y1--;
+   }
+
+   if (!u_rect_test_intersection(&setup->draw_regions[viewport_index], &bbox)) {
+      if (0) debug_printf("no intersection\n");
+      LP_COUNT(nr_culled_rects);
+      return TRUE;
+   }
+
+   u_rect_find_intersection(&setup->draw_regions[viewport_index], &bbox);
+
+   rect = lp_setup_alloc_rectangle(scene, key->num_inputs);
+   if (!rect)
+      return FALSE;
+
+#ifdef DEBUG
+   rect->v[0][0] = v0[0][0];
+   rect->v[0][1] = v0[0][1];
+   rect->v[1][0] = v1[0][0];
+   rect->v[1][1] = v1[0][1];
+#endif
+
+   rect->box.x0 = bbox.x0;
+   rect->box.x1 = bbox.x1;
+   rect->box.y0 = bbox.y0;
+   rect->box.y1 = bbox.y1;
+
+   /* Setup parameter interpolants:
+    */
+   setup->setup.variant->jit_function( v0,
+				       v1,
+				       v2,
+				       frontfacing,
+				       GET_A0(&rect->inputs),
+				       GET_DADX(&rect->inputs),
+				       GET_DADY(&rect->inputs),
+                                       &setup->setup.variant->key );
+
+   rect->inputs.frontfacing = frontfacing;
+   rect->inputs.disable = FALSE;
+   rect->inputs.is_blit = lp_setup_is_blit(setup, &rect->inputs);
+   rect->inputs.opaque = variant->opaque;
+   rect->inputs.layer = layer;
+   rect->inputs.viewport_index = viewport_index;
+   rect->inputs.view_index = setup->view_index;
+
+   return lp_setup_bin_rectangle(setup, rect);
+}
+
+
+boolean
+lp_setup_bin_rectangle(struct lp_setup_context *setup,
+                       struct lp_rast_rectangle *rect)
+{
+   struct lp_scene *scene = setup->scene;
+   unsigned ix0, iy0, ix1, iy1;
+   unsigned i, j;
+   unsigned left_mask = 0;
+   unsigned right_mask = 0;
+   unsigned top_mask = 0;
+   unsigned bottom_mask = 0;
+
+   /*
+    * All fields of 'rect' are now set.  The remaining code here is
+    * concerned with binning.
+    */
+
+   /* Convert to inclusive tile coordinates:
+    */
+   ix0 = rect->box.x0 / TILE_SIZE;
+   iy0 = rect->box.y0 / TILE_SIZE;
+   ix1 = rect->box.x1 / TILE_SIZE;
+   iy1 = rect->box.y1 / TILE_SIZE;
+
+   /*
+    * Clamp to framebuffer size
+    */
+   assert(ix0 == MAX2(ix0, 0));
+   assert(iy0 == MAX2(iy0, 0));
+   assert(ix1 == MIN2(ix1, scene->tiles_x - 1));
+   assert(iy1 == MIN2(iy1, scene->tiles_y - 1));
+
+   if (ix0 * TILE_SIZE != rect->box.x0)
+      left_mask = RECT_PLANE_LEFT;
+
+   if (ix1 * TILE_SIZE + TILE_SIZE - 1 != rect->box.x1)
+      right_mask  = RECT_PLANE_RIGHT;
+
+   if (iy0 * TILE_SIZE != rect->box.y0)
+      top_mask    = RECT_PLANE_TOP;
+
+   if (iy1 * TILE_SIZE + TILE_SIZE - 1 != rect->box.y1)
+      bottom_mask = RECT_PLANE_BOTTOM;
+
+   /* Determine which tile(s) intersect the rectangle's bounding box
+    */
+   if (iy0 == iy1 && ix0 == ix1) {
+      partial(setup, rect, ix0, iy0,
+              (left_mask | right_mask | top_mask | bottom_mask));
+   }
+   else if (ix0 == ix1) {
+      unsigned mask = left_mask | right_mask;
+      partial(setup, rect, ix0, iy0, mask | top_mask);
+      for (i = iy0 + 1; i < iy1; i++)
+         partial(setup, rect, ix0, i, mask);
+      partial(setup, rect, ix0, iy1, mask | bottom_mask);
+   }
+   else if (iy0 == iy1) {
+      unsigned mask = top_mask | bottom_mask;
+      partial(setup, rect, ix0, iy0, mask | left_mask);
+      for (i = ix0 + 1; i < ix1; i++)
+         partial(setup, rect, i, iy0, mask);
+      partial(setup, rect, ix1, iy0, mask | right_mask);
+   }
+   else {
+      partial(setup, rect, ix0, iy0, left_mask  | top_mask);
+      partial(setup, rect, ix0, iy1, left_mask  | bottom_mask);
+      partial(setup, rect, ix1, iy0, right_mask | top_mask);
+      partial(setup, rect, ix1, iy1, right_mask | bottom_mask);
+
+      /* Top/Bottom fringes
+       */
+      for (i = ix0 + 1; i < ix1; i++) {
+         partial(setup, rect, i, iy0, top_mask);
+         partial(setup, rect, i, iy1, bottom_mask);
+      }
+
+      /* Left/Right fringes
+       */
+      for (i = iy0 + 1; i < iy1; i++) {
+         partial(setup, rect, ix0, i, left_mask);
+         partial(setup, rect, ix1, i, right_mask);
+      }
+
+      /* Full interior tiles
+       */
+      for (j = iy0 + 1; j < iy1; j++) {
+         for (i = ix0 + 1; i < ix1; i++) {
+            lp_setup_whole_tile(setup, &rect->inputs, i, j);
+         }
+      }
+   }
+
+   /* Catch any out-of-memory which occurred during binning.  Do this
+    * once here rather than checking all the return values throughout.
+    */
+   if (lp_scene_is_oom(scene)) {
+      /* Disable rasterization of this partially-binned rectangle.
+       * We'll flush this scene and re-bin the entire rectangle:
+       */
+      rect->inputs.disable = TRUE;
+      return FALSE;
+   }
+
+   return TRUE;
+}
+
+
+void
+lp_rect_cw(struct lp_setup_context *setup,
+         const float (*v0)[4],
+         const float (*v1)[4],
+         const float (*v2)[4],
+         boolean frontfacing)
+{
+   if (!try_rect_cw(setup, v0, v1, v2, frontfacing)) {
+      if (!lp_setup_flush_and_restart(setup))
+         return;
+
+      if (!try_rect_cw(setup, v0, v1, v2, frontfacing))
+         return;
+   }
+}
+
+
+/**
+ * Take the six vertices for two triangles and try to determine if they
+ * form a screen-aligned quad/rectangle.  If so, draw the rect directly,
+ * else, draw as two regular triangles.
+ */
+static boolean
+do_rect_ccw(struct lp_setup_context *setup,
+            const float (*v0)[4],
+            const float (*v1)[4],
+            const float (*v2)[4],
+            const float (*v3)[4],
+            const float (*v4)[4],
+            const float (*v5)[4],
+            boolean front)
+{
+   const float (*rv0)[4], (*rv1)[4], (*rv2)[4], (*rv3)[4];  /* rect verts */
+
+#define SAME_POS(A, B)   (A[0][0] == B[0][0] && \
+                          A[0][1] == B[0][1] && \
+                          A[0][2] == B[0][2] && \
+                          A[0][3] == B[0][3])
+
+   /* Only need to consider CCW orientations.  There are nine ways
+    * that two counter-clockwise triangles can join up:
+    */
+   if (SAME_POS(v0, v3)) {
+      if (SAME_POS(v2, v4)) {
+         /*
+          *    v5   v4/v2
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v3/v0   v1
+          */
+         rv0 = v5;
+         rv1 = v0;
+         rv2 = v1;
+         rv3 = v2;
+      }
+      else if (SAME_POS(v1, v5)) {
+         /*
+          *    v4   v3/v0
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v5/v1   v2
+          */
+         rv0 = v4;
+         rv1 = v1;
+         rv2 = v2;
+         rv3 = v0;
+      }
+      else {
+         goto emit_triangles;
+      }
+   }
+   else if (SAME_POS(v0, v5)) {
+      if (SAME_POS(v2, v3)) {
+         /*
+          *    v4   v3/v2
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v5/v0   v1
+          */
+         rv0 = v4;
+         rv1 = v0;
+         rv2 = v1;
+         rv3 = v2;
+      }
+      else if (SAME_POS(v1, v4)) {
+         /*
+          *    v3   v5/v0
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v4/v1   v2
+          */
+         rv0 = v3;
+         rv1 = v1;
+         rv2 = v2;
+         rv3 = v0;
+      }
+      else {
+         goto emit_triangles;
+      }
+   }
+   else if (SAME_POS(v0, v4)) {
+      if (SAME_POS(v2, v5)) {
+         /*
+          *    v3   v5/v2
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v4/v0   v1
+          */
+         rv0 = v3;
+         rv1 = v0;
+         rv2 = v1;
+         rv3 = v2;
+      }
+      else if (SAME_POS(v1, v3)) {
+         /*
+          *    v5   v4/v0
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v3/v1   v2
+          */
+         rv0 = v5;
+         rv1 = v1;
+         rv2 = v2;
+         rv3 = v0;
+      }
+      else {
+         goto emit_triangles;
+      }
+   }
+   else if (SAME_POS(v2, v3)) {
+      if (SAME_POS(v1, v4)) {
+         /*
+          *    v5   v4/v1
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v3/v2   v0
+          */
+         rv0 = v5;
+         rv1 = v2;
+         rv2 = v0;
+         rv3 = v1;
+      }
+      else {
+         goto emit_triangles;
+      }
+   }
+   else if (SAME_POS(v2, v5)) {
+      if (SAME_POS(v1, v3)) {
+         /*
+          *    v4   v3/v1
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v5/v2   v0
+          */
+         rv0 = v4;
+         rv1 = v2;
+         rv2 = v0;
+         rv3 = v1;
+      }
+      else {
+         goto emit_triangles;
+      }
+   }
+   else if (SAME_POS(v2, v4)) {
+      if (SAME_POS(v1, v5)) {
+         /*
+          *    v3   v5/v1
+          *     +-----+
+          *     |   / |
+          *     |  /  |
+          *     | /   |
+          *     +-----+
+          *   v4/v2   v0
+          */
+         rv0 = v3;
+         rv1 = v2;
+         rv2 = v0;
+         rv3 = v1;
+      }
+      else {
+         goto emit_triangles;
+      }
+   }
+   else {
+      goto emit_triangles;
+   }
+
+
+#define SAME_X(A, B)   (A[0][0] == B[0][0])
+#define SAME_Y(A, B)   (A[0][1] == B[0][1])
+
+   /* The vertices are now counter clockwise, as such:
+    *
+    *  rv0 -------rv3
+    *    |        |
+    *  rv1 ------ rv2
+    *
+    * To render as a rectangle,
+    *   * The X values should be the same at v0, v1 and v2, v3.
+    *   * The Y values should be the same at v0, v3 and v1, v2.
+    */
+   if (SAME_Y(rv0, rv1)) {
+      const float (*tmp)[4];
+      tmp = rv0;
+      rv0 = rv1;
+      rv1 = rv2;
+      rv2 = rv3;
+      rv3 = tmp;
+   }
+
+   if (SAME_X(rv0, rv1) && SAME_X(rv2, rv3) &&
+       SAME_Y(rv0, rv3) && SAME_Y(rv1, rv2)) {
+
+      const struct lp_setup_variant_key *key = &setup->setup.variant->key;
+      const unsigned n = key->num_inputs;
+      unsigned i, j;
+
+      /* We have a rectangle.  Check that the other attributes are
+       * coplanar.
+       */
+      for (i = 0; i < n; i++) {
+         for (j = 0; j < 4; j++) {
+            if (key->inputs[i].usage_mask & (1<<j)) {
+               unsigned k = key->inputs[i].src_index;
+               float dxdx1, dxdx2, dxdy1, dxdy2;
+               dxdx1 = rv0[k][j] - rv3[k][j];
+               dxdx2 = rv1[k][j] - rv2[k][j];
+               dxdy1 = rv0[k][j] - rv1[k][j];
+               dxdy2 = rv3[k][j] - rv2[k][j];
+               if (dxdx1 != dxdx2 ||
+                   dxdy1 != dxdy2) {
+                  goto emit_triangles;
+               }
+            }
+         }
+      }
+
+      /* Note we're changing to clockwise here.  Fix this by reworking
+       * lp_rect_cw to expect/operate on ccw rects.  Note that
+       * function was previously misnamed.
+       */
+      lp_rect_cw(setup, rv0, rv2, rv1, front);
+      return TRUE;
+   }
+   else {
+      /* setup->quad(setup, rv0, rv1, rv2, rv3); */
+   }
+
+emit_triangles:
+   return FALSE;
+}
+
+
+enum winding {
+   WINDING_NONE = 0,
+   WINDING_CCW,
+   WINDING_CW
+};
+
+
+static inline enum winding
+winding(const float (*v0)[4],
+        const float (*v1)[4],
+        const float (*v2)[4])
+{
+   /* edge vectors e = v0 - v2, f = v1 - v2 */
+   const float ex = v0[0][0] - v2[0][0];
+   const float ey = v0[0][1] - v2[0][1];
+   const float fx = v1[0][0] - v2[0][0];
+   const float fy = v1[0][1] - v2[0][1];
+
+   /* det = cross(e,f).z */
+   const float det = ex * fy - ey * fx;
+
+   if (det < 0.0f)
+      return WINDING_CCW;
+   else if (det > 0.0f)
+      return WINDING_CW;
+   else
+      return WINDING_NONE;
+}
+
+
+static boolean
+setup_rect_cw(struct lp_setup_context *setup,
+              const float (*v0)[4],
+              const float (*v1)[4],
+              const float (*v2)[4],
+              const float (*v3)[4],
+              const float (*v4)[4],
+              const float (*v5)[4])
+{
+   enum winding winding0 = winding(v0, v1, v2);
+   enum winding winding1 = winding(v3, v4, v5);
+
+   if (winding0 == WINDING_CW &&
+       winding1 == WINDING_CW) {
+      return do_rect_ccw(setup, v0, v2, v1, v3, v5, v4, !setup->ccw_is_frontface);
+   } else if (winding0 == WINDING_CW) {
+      setup->triangle(setup, v0, v1, v2);
+      return TRUE;
+   } else if (winding1 == WINDING_CW) {
+      setup->triangle(setup, v3, v4, v5);
+      return TRUE;
+   } else {
+      return TRUE;
+   }
+}
+
+
+static boolean
+setup_rect_ccw(struct lp_setup_context *setup,
+               const float (*v0)[4],
+               const float (*v1)[4],
+               const float (*v2)[4],
+               const float (*v3)[4],
+               const float (*v4)[4],
+               const float (*v5)[4])
+{
+   enum winding winding0 = winding(v0, v1, v2);
+   enum winding winding1 = winding(v3, v4, v5);
+
+   if (winding0 == WINDING_CCW &&
+       winding1 == WINDING_CCW) {
+      return do_rect_ccw(setup, v0, v1, v2, v3, v4, v5, setup->ccw_is_frontface);
+   } else if (winding0 == WINDING_CCW) {
+      setup->triangle(setup, v0, v1, v2);
+      return TRUE;
+   } else if (winding1 == WINDING_CCW) {
+      return FALSE;
+      setup->triangle(setup, v3, v4, v5);
+      return TRUE;
+   } else {
+      return TRUE;
+   }
+}
+
+
+static boolean
+setup_rect_noop(struct lp_setup_context *setup,
+                const float (*v0)[4],
+                const float (*v1)[4],
+                const float (*v2)[4],
+                const float (*v3)[4],
+                const float (*v4)[4],
+                const float (*v5)[4])
+{
+   return TRUE;
+}
+
+
+static boolean
+setup_rect_both(struct lp_setup_context *setup,
+                const float (*v0)[4],
+                const float (*v1)[4],
+                const float (*v2)[4],
+                const float (*v3)[4],
+                const float (*v4)[4],
+                const float (*v5)[4])
+{
+   enum winding winding0 = winding(v0, v1, v2);
+   enum winding winding1 = winding(v3, v4, v5);
+
+   if (winding0 != winding1) {
+      /* If we knew that the "front" parameter wasn't going to be
+       * referenced, could rearrange one of the two triangles such
+       * that they were both CCW.  Aero actually does send mixed
+       * CW/CCW rectangles under some circumstances, but we catch them
+       * explicitly.
+       */
+      return FALSE;
+   }
+   else if (winding0 == WINDING_CCW) {
+      return do_rect_ccw(setup, v0, v1, v2, v3, v4, v5, setup->ccw_is_frontface);
+   }
+   else if (winding0 == WINDING_CW) {
+      return do_rect_ccw(setup, v0, v2, v1, v3, v5, v4, !setup->ccw_is_frontface);
+   } else {
+      return TRUE;
+   }
+}
+
+
+void
+lp_setup_choose_rect( struct lp_setup_context *setup )
+{
+   if (setup->rasterizer_discard) {
+      setup->rect = setup_rect_noop;
+      return;
+   }
+
+   switch (setup->cullmode) {
+   case PIPE_FACE_NONE:
+      setup->rect = setup_rect_both;
+      break;
+   case PIPE_FACE_BACK:
+      setup->rect = setup->ccw_is_frontface ? setup_rect_ccw : setup_rect_cw;
+      break;
+   case PIPE_FACE_FRONT:
+      setup->rect = setup->ccw_is_frontface ? setup_rect_cw : setup_rect_ccw;
+      break;
+   default:
+      setup->rect = setup_rect_noop;
+      break;
+   }
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_tri.c b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
index 4fb76dd22d2..56494d49a93 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_tri.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_tri.c
@@ -205,6 +205,7 @@ lp_rast_32_tri_tab[MAX_PLANES+1] = {
    LP_RAST_OP_TRIANGLE_32_8
 };
 
+
 static unsigned
 lp_rast_ms_tri_tab[MAX_PLANES+1] = {
    0,               /* should be impossible */
@@ -218,56 +219,46 @@ lp_rast_ms_tri_tab[MAX_PLANES+1] = {
    LP_RAST_OP_MS_TRIANGLE_8
 };
 
-/**
- * The primitive covers the whole tile- shade whole tile.
+/*
+ * Detect big primitives drawn with an alpha == 1.0.
  *
- * \param tx, ty  the tile position in tiles, not pixels
+ * This is used when simulating anti-aliasing primitives in shaders, e.g.,
+ * when drawing the windows client area in Aero's flip-3d effect.
  */
 static boolean
-lp_setup_whole_tile(struct lp_setup_context *setup,
-                    const struct lp_rast_shader_inputs *inputs,
-                    int tx, int ty)
+check_opaque(struct lp_setup_context *setup,
+             const float (*v1)[4],
+             const float (*v2)[4],
+             const float (*v3)[4])
 {
-   struct lp_scene *scene = setup->scene;
+   const struct lp_fragment_shader_variant *variant =
+      setup->fs.current.variant;
+   const struct lp_tgsi_channel_info *alpha_info = &variant->shader->info.cbuf[0][3];
 
-   LP_COUNT(nr_fully_covered_64);
+   if (variant->opaque)
+      return TRUE;
+   
+   if (!variant->potentially_opaque)
+      return FALSE;
 
-   /* if variant is opaque and scissor doesn't effect the tile */
-   if (inputs->opaque) {
-      /* Several things prevent this optimization from working:
-       * - For layered rendering we can't determine if this covers the same layer
-       * as previous rendering (or in case of clears those actually always cover
-       * all layers so optimization is impossible). Need to use fb_max_layer and
-       * not setup->layer_slot to determine this since even if there's currently
-       * no slot assigned previous rendering could have used one.
-       * - If there were any Begin/End query commands in the scene then those
-       * would get removed which would be very wrong. Furthermore, if queries
-       * were just active we also can't do the optimization since to get
-       * accurate query results we unfortunately need to execute the rendering
-       * commands.
-       */
-      if (!scene->fb.zsbuf && scene->fb_max_layer == 0 && !scene->had_queries) {
-         /*
-          * All previous rendering will be overwritten so reset the bin.
-          */
-         lp_scene_bin_reset( scene, tx, ty );
-      }
-
-      LP_COUNT(nr_shade_opaque_64);
-      return lp_scene_bin_cmd_with_state( scene, tx, ty,
-                                          setup->fs.stored,
-                                          LP_RAST_OP_SHADE_TILE_OPAQUE,
-                                          lp_rast_arg_inputs(inputs) );
-   } else {
-      LP_COUNT(nr_shade_64);
-      return lp_scene_bin_cmd_with_state( scene, tx, ty,
-                                          setup->fs.stored,
-                                          LP_RAST_OP_SHADE_TILE,
-                                          lp_rast_arg_inputs(inputs) );
+   if (alpha_info->file == TGSI_FILE_CONSTANT) {
+      const float *constants = setup->fs.current.jit_context.constants[0];
+      float alpha = constants[alpha_info->u.index*4 +
+                              alpha_info->swizzle];
+      return alpha == 1.0f;
    }
+
+   if (alpha_info->file == TGSI_FILE_INPUT) {
+      return (v1[1 + alpha_info->u.index][alpha_info->swizzle] == 1.0f &&
+              v2[1 + alpha_info->u.index][alpha_info->swizzle] == 1.0f &&
+              v3[1 + alpha_info->u.index][alpha_info->swizzle] == 1.0f);
+   }
+
+   return FALSE;
 }
 
 
+
 /**
  * Do basic setup for triangle rasterization and determine which
  * framebuffer tiles are touched.  Put the triangle in the scene's
@@ -333,15 +324,8 @@ do_triangle_ccw(struct lp_setup_context *setup,
       bbox.y1 = (MAX3(position->y[0], position->y[1], position->y[2]) - 1 + adj) >> FIXED_ORDER;
    }
 
-   if (bbox.x1 < bbox.x0 ||
-       bbox.y1 < bbox.y0) {
-      if (0) debug_printf("empty bounding box\n");
-      LP_COUNT(nr_culled_tris);
-      return TRUE;
-   }
-
    if (!u_rect_test_intersection(&setup->draw_regions[viewport_index], &bbox)) {
-      if (0) debug_printf("offscreen\n");
+      if (0) debug_printf("no intersection\n");
       LP_COUNT(nr_culled_tris);
       return TRUE;
    }
@@ -360,9 +344,11 @@ do_triangle_ccw(struct lp_setup_context *setup,
     * Determine how many scissor planes we need, that is drop scissor
     * edges if the bounding box of the tri is fully inside that edge.
     */
-   scissor = &setup->draw_regions[viewport_index];
-   scissor_planes_needed(s_planes, &bboxpos, scissor);
-   nr_planes += s_planes[0] + s_planes[1] + s_planes[2] + s_planes[3];
+   if (setup->scissor_or_vp_clip) {
+      scissor = &setup->draw_regions[viewport_index];
+      scissor_planes_needed(s_planes, &bboxpos, scissor);
+      nr_planes += s_planes[0] + s_planes[1] + s_planes[2] + s_planes[3];
+   }
 
    tri = lp_setup_alloc_triangle(scene,
                                  key->num_inputs,
@@ -382,17 +368,97 @@ do_triangle_ccw(struct lp_setup_context *setup,
 
    LP_COUNT(nr_tris);
 
+   /*
+    * Rotate the tri such that v0 is closest to the fb origin.
+    * This can give more accurate a0 value (which is at fb origin)
+    * when calculating the interpolants.
+    * It can't work when there's flat shading for instance in one
+    * of the attributes, hence restrict this to just a single attribute
+    * which is what causes some test failures.
+    * (This does not address the problem that interpolation may be
+    * inaccurate if gradients are relatively steep in small tris far
+    * away from the origin. It does however fix the (silly) wgf11rasterizer
+    * Interpolator test.)
+    * XXX This causes problems with mipgen -EmuTexture for not yet really
+    * understood reasons (if the vertices would be submitted in a different
+    * order, we'd also generate the same "wrong" results here without
+    * rotation). In any case, that we generate different values if a prim
+    * has the vertices rotated but is otherwise the same (which is due to
+    * numerical issues) is not a nice property. An additional problem by
+    * swapping the vertices here (which is possibly worse) is that
+    * the same primitive coming in twice might generate different values
+    * (in particular for z) due to the swapping potentially not happening
+    * both times, if the attributes to be interpolated are different. For now,
+    * just restrict this to not get used with dx9 (by checking pixel offset),
+    * could also restrict it further to only trigger with wgf11Interpolator
+    * Rasterizer test (the only place which needs it, with always the same
+    * vertices even).
+    */
+   if ((LP_DEBUG & DEBUG_ACCURATE_A0) &&
+       setup->pixel_offset == 0.5f &&
+       key->num_inputs == 1 &&
+       (key->inputs[0].interp == LP_INTERP_LINEAR ||
+        key->inputs[0].interp == LP_INTERP_PERSPECTIVE)) {
+      float dist0 = v0[0][0] * v0[0][0] + v0[0][1] * v0[0][1];
+      float dist1 = v1[0][0] * v1[0][0] + v1[0][1] * v1[0][1];
+      float dist2 = v2[0][0] * v2[0][0] + v2[0][1] * v2[0][1];
+      if (dist0 > dist1 && dist1 < dist2) {
+         const float (*vt)[4];
+         int x, y;
+         vt = v0;
+         v0 = v1;
+         v1 = v2;
+         v2 = vt;
+         x = position->x[0];
+         y = position->y[0];
+         position->x[0] = position->x[1];
+         position->y[0] = position->y[1];
+         position->x[1] = position->x[2];
+         position->y[1] = position->y[2];
+         position->x[2] = x;
+         position->y[2] = y;
+
+         position->dx20 = position->dx01;
+         position->dy20 = position->dy01;
+         position->dx01 = position->x[0] - position->x[1];
+         position->dy01 = position->y[0] - position->y[1];
+      }
+      else if (dist0 > dist2) {
+         const float (*vt)[4];
+         int x, y;
+         vt = v0;
+         v0 = v2;
+         v2 = v1;
+         v1 = vt;
+         x = position->x[0];
+         y = position->y[0];
+         position->x[0] = position->x[2];
+         position->y[0] = position->y[2];
+         position->x[2] = position->x[1];
+         position->y[2] = position->y[1];
+         position->x[1] = x;
+         position->y[1] = y;
+
+         position->dx01 = position->dx20;
+         position->dy01 = position->dy20;
+         position->dx20 = position->x[2] - position->x[0];
+         position->dy20 = position->y[2] - position->y[0];
+      }
+   }
+
    /* Setup parameter interpolants:
     */
    setup->setup.variant->jit_function(v0, v1, v2,
                                       frontfacing,
                                       GET_A0(&tri->inputs),
                                       GET_DADX(&tri->inputs),
-                                      GET_DADY(&tri->inputs));
+                                      GET_DADY(&tri->inputs),
+                                      &setup->setup.variant->key);
 
    tri->inputs.frontfacing = frontfacing;
    tri->inputs.disable = FALSE;
-   tri->inputs.opaque = setup->fs.current.variant->opaque;
+   tri->inputs.is_blit = FALSE;
+   tri->inputs.opaque = check_opaque(setup, v0, v1, v2);
    tri->inputs.layer = layer;
    tri->inputs.viewport_index = viewport_index;
    tri->inputs.view_index = setup->view_index;
@@ -693,7 +759,6 @@ do_triangle_ccw(struct lp_setup_context *setup,
     * (easier to evaluate) to ordinary planes.)
     */
    if (nr_planes > 3) {
-      /* why not just use draw_regions */
       struct lp_rast_plane *plane_s = &plane[3];
 
       if (s_planes[0]) {
@@ -912,8 +977,8 @@ lp_setup_bin_triangle(struct lp_setup_context *setup,
          ystep[i] = ((int64_t)plane[i].dcdy) << TILE_ORDER;
       }
 
-
-
+      tri->inputs.is_blit = lp_setup_is_blit(setup, &tri->inputs);
+     
       /* Test tile-sized blocks against the triangle.
        * Discard blocks fully outside the tri.  If the block is fully
        * contained inside the tri, bin an lp_rast_shade_tile command.
diff --git a/src/gallium/drivers/llvmpipe/lp_setup_vbuf.c b/src/gallium/drivers/llvmpipe/lp_setup_vbuf.c
index fcc27dc7e75..5ed569c4a97 100644
--- a/src/gallium/drivers/llvmpipe/lp_setup_vbuf.c
+++ b/src/gallium/drivers/llvmpipe/lp_setup_vbuf.c
@@ -41,9 +41,16 @@
 #include "draw/draw_vbuf.h"
 #include "draw/draw_vertex.h"
 #include "util/u_memory.h"
+#include "util/u_math.h"
+#include "lp_state_fs.h"
+#include "lp_perf.h"
 
 
-#define LP_MAX_VBUF_INDEXES 1024
+/* It should be a multiple of both 6 and 4 (in other words, a multiple of 12)
+ * to ensure draw splits between a whole number of rectangles.
+ */
+#define LP_MAX_VBUF_INDEXES 1020
+
 #define LP_MAX_VBUF_SIZE    4096
 
   
@@ -135,6 +142,22 @@ static inline const_float4_ptr get_vert( const void *vertex_buffer,
    return (const_float4_ptr)((char *)vertex_buffer + index * stride);
 }
 
+static inline void
+rect(struct lp_setup_context *setup,
+     const float (*v0)[4],
+     const float (*v1)[4],
+     const float (*v2)[4],
+     const float (*v3)[4],
+     const float (*v4)[4],
+     const float (*v5)[4])
+{
+   if (!setup->permit_linear_rasterizer ||
+       !setup->rect( setup, v0, v1, v2, v3, v4, v5)) {
+      setup->triangle(setup, v0, v1, v2);
+      setup->triangle(setup, v3, v4, v5);
+   }
+}
+
 /**
  * draw elements / indexed primitives
  */
@@ -145,6 +168,7 @@ lp_setup_draw_elements(struct vbuf_render *vbr, const ushort *indices, uint nr)
    const unsigned stride = setup->vertex_info->size * sizeof(float);
    const void *vertex_buffer = setup->vertex_buffer;
    const boolean flatshade_first = setup->flatshade_first;
+   boolean uses_constant_interp;
    unsigned i;
 
    assert(setup->setup.variant);
@@ -152,6 +176,8 @@ lp_setup_draw_elements(struct vbuf_render *vbr, const ushort *indices, uint nr)
    if (!lp_setup_update_state(setup, TRUE))
       return;
 
+   uses_constant_interp = setup->setup.variant->key.uses_constant_interp;
+
    switch (setup->prim) {
    case PIPE_PRIM_POINTS:
       for (i = 0; i < nr; i++) {
@@ -190,11 +216,24 @@ lp_setup_draw_elements(struct vbuf_render *vbr, const ushort *indices, uint nr)
       break;
 
    case PIPE_PRIM_TRIANGLES:
-      for (i = 2; i < nr; i += 3) {
-         setup->triangle( setup,
+      if (nr % 6 == 0 && !uses_constant_interp) {
+         for (i = 5; i < nr; i += 6) {
+            rect( setup,
+                          get_vert(vertex_buffer, indices[i-5], stride),
+                          get_vert(vertex_buffer, indices[i-4], stride),
+                          get_vert(vertex_buffer, indices[i-3], stride),
                           get_vert(vertex_buffer, indices[i-2], stride),
                           get_vert(vertex_buffer, indices[i-1], stride),
                           get_vert(vertex_buffer, indices[i-0], stride) );
+         }
+      }
+      else {
+         for (i = 2; i < nr; i += 3) {
+            setup->triangle( setup,
+                             get_vert(vertex_buffer, indices[i-2], stride),
+                             get_vert(vertex_buffer, indices[i-1], stride),
+                             get_vert(vertex_buffer, indices[i-0], stride) );
+         }
       }
       break;
 
@@ -345,11 +384,14 @@ lp_setup_draw_arrays(struct vbuf_render *vbr, uint start, uint nr)
    const void *vertex_buffer =
       (void *) get_vert(setup->vertex_buffer, start, stride);
    const boolean flatshade_first = setup->flatshade_first;
+   boolean uses_constant_interp;
    unsigned i;
 
    if (!lp_setup_update_state(setup, TRUE))
       return;
 
+   uses_constant_interp = setup->setup.variant->key.uses_constant_interp;
+
    switch (setup->prim) {
    case PIPE_PRIM_POINTS:
       for (i = 0; i < nr; i++) {
@@ -388,22 +430,74 @@ lp_setup_draw_arrays(struct vbuf_render *vbr, uint start, uint nr)
       break;
 
    case PIPE_PRIM_TRIANGLES:
-      for (i = 2; i < nr; i += 3) {
-         setup->triangle( setup,
+      if (nr % 6 == 0 && !uses_constant_interp) {
+         for (i = 5; i < nr; i += 6) {
+            rect( setup,
+                          get_vert(vertex_buffer, i-5, stride),
+                          get_vert(vertex_buffer, i-4, stride),
+                          get_vert(vertex_buffer, i-3, stride),
                           get_vert(vertex_buffer, i-2, stride),
                           get_vert(vertex_buffer, i-1, stride),
                           get_vert(vertex_buffer, i-0, stride) );
+         }
+      }
+      else if (!uses_constant_interp &&
+               lp_setup_analyse_triangles(setup, vertex_buffer, stride, nr)) {
+         /* If lp_setup_analyse_triangles() returned true, it also
+          * emitted (setup) the rect or triangles.
+          */
+      }
+      else {
+         for (i = 2; i < nr; i += 3) {
+            setup->triangle( setup,
+                             get_vert(vertex_buffer, i-2, stride),
+                             get_vert(vertex_buffer, i-1, stride),
+                             get_vert(vertex_buffer, i-0, stride) );
+         }
       }
       break;
 
    case PIPE_PRIM_TRIANGLE_STRIP:
       if (flatshade_first) {
-         for (i = 2; i < nr; i++) {
-            /* emit first triangle vertex as first triangle vertex */
-            setup->triangle( setup,
-                             get_vert(vertex_buffer, i-2, stride),
-                             get_vert(vertex_buffer, i+(i&1)-1, stride),
-                             get_vert(vertex_buffer, i-(i&1), stride) );
+         if (!uses_constant_interp) {
+            int j;
+            i = 2;
+            j = 3;
+            while (j < nr) {
+               /* emit first triangle vertex as first triangle vertex */
+               const float (*v0)[4] = get_vert(vertex_buffer, i-2, stride);
+               const float (*v1)[4] = get_vert(vertex_buffer, i+(i&1)-1, stride);
+               const float (*v2)[4] = get_vert(vertex_buffer, i-(i&1), stride);
+               const float (*v3)[4] = get_vert(vertex_buffer, j-2, stride);
+               const float (*v4)[4] = get_vert(vertex_buffer, j+(j&1)-1, stride);
+               const float (*v5)[4] = get_vert(vertex_buffer, j-(j&1), stride);
+               if (setup->permit_linear_rasterizer &&
+                   setup->rect(setup, v0, v1, v2, v3, v4, v5)) {
+                  i += 2;
+                  j += 2;
+               } else {
+                  /* emit one triangle, and retry rectangle in the next one */
+                  setup->triangle(setup, v0, v1, v2);
+                  i += 1;
+                  j += 1;
+               }
+            }
+            if (i < nr) {
+               /* emit last triangle */
+               setup->triangle( setup,
+                                get_vert(vertex_buffer, i-2, stride),
+                                get_vert(vertex_buffer, i+(i&1)-1, stride),
+                                get_vert(vertex_buffer, i-(i&1), stride) );
+            }
+         }
+         else {
+            for (i = 2; i < nr; i++) {
+               /* emit first triangle vertex as first triangle vertex */
+               setup->triangle( setup,
+                                get_vert(vertex_buffer, i-2, stride),
+                                get_vert(vertex_buffer, i+(i&1)-1, stride),
+                                get_vert(vertex_buffer, i-(i&1), stride) );
+            }
          }
       }
       else {
@@ -418,7 +512,16 @@ lp_setup_draw_arrays(struct vbuf_render *vbr, uint start, uint nr)
       break;
 
    case PIPE_PRIM_TRIANGLE_FAN:
-      if (flatshade_first) {
+      if (nr == 4 && !uses_constant_interp) {
+         rect( setup,
+                      get_vert(vertex_buffer, 0, stride),
+                      get_vert(vertex_buffer, 1, stride),
+                      get_vert(vertex_buffer, 2, stride),
+                      get_vert(vertex_buffer, 0, stride),
+                      get_vert(vertex_buffer, 2, stride),
+                      get_vert(vertex_buffer, 3, stride) );
+      }
+      else if (flatshade_first) {
          for (i = 2; i < nr; i += 1) {
             /* emit first non-spoke vertex as first vertex */
             setup->triangle( setup,
@@ -455,15 +558,28 @@ lp_setup_draw_arrays(struct vbuf_render *vbr, uint start, uint nr)
       }
       else {
          /* emit last quad vertex as last triangle vertex */
-         for (i = 3; i < nr; i += 4) {
-            setup->triangle( setup,
+         if (!uses_constant_interp) {
+            for (i = 3; i < nr; i += 4) {
+               rect( setup,
                              get_vert(vertex_buffer, i-3, stride),
                              get_vert(vertex_buffer, i-2, stride),
-                             get_vert(vertex_buffer, i-0, stride) );
-            setup->triangle( setup,
-                             get_vert(vertex_buffer, i-2, stride),
+                             get_vert(vertex_buffer, i-1, stride),
+                             get_vert(vertex_buffer, i-3, stride),
                              get_vert(vertex_buffer, i-1, stride),
                              get_vert(vertex_buffer, i-0, stride) );
+            }
+         }
+         else {
+            for (i = 3; i < nr; i += 4) {
+               setup->triangle( setup,
+                                get_vert(vertex_buffer, i-3, stride),
+                                get_vert(vertex_buffer, i-2, stride),
+                                get_vert(vertex_buffer, i-0, stride) );
+               setup->triangle( setup,
+                                get_vert(vertex_buffer, i-2, stride),
+                                get_vert(vertex_buffer, i-1, stride),
+                                get_vert(vertex_buffer, i-0, stride) );
+            }
          }
       }
       break;
diff --git a/src/gallium/drivers/llvmpipe/lp_state.h b/src/gallium/drivers/llvmpipe/lp_state.h
index 6887d98487d..fab180a8cfb 100644
--- a/src/gallium/drivers/llvmpipe/lp_state.h
+++ b/src/gallium/drivers/llvmpipe/lp_state.h
@@ -116,6 +116,9 @@ llvmpipe_update_fs(struct llvmpipe_context *lp);
 void 
 llvmpipe_update_setup(struct llvmpipe_context *lp);
 
+void
+llvmpipe_update_derived_clear(struct llvmpipe_context *llvmpipe);
+
 void
 llvmpipe_update_derived(struct llvmpipe_context *llvmpipe);
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_blend.c b/src/gallium/drivers/llvmpipe/lp_state_blend.c
index a1ba3e3f5dc..2c4760a574e 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_blend.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_blend.c
@@ -177,6 +177,8 @@ llvmpipe_set_sample_mask(struct pipe_context *pipe,
    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
 
    if (sample_mask != llvmpipe->sample_mask) {
+      draw_flush(llvmpipe->draw);
+
       llvmpipe->sample_mask = sample_mask;
 
       llvmpipe->dirty |= LP_NEW_SAMPLE_MASK;
diff --git a/src/gallium/drivers/llvmpipe/lp_state_derived.c b/src/gallium/drivers/llvmpipe/lp_state_derived.c
index 04899dd9bfd..3baf509a73c 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_derived.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_derived.c
@@ -173,6 +173,73 @@ compute_vertex_info(struct llvmpipe_context *llvmpipe)
 }
 
 
+static void
+check_linear_rasterizer( struct llvmpipe_context *lp )
+{
+   boolean bgr8;
+   boolean permit_linear;
+   boolean single_vp;
+   boolean clipping_changed = FALSE;
+
+   bgr8 = (lp->framebuffer.nr_cbufs == 1 &&
+           lp->framebuffer.cbufs[0]->texture->target == PIPE_TEXTURE_2D &&
+           (lp->framebuffer.cbufs[0]->format == PIPE_FORMAT_B8G8R8A8_UNORM ||
+            lp->framebuffer.cbufs[0]->format == PIPE_FORMAT_B8G8R8X8_UNORM));
+
+   /* permit_linear means guardband, hence fake scissor, which we can only
+    * handle if there's just one vp. */
+   single_vp = lp->viewport_index_slot < 0;
+   permit_linear = (!lp->framebuffer.zsbuf &&
+                    bgr8 &&
+                    single_vp);
+
+   /* Tell draw that we're happy doing our own x/y clipping.
+    */
+   if (lp->permit_linear_rasterizer != permit_linear) {
+      lp->permit_linear_rasterizer = permit_linear;
+      lp_setup_set_linear_mode(lp->setup, permit_linear);
+      clipping_changed = TRUE;
+   }
+
+   if (lp->single_vp != single_vp) {
+      lp->single_vp = single_vp;
+      clipping_changed = TRUE;
+   }
+
+   /* Disable xy clipping in linear mode.
+    *
+    * Use a guard band if we don't have zsbuf.  Could enable
+    * guardband always - this just to be conservative.
+    *
+    * Because we have a layering violation where the draw module emits
+    * state changes to the driver while we're already inside a draw
+    * call, need to be careful about when we make calls back to the
+    * draw module.  Hence the clipping_changed flag which is as much
+    * to prevent flush recursion as it is to short-circuit noop state
+    * changes.
+    */
+   if (clipping_changed) {
+      draw_set_driver_clipping(lp->draw,
+                               FALSE,
+                               FALSE,
+                               permit_linear,
+                               single_vp);
+   }
+}
+
+
+/**
+ * Handle state changes before clears.
+ * Called just prior to clearing (pipe::clear()).
+ */
+void llvmpipe_update_derived_clear( struct llvmpipe_context *llvmpipe )
+{
+   if (llvmpipe->dirty & (LP_NEW_FS |
+                          LP_NEW_FRAMEBUFFER))
+      check_linear_rasterizer(llvmpipe);
+}
+
+
 /**
  * Handle state changes.
  * Called just prior to drawing anything (pipe::draw_arrays(), etc).
@@ -293,6 +360,8 @@ void llvmpipe_update_derived( struct llvmpipe_context *llvmpipe )
                              llvmpipe->viewports);
    }
 
+   llvmpipe_update_derived_clear(llvmpipe);
+
    llvmpipe->dirty = 0;
 }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.c b/src/gallium/drivers/llvmpipe/lp_state_fs.c
index d577990c761..cf36d2d14dc 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.c
@@ -589,7 +589,7 @@ generate_fs_loop(struct gallivm_state *gallivm,
    LLVMValueRef stencil_refs[2];
    LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS][TGSI_NUM_CHANNELS];
    LLVMValueRef zs_samples = lp_build_const_int32(gallivm, key->zsbuf_nr_samples);
-   struct lp_build_for_loop_state loop_state, sample_loop_state;
+   struct lp_build_for_loop_state loop_state, sample_loop_state = {0};
    struct lp_build_mask_context mask;
    /*
     * TODO: figure out if simple_shader optimization is really worthwile to
@@ -3441,6 +3441,24 @@ dump_fs_variant_key(struct lp_fragment_shader_variant_key *key)
    }
 }
 
+const char *
+lp_debug_fs_kind(enum lp_fs_kind kind)
+{
+   switch(kind) {
+   case LP_FS_KIND_GENERAL:
+      return "GENERAL";
+   case LP_FS_KIND_BLIT_RGBA:
+      return "BLIT_RGBA";
+   case LP_FS_KIND_BLIT_RGB1:
+      return "BLIT_RGB1";
+   case LP_FS_KIND_AERO_MINIFICATION:
+      return "AERO_MINIFICATION";
+   case LP_FS_KIND_LLVM_LINEAR:
+      return "LLVM_LINEAR";
+   default:
+      return "unknown";
+   }
+}
 
 void
 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
@@ -3453,6 +3471,9 @@ lp_debug_fs_variant(struct lp_fragment_shader_variant *variant)
       nir_print_shader(variant->shader->base.ir.nir, stderr);
    dump_fs_variant_key(&variant->key);
    debug_printf("variant->opaque = %u\n", variant->opaque);
+   debug_printf("variant->potentially_opaque = %u\n", variant->potentially_opaque);
+   debug_printf("variant->blit = %u\n", variant->blit);
+   debug_printf("shader->kind = %s\n", lp_debug_fs_kind(variant->shader->kind));
    debug_printf("\n");
 }
 
@@ -3491,6 +3512,8 @@ generate_variant(struct llvmpipe_context *lp,
    struct lp_fragment_shader_variant *variant;
    const struct util_format_description *cbuf0_format_desc = NULL;
    boolean fullcolormask;
+   boolean no_kill;
+   boolean linear;
    char module_name[64];
    unsigned char ir_sha1_cache_key[20];
    struct lp_cached_code cached = { 0 };
@@ -3536,9 +3559,9 @@ generate_variant(struct llvmpipe_context *lp,
       fullcolormask = util_format_colormask_full(cbuf0_format_desc, key->blend.rt[0].colormask);
    }
 
-   variant->opaque =
-         !key->blend.logicop_enable &&
-         !key->blend.rt[0].blend_enable &&
+   /* The scissor is ignored here as only tiles inside the scissoring
+    * rectangle will refer to this */
+   no_kill =
          fullcolormask &&
          !key->stencil[0].enabled &&
          !key->alpha.enabled &&
@@ -3546,13 +3569,82 @@ generate_variant(struct llvmpipe_context *lp,
          !key->blend.alpha_to_coverage &&
          !key->depth.enabled &&
          !shader->info.base.uses_kill &&
-         !shader->info.base.writes_samplemask
-      ? TRUE : FALSE;
+         !shader->info.base.writes_samplemask;
+
+   variant->opaque =
+         no_kill &&
+         !key->blend.logicop_enable &&
+         !key->blend.rt[0].blend_enable
+         ? TRUE : FALSE;
+
+   variant->potentially_opaque =
+         no_kill &&
+         !key->blend.logicop_enable &&
+         key->blend.rt[0].blend_enable &&
+         key->blend.rt[0].rgb_func == PIPE_BLEND_ADD &&
+         key->blend.rt[0].rgb_dst_factor == PIPE_BLENDFACTOR_INV_SRC_ALPHA &&
+         key->blend.rt[0].alpha_func == key->blend.rt[0].rgb_func &&
+         key->blend.rt[0].alpha_dst_factor == key->blend.rt[0].rgb_dst_factor &&
+         shader->base.type == PIPE_SHADER_IR_TGSI &&
+         /*
+          * FIXME: for NIR, all of the fields of info.xxx (except info.base)
+          * are zeros, hence shader analysis (here and elsewhere) using these
+          * bits cannot work and will silently fail (cbuf is the only pointer
+          * field, hence causing a crash).
+          */
+         shader->info.cbuf[0][3].file != TGSI_FILE_NULL
+         ? TRUE : FALSE;
+
+   /* We only care about opaque blits for now */
+   if (variant->opaque &&
+       (shader->kind == LP_FS_KIND_BLIT_RGBA ||
+        shader->kind == LP_FS_KIND_BLIT_RGB1)) {
+      unsigned target, min_img_filter, mag_img_filter, min_mip_filter;
+      enum pipe_format texture_format;
+
+      texture_format = key->samplers[0].texture_state.format;
+      target = key->samplers[0].texture_state.target;
+      min_img_filter = key->samplers[0].sampler_state.min_img_filter;
+      mag_img_filter = key->samplers[0].sampler_state.mag_img_filter;
+      if (key->samplers[0].texture_state.level_zero_only) {
+         min_mip_filter = PIPE_TEX_MIPFILTER_NONE;
+      } else {
+         min_mip_filter = key->samplers[0].sampler_state.min_mip_filter;
+      }
+
+      if (target == PIPE_TEXTURE_2D &&
+          min_img_filter == PIPE_TEX_FILTER_NEAREST &&
+          mag_img_filter == PIPE_TEX_FILTER_NEAREST &&
+          min_mip_filter == PIPE_TEX_MIPFILTER_NONE &&
+          ((texture_format &&
+            util_is_format_compatible(util_format_description(texture_format),
+                                      cbuf0_format_desc)) ||
+           (shader->kind == LP_FS_KIND_BLIT_RGB1 &&
+            (texture_format == PIPE_FORMAT_B8G8R8A8_UNORM ||
+             texture_format == PIPE_FORMAT_B8G8R8X8_UNORM) &&
+            (key->cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM ||
+             key->cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM))))
+         variant->blit = 1;
+   }
+
+
+   /* Whether this is a candidate for the linear path */
+   linear =
+         !key->stencil[0].enabled &&
+         !key->depth.enabled &&
+         !shader->info.base.uses_kill &&
+         !key->blend.logicop_enable &&
+         (key->cbuf_format[0] == PIPE_FORMAT_B8G8R8A8_UNORM ||
+          key->cbuf_format[0] == PIPE_FORMAT_B8G8R8X8_UNORM);
+
+   memcpy(&variant->key, key, sizeof *key);
 
    if ((LP_DEBUG & DEBUG_FS) || (gallivm_debug & GALLIVM_DEBUG_IR)) {
       lp_debug_fs_variant(variant);
    }
 
+   llvmpipe_fs_variant_fastpath(variant);
+
    lp_jit_init_types(variant);
    
    if (variant->jit_function[RAST_EDGE_TEST] == NULL)
@@ -3565,6 +3657,36 @@ generate_variant(struct llvmpipe_context *lp,
       }
    }
 
+   if (linear) {
+      /* Currently keeping both the old fastpaths and new linear path
+       * active.  The older code is still somewhat faster for the cases
+       * it covers.
+       *
+       * XXX: consider restricting this to aero-mode only.
+       */
+      if (fullcolormask &&
+          !key->alpha.enabled &&
+          !key->blend.alpha_to_coverage) {
+         llvmpipe_fs_variant_linear_fastpath(variant);
+      }
+
+      /* If the original fastpath doesn't cover this variant, try the new
+       * code:
+       */
+      if (variant->jit_linear == NULL) {
+         if (shader->kind == LP_FS_KIND_BLIT_RGBA ||
+             shader->kind == LP_FS_KIND_BLIT_RGB1 ||
+             shader->kind == LP_FS_KIND_LLVM_LINEAR) {
+            llvmpipe_fs_variant_linear_llvm(lp, shader, variant);
+         }
+      }
+   } else {
+      if (LP_DEBUG & DEBUG_LINEAR) {
+         lp_debug_fs_variant(variant);
+         debug_printf("    ----> no linear path for this variant\n");
+      }
+   }
+
    /*
     * Compile everything
     */
@@ -3587,6 +3709,19 @@ generate_variant(struct llvmpipe_context *lp,
       variant->jit_function[RAST_WHOLE] = variant->jit_function[RAST_EDGE_TEST];
    }
 
+   if (linear) {
+      if (variant->linear_function) {
+         variant->jit_linear_llvm = (lp_jit_linear_llvm_func)
+               gallivm_jit_function(variant->gallivm, variant->linear_function);
+      }
+
+      /*
+       * This must be done after LLVM compilation, as it will call the JIT'ed
+       * code to determine active inputs.
+       */
+      lp_linear_check_variant(variant);
+   }
+
    if (needs_caching) {
       lp_disk_cache_insert_shader(screen, &cached, ir_sha1_cache_key);
    }
@@ -3697,6 +3832,9 @@ llvmpipe_create_fs_state(struct pipe_context *pipe,
       debug_printf("\n");
    }
 
+   /* This will put a derived copy of the tokens into shader->base.tokens */
+   llvmpipe_fs_analyse(shader, templ->tokens);
+
    return shader;
 }
 
@@ -4155,11 +4293,16 @@ make_variant_key(struct llvmpipe_context *lp,
                                                &lp->images[PIPE_SHADER_FRAGMENT][i]);
       }
    }
+
+   if (shader->kind == LP_FS_KIND_AERO_MINIFICATION) {
+      key->samplers[0].sampler_state.min_img_filter = PIPE_TEX_FILTER_NEAREST;
+      key->samplers[0].sampler_state.mag_img_filter = PIPE_TEX_FILTER_NEAREST;
+   }
+
    return key;
 }
 
 
-
 /**
  * Update fragment shader state.  This is called just prior to drawing
  * something when some fragment-related state has changed.
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs.h b/src/gallium/drivers/llvmpipe/lp_state_fs.h
index d5dd49901e6..9f43665d839 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_fs.h
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs.h
@@ -48,6 +48,16 @@ struct lp_fragment_shader;
 #define RAST_EDGE_TEST 1
 
 
+enum lp_fs_kind
+{
+   LP_FS_KIND_GENERAL = 0,
+   LP_FS_KIND_BLIT_RGBA,
+   LP_FS_KIND_BLIT_RGB1,
+   LP_FS_KIND_AERO_MINIFICATION,
+   LP_FS_KIND_LLVM_LINEAR
+};
+
+
 struct lp_sampler_static_state
 {
    /*
@@ -137,6 +147,13 @@ struct lp_fs_variant_list_item
 
 struct lp_fragment_shader_variant
 {
+   /*
+    * Whether some primitives can be opaque.
+    */
+   unsigned potentially_opaque:1;
+
+   unsigned blit:1;
+   unsigned linear_input_mask:16;
    struct pipe_reference reference;
    boolean opaque;
 
@@ -150,6 +167,17 @@ struct lp_fragment_shader_variant
 
    lp_jit_frag_func jit_function[2];
 
+   lp_jit_linear_func jit_linear;
+   lp_jit_linear_func jit_linear_blit;
+
+   /* Functions within the linear path:
+    */
+   LLVMValueRef linear_function;
+   lp_jit_linear_llvm_func jit_linear_llvm;
+
+   /* Bitmask to say what cbufs are unswizzled */
+   unsigned unswizzled_cbufs;
+
    /* Total number of LLVM instructions generated */
    unsigned nr_instrs;
 
@@ -172,6 +200,13 @@ struct lp_fragment_shader
    struct pipe_reference reference;
    struct lp_tgsi_info info;
 
+   /*
+    * Analysis results
+    */
+
+   enum lp_fs_kind kind;
+
+
    struct lp_fs_variant_list_item variants;
 
    struct draw_fragment_shader *draw_data;
@@ -187,9 +222,31 @@ struct lp_fragment_shader
 };
 
 
+void
+llvmpipe_fs_analyse(struct lp_fragment_shader *shader,
+                    const struct tgsi_token *tokens);
+
+void
+llvmpipe_fs_variant_fastpath(struct lp_fragment_shader_variant *variant);
+
+void
+llvmpipe_fs_variant_linear_fastpath(struct lp_fragment_shader_variant *variant);
+
+void
+llvmpipe_fs_variant_linear_llvm(struct llvmpipe_context *lp,
+                                struct lp_fragment_shader *shader,
+                                struct lp_fragment_shader_variant *variant);
+
 void
 lp_debug_fs_variant(struct lp_fragment_shader_variant *variant);
 
+const char *
+lp_debug_fs_kind(enum lp_fs_kind kind);
+
+
+void
+lp_linear_check_variant(struct lp_fragment_shader_variant *variant);
+
 void
 llvmpipe_destroy_fs(struct llvmpipe_context *llvmpipe,
                     struct lp_fragment_shader *shader);
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs_analysis.c b/src/gallium/drivers/llvmpipe/lp_state_fs_analysis.c
new file mode 100644
index 00000000000..cd397e10824
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs_analysis.c
@@ -0,0 +1,204 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "util/u_memory.h"
+#include "util/u_math.h"
+#include "tgsi/tgsi_parse.h"
+#include "tgsi/tgsi_text.h"
+#include "tgsi/tgsi_util.h"
+#include "tgsi/tgsi_dump.h"
+#include "lp_debug.h"
+#include "lp_state.h"
+
+
+/*
+ * Detect Aero minification shaders.
+ *
+ * Aero does not use texture mimaps when a window gets animated and its shaped
+ * bended. Instead it uses the average of 4 nearby texels. This is the simplest
+ * of such shader, but there are several variations:
+ *
+ *   FRAG
+ *   DCL IN[0], GENERIC[1], PERSPECTIVE
+ *   DCL IN[1], GENERIC[2], PERSPECTIVE
+ *   DCL IN[2], GENERIC[3], PERSPECTIVE
+ *   DCL OUT[0], COLOR
+ *   DCL SAMP[0]
+ *   DCL TEMP[0..3]
+ *   IMM FLT32 {     0.2500,     0.0000,     0.0000,     0.0000 }
+ *   MOV TEMP[0].x, IN[0].zzzz
+ *   MOV TEMP[0].y, IN[0].wwww
+ *   MOV TEMP[1].x, IN[1].zzzz
+ *   MOV TEMP[1].y, IN[1].wwww
+ *   TEX TEMP[0], TEMP[0], SAMP[0], 2D
+ *   TEX TEMP[2], IN[0], SAMP[0], 2D
+ *   TEX TEMP[3], IN[1], SAMP[0], 2D
+ *   TEX TEMP[1], TEMP[1], SAMP[0], 2D
+ *   ADD TEMP[0], TEMP[0], TEMP[2]
+ *   ADD TEMP[0], TEMP[3], TEMP[0]
+ *   ADD TEMP[0], TEMP[1], TEMP[0]
+ *   MUL TEMP[0], TEMP[0], IN[2]
+ *   MUL TEMP[0], TEMP[0], IMM[0].xxxx
+ *   MOV OUT[0], TEMP[0]
+ *   END
+ *
+ * Texture coordinates are interleaved like the Gaussian blur shaders, but
+ * unlike the later there isn't structure in the sub-pixel positioning of the
+ * texels, other than being disposed in a diamond-like shape. For example,
+ * these are the relative offsets of the texels relative to the average:
+ *
+ *    x offset   y offset
+ *   --------------------
+ *    0.691834   -0.21360
+ *   -0.230230   -0.64160
+ *   -0.692406    0.21356
+ *    0.230802    0.64160
+ *
+ *  These shaders are typically used with linear min/mag filtering, but the
+ *  linear filtering provides very little visual improvement compared to the
+ *  performance impact it has. The ultimate purpose of detecting these shaders
+ *  is to override with nearest texture filtering.
+ */
+static inline boolean
+match_aero_minification_shader(const struct tgsi_token *tokens,
+                               const struct lp_tgsi_info *info)
+{
+   struct tgsi_parse_context parse;
+   unsigned coord_mask;
+   boolean has_quarter_imm;
+   unsigned index, chan;
+
+   if ((info->base.opcode_count[TGSI_OPCODE_TEX] != 4 &&
+        info->base.opcode_count[TGSI_OPCODE_SAMPLE] != 4) ||
+       info->num_texs != 4) {
+      return FALSE;
+   }
+
+   /*
+    * Ensure the texture coordinates are interleaved as in the example above.
+    */
+
+   coord_mask = 0;
+   for (index = 0; index < 4; ++index) {
+      const struct lp_tgsi_texture_info *tex = &info->tex[index];
+      if (tex->sampler_unit != 0 ||
+          tex->texture_unit != 0 ||
+          tex->coord[0].file != TGSI_FILE_INPUT ||
+          tex->coord[1].file != TGSI_FILE_INPUT ||
+          tex->coord[0].u.index != tex->coord[1].u.index ||
+          (tex->coord[0].swizzle % 2) != 0 ||
+          tex->coord[1].swizzle != tex->coord[0].swizzle + 1) {
+         return FALSE;
+      }
+
+      coord_mask |= 1 << (tex->coord[0].u.index*2 + tex->coord[0].swizzle/2);
+   }
+   if (coord_mask != 0xf) {
+      return FALSE;
+   }
+
+   /*
+    * Ensure it has the 0.25 immediate.
+    */
+
+   has_quarter_imm = FALSE;
+
+   tgsi_parse_init(&parse, tokens);
+
+   while (!tgsi_parse_end_of_tokens(&parse)) {
+      tgsi_parse_token(&parse);
+
+      switch (parse.FullToken.Token.Type) {
+      case TGSI_TOKEN_TYPE_DECLARATION:
+         break;
+
+      case TGSI_TOKEN_TYPE_INSTRUCTION:
+         goto finished;
+
+      case TGSI_TOKEN_TYPE_IMMEDIATE:
+         {
+            const unsigned size =
+                  parse.FullToken.FullImmediate.Immediate.NrTokens - 1;
+            assert(size <= 4);
+            for (chan = 0; chan < size; ++chan) {
+               if (parse.FullToken.FullImmediate.u[chan].Float == 0.25f) {
+                  has_quarter_imm = TRUE;
+                  goto finished;
+               }
+            }
+         }
+         break;
+
+      case TGSI_TOKEN_TYPE_PROPERTY:
+         break;
+
+      default:
+         assert(0);
+         goto finished;
+      }
+   }
+finished:
+
+   tgsi_parse_free(&parse);
+
+   if (!has_quarter_imm) {
+      return FALSE;
+   }
+
+   return TRUE;
+}
+
+
+void
+llvmpipe_fs_analyse(struct lp_fragment_shader *shader,
+                    const struct tgsi_token *tokens)
+{
+   shader->kind = LP_FS_KIND_GENERAL;
+
+   if (shader->kind == LP_FS_KIND_GENERAL &&
+       shader->info.base.num_inputs <= LP_MAX_LINEAR_INPUTS &&
+       shader->info.base.num_outputs == 1 &&
+       !shader->info.indirect_textures &&
+       !shader->info.sampler_texture_units_different &&
+       !shader->info.unclamped_immediates &&
+       shader->info.num_texs <= LP_MAX_LINEAR_TEXTURES &&
+       (shader->info.base.opcode_count[TGSI_OPCODE_TEX] +
+        shader->info.base.opcode_count[TGSI_OPCODE_SAMPLE] +
+        shader->info.base.opcode_count[TGSI_OPCODE_MOV] +
+        shader->info.base.opcode_count[TGSI_OPCODE_MUL] +
+        shader->info.base.opcode_count[TGSI_OPCODE_RET] +
+        shader->info.base.opcode_count[TGSI_OPCODE_END] ==
+        shader->info.base.num_instructions)) {
+      shader->kind = LP_FS_KIND_LLVM_LINEAR;
+   }
+
+   if (shader->kind == LP_FS_KIND_GENERAL &&
+       match_aero_minification_shader(tokens, &shader->info)) {
+      shader->kind = LP_FS_KIND_AERO_MINIFICATION;
+   }
+}
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs_fastpath.c b/src/gallium/drivers/llvmpipe/lp_state_fs_fastpath.c
new file mode 100644
index 00000000000..3fbb9142c14
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs_fastpath.c
@@ -0,0 +1,238 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "pipe/p_config.h"
+
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_sse.h"
+
+#include "lp_jit.h"
+#include "lp_state_fs.h"
+#include "lp_debug.h"
+
+
+#if defined(PIPE_ARCH_SSE)
+
+#include <emmintrin.h>
+
+
+static void
+no_op(const struct lp_jit_context *context,
+      uint32_t x,
+      uint32_t y,
+      uint32_t facing,
+      const void *a0,
+      const void *dadx,
+      const void *dady,
+      uint8_t **cbufs,
+      uint8_t *depth,
+      uint64_t mask,
+      struct lp_jit_thread_data *thread_data,
+      unsigned *strides,
+      unsigned depth_stride,
+      unsigned *color_sample_stride,
+      unsigned depth_sample_stride)
+{
+}
+
+
+/*
+ * m ? a : b
+ */
+static inline __m128i
+mm_select_si128(__m128i m, __m128i a, __m128i b)
+{
+   __m128i res;
+
+   /*
+    * TODO: use PBLENVB when available.
+    */
+
+   res = _mm_or_si128(_mm_and_si128(m, a),
+                      _mm_andnot_si128(m, b));
+
+   return res;
+}
+
+
+/*
+ * *p = m ? a : *p;
+ */
+static inline void
+mm_store_mask_si128(__m128i *p, __m128i m, __m128i a)
+{
+   _mm_store_si128(p, mm_select_si128(m, a, _mm_load_si128(p)));
+}
+
+
+/**
+ * Expand the mask from a 16 bit integer to a 4 x 4 x 32 bit vector mask, ie.
+ * 1 bit -> 32bits.
+ */
+static inline void
+expand_mask(uint32_t int_mask,
+            __m128i *vec_mask)
+{
+   __m128i inv_mask = _mm_set1_epi32(~int_mask & 0xffff);
+   __m128i zero = _mm_setzero_si128();
+
+   vec_mask[0] = _mm_and_si128(inv_mask, _mm_setr_epi32(0x0001, 0x0002, 0x0004, 0x0008));
+   vec_mask[1] = _mm_and_si128(inv_mask, _mm_setr_epi32(0x0010, 0x0020, 0x0040, 0x0080));
+   inv_mask = _mm_srli_epi32(inv_mask, 8);
+   vec_mask[2] = _mm_and_si128(inv_mask, _mm_setr_epi32(0x0001, 0x0002, 0x0004, 0x0008));
+   vec_mask[3] = _mm_and_si128(inv_mask, _mm_setr_epi32(0x0010, 0x0020, 0x0040, 0x0080));
+
+   vec_mask[0] = _mm_cmpeq_epi32(vec_mask[0], zero);
+   vec_mask[1] = _mm_cmpeq_epi32(vec_mask[1], zero);
+   vec_mask[2] = _mm_cmpeq_epi32(vec_mask[2], zero);
+   vec_mask[3] = _mm_cmpeq_epi32(vec_mask[3], zero);
+}
+
+
+/**
+ * Draw opaque color (for debugging).
+ */
+static void
+opaque_color(uint8_t **cbufs, unsigned *strides,
+             uint32_t int_mask,
+             uint32_t color)
+{
+   __m128i *cbuf = (__m128i *)cbufs[0];
+   unsigned stride = strides[0] / sizeof *cbuf;
+   __m128i vec_mask[4];
+   __m128i vec_color = _mm_set1_epi32(color);
+
+   expand_mask(int_mask, vec_mask);
+
+   mm_store_mask_si128(cbuf, vec_mask[0], vec_color); cbuf += stride;
+   mm_store_mask_si128(cbuf, vec_mask[1], vec_color); cbuf += stride;
+   mm_store_mask_si128(cbuf, vec_mask[2], vec_color); cbuf += stride;
+   mm_store_mask_si128(cbuf, vec_mask[3], vec_color);
+}
+
+
+/**
+ * Draw opaque red (for debugging).
+ */
+static void
+red(const struct lp_jit_context *context,
+    uint32_t x,
+    uint32_t y,
+    uint32_t facing,
+    const void *a0,
+    const void *dadx,
+    const void *dady,
+    uint8_t **cbufs,
+    uint8_t *depth,
+    uint64_t int_mask,
+    struct lp_jit_thread_data *thread_data,
+    unsigned *strides,
+    unsigned depth_stride,
+    unsigned *sample_stride,
+    unsigned depth_sample_stride)
+{
+   opaque_color(cbufs, strides, int_mask, 0xffff0000);
+   (void)facing;
+   (void)depth;
+   (void)thread_data;
+}
+
+
+/**
+ * Draw opaque green (for debugging).
+ */
+static void
+green(const struct lp_jit_context *context,
+      uint32_t x,
+      uint32_t y,
+      uint32_t facing,
+      const void *a0,
+      const void *dadx,
+      const void *dady,
+      uint8_t **cbufs,
+      uint8_t *depth,
+      uint64_t int_mask,
+      struct lp_jit_thread_data *thread_data,
+      unsigned *strides,
+      unsigned depth_stride,
+      unsigned *sample_stride,
+      unsigned depth_sample_stride)
+{
+   opaque_color(cbufs, strides, int_mask, 0xff00ff00);
+   (void)facing;
+   (void)depth;
+   (void)thread_data;
+}
+
+
+void
+llvmpipe_fs_variant_fastpath(struct lp_fragment_shader_variant *variant)
+{
+   variant->jit_function[RAST_WHOLE]     = NULL;
+   variant->jit_function[RAST_EDGE_TEST] = NULL;
+
+   if (LP_DEBUG & DEBUG_NO_FASTPATH)
+      return;
+
+   if (variant->key.cbuf_format[0] != PIPE_FORMAT_B8G8R8A8_UNORM &&
+       variant->key.cbuf_format[0] != PIPE_FORMAT_B8G8R8X8_UNORM) {
+      return;
+   }
+
+   if (0) {
+      variant->jit_function[RAST_WHOLE]     = red;
+      variant->jit_function[RAST_EDGE_TEST] = red;
+   }
+
+   if (0) {
+      variant->jit_function[RAST_WHOLE]     = green;
+      variant->jit_function[RAST_EDGE_TEST] = green;
+   }
+
+   if (0) {
+      variant->jit_function[RAST_WHOLE]     = no_op;
+      variant->jit_function[RAST_EDGE_TEST] = no_op;
+   }
+
+   /* Make it easier to see triangles:
+    */
+   if ((LP_DEBUG & DEBUG_LINEAR) || (LP_PERF & PERF_NO_SHADE)) {
+      variant->jit_function[RAST_EDGE_TEST] = red;
+      variant->jit_function[RAST_WHOLE] = green;
+   }
+}
+
+#else
+
+void
+llvmpipe_fs_variant_fastpath(struct lp_fragment_shader_variant *variant)
+{
+}
+
+#endif
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs_linear.c b/src/gallium/drivers/llvmpipe/lp_state_fs_linear.c
new file mode 100644
index 00000000000..a8d10ae300d
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs_linear.c
@@ -0,0 +1,713 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL
+ * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM,
+ * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR
+ * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
+ * USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ **************************************************************************/
+
+
+#include "pipe/p_config.h"
+
+#include "util/u_math.h"
+#include "util/u_cpu_detect.h"
+#include "util/u_pack_color.h"
+#include "util/u_surface.h"
+#include "util/u_sse.h"
+
+#include "lp_jit.h"
+#include "lp_rast.h"
+#include "lp_debug.h"
+#include "lp_state_fs.h"
+#include "lp_linear_priv.h"
+
+
+#if defined(PIPE_ARCH_SSE)
+
+#include <emmintrin.h>
+
+
+struct nearest_sampler {
+   PIPE_ALIGN_VAR(16) uint32_t out[64];
+
+   const struct lp_jit_texture *texture;
+   float fsrc_x;                /* src_x0 */
+   float fsrc_y;                /* src_y0 */
+   float fdsdx;              /* sx */
+   float fdsdy;              /* sx */
+   float fdtdx;              /* sy */
+   float fdtdy;              /* sy */
+   int width;
+   int y;
+
+   const uint32_t *(*fetch)(struct nearest_sampler *samp);
+};
+
+
+struct linear_interp {
+   PIPE_ALIGN_VAR(16) uint32_t out[64];
+   __m128i a0;
+   __m128i dadx;
+   __m128i dady;
+   int width;                   /* rounded up to multiple of 4 */
+   boolean is_constant;
+};
+
+/* Organize all the information needed for blending in one place.
+ * Could have blend function pointer here, but we currently always
+ * know which one we want to call.
+ */
+struct color_blend {
+   const uint32_t *src;
+   uint8_t *color;
+   int stride;
+   int width;                   /* the exact width */
+};
+
+
+/* Organize all the information needed for running each of the shaders
+ * in one place.
+ */
+struct shader {
+   PIPE_ALIGN_VAR(16) uint32_t out0[64];
+   const uint32_t *src0;
+   const uint32_t *src1;
+   __m128i const0;
+   int width;                   /* rounded up to multiple of 4 */
+};
+
+
+/* For a row of pixels, perform add/one/inv_src_alpha (ie
+ * premultiplied alpha) blending between the incoming pixels and the
+ * destination buffer.
+ *
+ * Used to implement the BLIT_RGBA + blend shader, there are no
+ * operations from the pixel shader left to implement at this level -
+ * effectively the pixel shader was just a texture fetch which has
+ * already been performed.  This routine then purely implements
+ * blending.
+ */
+static void
+blend_premul(struct color_blend *blend)
+{
+   const uint32_t *src = blend->src;  /* aligned */
+   uint32_t *dst = (uint32_t *)blend->color;      /* unaligned */
+   int width = blend->width;
+   int i;
+   __m128i tmp;
+   union { __m128i m128; uint ui[4]; } dstreg;
+
+   blend->color += blend->stride;
+
+   for (i = 0; i + 3 < width; i += 4) {
+      tmp = _mm_loadu_si128((const __m128i *)&dst[i]);  /* UNALIGNED READ */
+      dstreg.m128 = util_sse2_blend_premul_4(*(const __m128i *)&src[i],
+                                             tmp);
+      _mm_storeu_si128((__m128i *)&dst[i], dstreg.m128); /* UNALIGNED WRITE */
+   }
+
+   if (i < width) {
+      int j;
+      for (j = 0; j < width - i ; j++) {
+         dstreg.ui[j] = dst[i+j];
+      }
+      dstreg.m128 = util_sse2_blend_premul_4(*(const __m128i *)&src[i],
+                                             dstreg.m128);
+      for (; i < width; i++)
+         dst[i] = dstreg.ui[i&3];
+   }
+}
+
+
+static void
+blend_noop(struct color_blend *blend)
+{
+   memcpy(blend->color, blend->src, blend->width * sizeof(unsigned));
+   blend->color += blend->stride;
+}
+
+
+static void
+init_blend(struct color_blend *blend,
+           int x, int y, int width, int height,
+           uint8_t *color,
+           int stride)
+{
+   blend->color = color + x * 4 + y * stride;
+   blend->stride = stride;
+   blend->width = width;
+}
+
+
+/*
+ * Perform nearest filtered lookup of a row of texels.  Texture lookup
+ * is assumed to be axis aligned but with arbitrary scaling.
+ *
+ * Texture coordinate interpolation is performed in 24.8 fixed point.
+ * Note that the longest span we will encounter is 64 pixels long,
+ * meaning that 8 fractional bits is more than sufficient to represent
+ * the shallowest gradient possible within this span.
+ *
+ * After 64 pixels (ie. in the next tile), the starting point will be
+ * recalculated with floating point arithmetic.
+ *
+ * XXX: migrate this to use Jose's quad blitter texture fetch routines.
+ */
+static const uint32_t *
+fetch_row(struct nearest_sampler *samp)
+{
+   int y = samp->y++;
+   uint32_t *row = samp->out;
+   const struct lp_jit_texture *texture = samp->texture;
+   int yy = util_iround(samp->fsrc_y + samp->fdtdy * y);
+   const uint32_t *src_row =
+      (const uint32_t *)((const uint8_t *)texture->base +
+                         yy * texture->row_stride[0]);
+   int iscale_x = samp->fdsdx * 256;
+   int acc      = samp->fsrc_x * 256 + 128;
+   int width    = samp->width;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      row[i] = src_row[acc>>8];
+      acc += iscale_x;
+   }
+
+   return row;
+}
+
+/* Version of fetch_row which can cope with texture edges.  In
+ * practise, aero never triggers this.
+ */
+static const uint32_t *
+fetch_row_clamped(struct nearest_sampler *samp)
+{
+   int y = samp->y++;
+   uint32_t *row = samp->out;
+   const struct lp_jit_texture *texture = samp->texture;
+
+   int yy = util_iround(samp->fsrc_y + samp->fdtdy * y);
+
+   const uint32_t *src_row =
+      (const uint32_t *)((const uint8_t *)texture->base +
+                         CLAMP(yy, 0, texture->height-1) *
+                         texture->row_stride[0]);
+   float src_x0 = samp->fsrc_x;
+   float scale_x = samp->fdsdx;
+   int width    = samp->width;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      row[i] = src_row[CLAMP(util_iround(src_x0 + i*scale_x),0,texture->width-1)];
+   }
+
+   return row;
+}
+
+/* It vary rarely happens that some non-axis-aligned texturing creeps
+ * into the linear path.  Handle it here.  The alternative would be
+ * more pre-checking or an option to fallback by returning false from
+ * jit_linear.
+ */
+static const uint32_t *
+fetch_row_xy_clamped(struct nearest_sampler *samp)
+{
+   int y = samp->y++;
+   uint32_t *row = samp->out;
+   const struct lp_jit_texture *texture = samp->texture;
+   float yrow = samp->fsrc_y + samp->fdtdy * y;
+   float xrow = samp->fsrc_x + samp->fdsdy * y;
+   int width  = samp->width;
+   int i;
+
+   for (i = 0; i < width; i++) {
+      int yy = util_iround(yrow + samp->fdtdx * i);
+      int xx = util_iround(xrow + samp->fdsdx * i);
+
+      const uint32_t *src_row =
+         (const uint32_t *)((const uint8_t *)texture->base +
+                            CLAMP(yy, 0, texture->height-1) *
+                            texture->row_stride[0]);
+
+      row[i] = src_row[CLAMP(xx,0,texture->width-1)];
+   }
+
+   return row;
+}
+
+
+static boolean
+init_nearest_sampler(struct nearest_sampler *samp,
+                     const struct lp_jit_texture *texture,
+                     int x0, int y0,
+                     int width, int height,
+                     float s0, float dsdx, float dsdy,
+                     float t0, float dtdx, float dtdy,
+                     float w0, float dwdx, float dwdy)
+{
+   int i;
+   float oow = 1.0f / w0;
+
+   if (dwdx != 0.0 || dwdy != 0.0)
+      return FALSE;
+
+   samp->texture = texture;
+   samp->width = width;
+   samp->fdsdx = dsdx * texture->width * oow;
+   samp->fdsdy = dsdy * texture->width * oow;
+   samp->fdtdx = dtdx * texture->height * oow;
+   samp->fdtdy = dtdy * texture->height * oow;
+   samp->fsrc_x = (samp->fdsdx * x0 +
+                   samp->fdsdy * y0 +
+                   s0 * texture->width * oow - 0.5f);
+
+   samp->fsrc_y = (samp->fdtdx * x0 +
+                   samp->fdtdy * y0 +
+                   t0 * texture->height * oow - 0.5f);
+   samp->y = 0;
+
+   /* Because we want to permit consumers of this data to round up to
+    * the next multiple of 4, and because we don't want valgrind to
+    * complain about uninitialized reads, set the last bit of the
+    * buffer to zero:
+    */
+   for (i = width; i & 3; i++)
+      samp->out[i] = 0;
+
+   if (dsdy != 0 || dtdx != 0)
+   {
+      /* Arbitrary texture lookup:
+       */
+      samp->fetch = fetch_row_xy_clamped;
+   }
+   else
+   {
+      /* Axis aligned stretch blit, abitrary scaling factors including
+       * flipped, minifying and magnifying:
+       */
+      int isrc_x = util_iround(samp->fsrc_x);
+      int isrc_y = util_iround(samp->fsrc_y);
+      int isrc_x1 = util_iround(samp->fsrc_x + width * samp->fdsdx);
+      int isrc_y1 = util_iround(samp->fsrc_y + height * samp->fdtdy);
+
+      /* Look at the maximum and minimum texture coordinates we will be
+       * fetching and figure out if we need to use clamping.  There is
+       * similar code in u_blit_sw.c which takes a better approach to
+       * this which could be substituted later.
+       */
+      if (isrc_x  <= texture->width  && isrc_x  >= 0 &&
+          isrc_y  <= texture->height && isrc_y  >= 0 &&
+          isrc_x1 <= texture->width  && isrc_x1 >= 0 &&
+          isrc_y1 <= texture->height && isrc_y1 >= 0)
+      {
+         samp->fetch = fetch_row;
+      }
+      else {
+         samp->fetch = fetch_row_clamped;
+      }
+   }
+
+   return TRUE;
+}
+
+
+static const uint32_t *
+shade_rgb1(struct shader *shader)
+{
+   const __m128i rgb1 = _mm_set1_epi32(0xff000000);
+   const uint32_t *src0 = shader->src0;
+   uint32_t *dst = shader->out0;
+   int width = shader->width;
+   int i;
+
+   for (i = 0; i + 3 < width; i += 4) {
+      __m128i s = *(const __m128i *)&src0[i];
+      *(__m128i *)&dst[i] = _mm_or_si128(s, rgb1);
+   }
+
+   return shader->out0;
+}
+
+
+static void
+init_shader(struct shader *shader,
+           int x, int y, int width, int height)
+{
+   shader->width = align(width, 4);
+}
+
+
+/* Linear shader which implements the BLIT_RGBA shader with the
+ * additional constraints imposed by lp_setup_is_blit().
+ */
+static boolean
+blit_rgba_blit(const struct lp_rast_state *state,
+               unsigned x, unsigned y,
+               unsigned width, unsigned height,
+               const float (*a0)[4],
+               const float (*dadx)[4],
+               const float (*dady)[4],
+               uint8_t *color,
+               unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   const struct lp_jit_texture *texture = &context->textures[0];
+   const uint8_t *src;
+   unsigned src_stride;
+   int src_x, src_y;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   /* Require w==1.0:
+    */
+   if (a0[0][3] != 1.0 ||
+       dadx[0][3] != 0.0 ||
+       dady[0][3] != 0.0)
+      return FALSE;
+
+   src_x = x + util_iround(a0[1][0]*texture->width - 0.5f);
+   src_y = y + util_iround(a0[1][1]*texture->height - 0.5f);
+
+   src = texture->base;
+   src_stride = texture->row_stride[0];
+
+   /* Fall back to blit_rgba() if clamping required:
+    */
+   if (src_x < 0 ||
+       src_y < 0 ||
+       src_x + width > texture->width ||
+       src_y + height > texture->height)
+      return FALSE;
+
+   util_copy_rect(color, PIPE_FORMAT_B8G8R8A8_UNORM, stride,
+                  x, y,
+                  width, height,
+                  src, src_stride,
+                  src_x, src_y);
+
+   return TRUE;
+}
+
+
+/* Linear shader which implements the BLIT_RGB1 shader, with the
+ * additional constraints imposed by lp_setup_is_blit().
+ */
+static boolean
+blit_rgb1_blit(const struct lp_rast_state *state,
+               unsigned x, unsigned y,
+               unsigned width, unsigned height,
+               const float (*a0)[4],
+               const float (*dadx)[4],
+               const float (*dady)[4],
+               uint8_t *color,
+               unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   const struct lp_jit_texture *texture = &context->textures[0];
+   const uint8_t *src;
+   unsigned src_stride;
+   int src_x, src_y;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   /* Require w==1.0:
+    */
+   if (a0[0][3] != 1.0 ||
+       dadx[0][3] != 0.0 ||
+       dady[0][3] != 0.0)
+      return FALSE;
+
+   color += x * 4 + y * stride;
+
+   src_x = x + util_iround(a0[1][0]*texture->width - 0.5f);
+   src_y = y + util_iround(a0[1][1]*texture->height - 0.5f);
+
+   src = texture->base;
+   src_stride = texture->row_stride[0];
+   src += src_x * 4;
+   src += src_y * src_stride;
+
+   if (src_x < 0 ||
+       src_y < 0 ||
+       src_x + width > texture->width ||
+       src_y + height > texture->height)
+      return FALSE;
+
+   for (y = 0; y < height; y++) {
+      const uint32_t *src_row = (const uint32_t *)src;
+      uint32_t *dst_row = (uint32_t *)color;
+
+      for (x = 0; x < width; x++) {
+         *dst_row++ = *src_row++ | 0xff000000;
+      }
+
+      color += stride;
+      src += src_stride;
+   }
+
+   return TRUE;
+}
+
+
+/* Linear shader variant implementing the BLIT_RGBA shader without
+ * blending.
+ */
+static boolean
+blit_rgba(const struct lp_rast_state *state,
+          unsigned x, unsigned y,
+          unsigned width, unsigned height,
+          const float (*a0)[4],
+          const float (*dadx)[4],
+          const float (*dady)[4],
+          uint8_t *color,
+          unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   struct nearest_sampler samp;
+   struct color_blend blend;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   if (!init_nearest_sampler(&samp,
+                             &context->textures[0],
+                             x, y, width, height,
+                             a0[1][0], dadx[1][0], dady[1][0],
+                             a0[1][1], dadx[1][1], dady[1][1],
+                             a0[0][3], dadx[0][3], dady[0][3]))
+      return FALSE;
+
+   init_blend(&blend,
+              x, y, width, height,
+              color, stride);
+
+   /* Rasterize the rectangle and run the shader:
+    */
+   for (y = 0; y < height; y++) {
+      blend.src = samp.fetch(&samp);
+      blend_noop(&blend);
+   }
+
+   return TRUE;
+}
+
+
+static boolean
+blit_rgb1(const struct lp_rast_state *state,
+          unsigned x, unsigned y,
+          unsigned width, unsigned height,
+          const float (*a0)[4],
+          const float (*dadx)[4],
+          const float (*dady)[4],
+          uint8_t *color,
+          unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   struct nearest_sampler samp;
+   struct color_blend blend;
+   struct shader shader;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   if (!init_nearest_sampler(&samp,
+                             &context->textures[0],
+                             x, y, width, height,
+                             a0[1][0], dadx[1][0], dady[1][0],
+                             a0[1][1], dadx[1][1], dady[1][1],
+                             a0[0][3], dadx[0][3], dady[0][3]))
+      return FALSE;
+
+   init_blend(&blend,
+              x, y, width, height,
+              color, stride);
+
+
+   init_shader(&shader,
+               x, y, width, height);
+
+   /* Rasterize the rectangle and run the shader:
+    */
+   for (y = 0; y < height; y++) {
+      shader.src0 = samp.fetch(&samp);
+      blend.src = shade_rgb1(&shader);
+      blend_noop(&blend);
+   }
+
+   return TRUE;
+}
+
+
+/* Linear shader variant implementing the BLIT_RGBA shader with
+ * one/inv_src_alpha blending.
+ */
+static boolean
+blit_rgba_blend_premul(const struct lp_rast_state *state,
+                       unsigned x, unsigned y,
+                       unsigned width, unsigned height,
+                       const float (*a0)[4],
+                       const float (*dadx)[4],
+                       const float (*dady)[4],
+                       uint8_t *color,
+                       unsigned stride)
+{
+   const struct lp_jit_context *context = &state->jit_context;
+   struct nearest_sampler samp;
+   struct color_blend blend;
+
+   LP_DBG(DEBUG_RAST, "%s\n", __FUNCTION__);
+
+   if (!init_nearest_sampler(&samp,
+                             &context->textures[0],
+                             x, y, width, height,
+                             a0[1][0], dadx[1][0], dady[1][0],
+                             a0[1][1], dadx[1][1], dady[1][1],
+                             a0[0][3], dadx[0][3], dady[0][3]))
+      return FALSE;
+
+
+   init_blend(&blend,
+              x, y, width, height,
+              color, stride);
+
+   /* Rasterize the rectangle and run the shader:
+    */
+   for (y = 0; y < height; y++) {
+      blend.src = samp.fetch(&samp);
+      blend_premul(&blend);
+   }
+
+   return TRUE;
+}
+
+
+/* Linear shader which always emits red.  Used for debugging.
+ */
+static boolean
+linear_red(const struct lp_rast_state *state,
+           unsigned x, unsigned y,
+           unsigned width, unsigned height,
+           const float (*a0)[4],
+           const float (*dadx)[4],
+           const float (*dady)[4],
+           uint8_t *color,
+           unsigned stride)
+{
+   union util_color uc;
+
+   util_pack_color_ub(0xff, 0, 0, 0xff,
+                      PIPE_FORMAT_B8G8R8A8_UNORM, &uc);
+
+   util_fill_rect(color,
+                  PIPE_FORMAT_B8G8R8A8_UNORM,
+                  stride,
+                  x,
+                  y,
+                  width,
+                  height,
+                  &uc);
+
+   return TRUE;
+}
+
+
+/* Noop linear shader variant, for debugging.
+ */
+static boolean
+linear_no_op(const struct lp_rast_state *state,
+             unsigned x, unsigned y,
+             unsigned width, unsigned height,
+             const float (*a0)[4],
+             const float (*dadx)[4],
+             const float (*dady)[4],
+             uint8_t *color,
+             unsigned stride)
+{
+   return TRUE;
+}
+
+/* Check for ADD/ONE/INV_SRC_ALPHA, ie premultiplied-alpha blending.
+ */
+static boolean
+is_one_inv_src_alpha_blend(const struct lp_fragment_shader_variant *variant)
+{
+   return
+      !variant->key.blend.logicop_enable &&
+      variant->key.blend.rt[0].blend_enable &&
+      variant->key.blend.rt[0].rgb_func == PIPE_BLEND_ADD &&
+      variant->key.blend.rt[0].rgb_src_factor == PIPE_BLENDFACTOR_ONE &&
+      variant->key.blend.rt[0].rgb_dst_factor == PIPE_BLENDFACTOR_INV_SRC_ALPHA &&
+      variant->key.blend.rt[0].alpha_func == PIPE_BLEND_ADD &&
+      variant->key.blend.rt[0].alpha_src_factor == PIPE_BLENDFACTOR_ONE &&
+      variant->key.blend.rt[0].alpha_dst_factor == PIPE_BLENDFACTOR_INV_SRC_ALPHA &&
+      variant->key.blend.rt[0].colormask == 0xf;
+}
+
+
+/* Examine the fragment shader varient and determine whether we can
+ * substitute a fastpath linear shader implementation.
+ */
+void
+llvmpipe_fs_variant_linear_fastpath(struct lp_fragment_shader_variant *variant)
+{
+   enum pipe_format tex_format = variant->key.samplers[0].texture_state.format;
+
+   if (LP_PERF & PERF_NO_SHADE) {
+      variant->jit_linear                   = linear_red;
+      return;
+   }
+
+   if (variant->shader->kind == LP_FS_KIND_BLIT_RGBA &&
+       tex_format == PIPE_FORMAT_B8G8R8A8_UNORM &&
+       is_nearest_clamp_sampler(&variant->key.samplers[0])) {
+      if (variant->opaque) {
+         variant->jit_linear_blit             = blit_rgba_blit;
+         variant->jit_linear                  = blit_rgba;
+      }
+      else if (is_one_inv_src_alpha_blend(variant) &&
+               util_get_cpu_caps()->has_sse2) {
+         variant->jit_linear                  = blit_rgba_blend_premul;
+      }
+      return;
+   }
+
+   if (variant->shader->kind == LP_FS_KIND_BLIT_RGB1 &&
+       variant->opaque &&
+       (tex_format == PIPE_FORMAT_B8G8R8A8_UNORM ||
+        tex_format == PIPE_FORMAT_B8G8R8X8_UNORM) &&
+       is_nearest_clamp_sampler(&variant->key.samplers[0])) {
+      variant->jit_linear_blit             = blit_rgb1_blit;
+      variant->jit_linear                  = blit_rgb1;
+      return;
+   }
+
+   if (0) {
+      variant->jit_linear                   = linear_no_op;
+      return;
+   }
+}
+#else
+void
+llvmpipe_fs_variant_linear_fastpath(struct lp_fragment_shader_variant *variant)
+{
+   /* don't bother if there is no SSE */
+}
+#endif
+
diff --git a/src/gallium/drivers/llvmpipe/lp_state_fs_linear_llvm.c b/src/gallium/drivers/llvmpipe/lp_state_fs_linear_llvm.c
new file mode 100644
index 00000000000..1333dfef846
--- /dev/null
+++ b/src/gallium/drivers/llvmpipe/lp_state_fs_linear_llvm.c
@@ -0,0 +1,522 @@
+/**************************************************************************
+ *
+ * Copyright 2010-2021 VMware, Inc.
+ * All Rights Reserved.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the
+ * "Software"), to deal in the Software without restriction, including
+ * without limitation the rights to use, copy, modify, merge, publish,
+ * distribute, sub license, and/or sell copies of the Software, and to
+ * permit persons to whom the Software is furnished to do so, subject to
+ * the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the
+ * next paragraph) shall be included in all copies or substantial portions
+ * of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS
+ * OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT.
+ * IN NO EVENT SHALL VMWARE AND/OR ITS SUPPLIERS BE LIABLE FOR
+ * ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT,
+ * TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE
+ * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+ *
+ **************************************************************************/
+
+#include <limits.h>
+#include "util/simple_list.h"
+
+#include "pipe/p_defines.h"
+#include "util/u_inlines.h"
+#include "util/u_memory.h"
+#include "util/u_pointer.h"
+#include "util/format/u_format.h"
+#include "util/u_dump.h"
+#include "util/u_string.h"
+#include "util/os_time.h"
+#include "pipe/p_shader_tokens.h"
+#include "draw/draw_context.h"
+#include "tgsi/tgsi_dump.h"
+#include "tgsi/tgsi_scan.h"
+#include "tgsi/tgsi_parse.h"
+#include "gallivm/lp_bld_type.h"
+#include "gallivm/lp_bld_const.h"
+#include "gallivm/lp_bld_conv.h"
+#include "gallivm/lp_bld_init.h"
+#include "gallivm/lp_bld_intr.h"
+#include "gallivm/lp_bld_logic.h"
+#include "gallivm/lp_bld_tgsi.h"
+#include "gallivm/lp_bld_swizzle.h"
+#include "gallivm/lp_bld_flow.h"
+#include "gallivm/lp_bld_printf.h"
+#include "gallivm/lp_bld_debug.h"
+
+#include "lp_bld_alpha.h"
+#include "lp_bld_blend.h"
+#include "lp_bld_depth.h"
+#include "lp_bld_interp.h"
+#include "lp_context.h"
+#include "lp_debug.h"
+#include "lp_perf.h"
+#include "lp_screen.h"
+#include "lp_setup.h"
+#include "lp_state.h"
+#include "lp_tex_sample.h"
+#include "lp_flush.h"
+#include "lp_state_fs.h"
+
+
+/**
+ * Sampler.
+ */
+struct linear_sampler
+{
+   struct lp_build_sampler_aos base;
+
+   LLVMValueRef texels_ptrs[LP_MAX_LINEAR_TEXTURES];
+
+   LLVMValueRef counter;
+
+   unsigned instance;
+};
+
+
+/**
+ * Provide texels to the TGSI translation.
+ *
+ * We don't actually do any texture sampling here, but simply hand the
+ * precomputed row of texels.
+ */
+static LLVMValueRef
+emit_fetch_texel_linear(const struct lp_build_sampler_aos *base,
+                        struct lp_build_context *bld,
+                        unsigned target, /* TGSI_TEXTURE_* */
+                        unsigned unit,
+                        LLVMValueRef coords,
+                        const struct lp_derivatives derivs,
+                        enum lp_build_tex_modifier modifier)
+{
+   struct linear_sampler *sampler = (struct linear_sampler *)base;
+   LLVMValueRef texels_ptr;
+   LLVMValueRef texel;
+
+   if (sampler->instance >= LP_MAX_LINEAR_TEXTURES) {
+      assert(FALSE);
+      return bld->undef;
+   }
+
+   /* Pointer to a row of texels */
+   texels_ptr = sampler->texels_ptrs[sampler->instance];
+
+   texel = lp_build_pointer_get(bld->gallivm->builder, texels_ptr,
+                                sampler->counter);
+   assert(LLVMTypeOf(texel) == bld->vec_type);
+
+   /*
+    * We have a struct lp_linear_sampler instance per TEX instruction,
+    * _not_ per unit, as each TEX intruction will need separate storage for the
+    * texels.
+    */
+   (void)unit;
+   ++sampler->instance;
+
+   return texel;
+}
+
+/**
+ * Generates the main body of the fragment shader
+ * Supports generating code for 4 pixel blocks and individual pixels
+ */
+static LLVMValueRef
+llvm_fragment_body(struct lp_build_context *bld,
+                   struct lp_fragment_shader *shader,
+                   struct lp_fragment_shader_variant *variant,
+                   struct linear_sampler* sampler,
+                   LLVMValueRef *inputs_ptrs,
+                   LLVMValueRef consts_ptr,
+                   LLVMValueRef blend_color,
+                   LLVMValueRef alpha_ref,
+                   struct lp_type fs_type,
+                   LLVMValueRef dst)
+{
+   const unsigned char bgra_swizzles[4] = {2, 1, 0, 3};
+
+   LLVMValueRef inputs[PIPE_MAX_SHADER_INPUTS];
+   LLVMValueRef outputs[PIPE_MAX_SHADER_OUTPUTS];
+
+   LLVMBuilderRef builder = bld->gallivm->builder;
+   struct gallivm_state *gallivm = bld->gallivm;
+
+   LLVMValueRef src1 = lp_build_zero(gallivm, fs_type);
+
+   LLVMValueRef result = NULL;
+   unsigned i;
+
+   sampler->instance = 0;
+
+
+   /*
+    * Advance inputs
+    */
+
+   for (i = 0; i < shader->info.base.num_inputs; ++i) {
+      LLVMValueRef inputs_ptr;
+      LLVMValueRef input;
+
+      inputs_ptr = inputs_ptrs[i];
+
+      input = lp_build_pointer_get(builder, inputs_ptr, sampler->counter);
+      assert(LLVMTypeOf(input) == bld->vec_type);
+
+      inputs[i] = input;
+   }
+
+   for ( ; i < PIPE_MAX_SHADER_INPUTS; ++i) {
+      inputs[i] = bld->undef;
+   }
+
+
+   /*
+    * Translate the TGSI
+    */
+
+   for (i = 0; i < PIPE_MAX_SHADER_OUTPUTS; ++i) {
+      outputs[i] = bld->undef;
+   }
+
+   lp_build_tgsi_aos(gallivm, shader->base.tokens, fs_type,
+                     bgra_swizzles,
+                     consts_ptr, inputs, outputs,
+                     &sampler->base,
+                     &shader->info.base);
+
+   /*
+    * Blend output color
+    */
+
+   for (i = 0; i < shader->info.base.num_outputs; ++i) {
+      LLVMValueRef mask = NULL;
+      LLVMValueRef output;
+      unsigned cbuf;
+
+      if (!outputs[i])
+         continue;
+
+      output = LLVMBuildLoad(builder, outputs[i], "");
+      lp_build_name(output, "output%u", i);
+
+      cbuf = shader->info.base.output_semantic_index[i];
+      lp_build_name(output, "cbuf%u", cbuf);
+
+      if (shader->info.base.output_semantic_name[i] != TGSI_SEMANTIC_COLOR || cbuf != 0)
+         continue;
+
+      /* Perform alpha test if necessary */
+      if (variant->key.alpha.enabled) {
+         LLVMTypeRef vec_type = lp_build_vec_type(gallivm, fs_type);
+         LLVMValueRef broadcast_alpha = lp_build_broadcast(gallivm, vec_type, alpha_ref);
+
+         mask = lp_build_cmp(bld, variant->key.alpha.func, output, broadcast_alpha);
+         /* XXX is 4 correct? */
+         mask = lp_build_swizzle_scalar_aos(bld, mask, bgra_swizzles[3], 4);
+
+         lp_build_name(mask, "alpha_test_mask");
+      }
+
+      result = lp_build_blend_aos(gallivm,
+                                  &variant->key.blend,
+                                  variant->key.cbuf_format[i],
+                                  fs_type,
+                                  cbuf,   /* rt */
+                                  output, /* src */
+                                  NULL,   /* src_alpha */
+                                  src1,   /* src1 */
+                                  NULL,   /* src1_alpha */
+                                  dst,
+                                  mask,
+                                  blend_color,  /* const_ */
+                                  NULL,         /* const_alpha */
+                                  bgra_swizzles,
+                                  4);
+   }
+
+   return result;
+}
+
+/**
+ * Generate a function that executes the fragment shader in a linear fashion.
+ */
+void
+llvmpipe_fs_variant_linear_llvm(struct llvmpipe_context *lp,
+                                struct lp_fragment_shader *shader,
+                                struct lp_fragment_shader_variant *variant)
+{
+   struct gallivm_state *gallivm = variant->gallivm;
+   char func_name[256];
+   struct lp_type fs_type;
+   LLVMTypeRef ret_type;
+   LLVMTypeRef arg_types[4];
+   LLVMTypeRef func_type;
+   LLVMValueRef context_ptr;
+   LLVMValueRef x;
+   LLVMValueRef y;
+   LLVMValueRef width;
+   LLVMValueRef excess;
+   LLVMValueRef function;
+   LLVMBasicBlockRef block;
+   LLVMBuilderRef builder;
+   struct lp_build_context bld;
+   struct linear_sampler sampler;
+   struct lp_build_if_state ifstate;
+   struct lp_build_for_loop_state loop;
+
+   LLVMValueRef consts_ptr;
+   LLVMValueRef interpolators_ptr;
+   LLVMValueRef samplers_ptr;
+   LLVMValueRef color0_ptr;
+   LLVMValueRef blend_color;
+   LLVMValueRef alpha_ref;
+
+   LLVMValueRef inputs_ptrs[LP_MAX_LINEAR_INPUTS];
+
+   LLVMTypeRef int8t = LLVMInt8TypeInContext(gallivm->context);
+   LLVMTypeRef int32t = LLVMInt32TypeInContext(gallivm->context);
+   LLVMTypeRef pint8t = LLVMPointerType(int8t, 0);
+   LLVMTypeRef pixelt = LLVMVectorType(int32t, 4);
+
+   unsigned attrib;
+   unsigned i;
+
+   memset(&fs_type, 0, sizeof fs_type);
+   fs_type.floating = FALSE;
+   fs_type.sign = FALSE;
+   fs_type.norm = TRUE;
+   fs_type.width = 8;
+   fs_type.length = 16;
+
+   if (LP_DEBUG & DEBUG_TGSI) {
+      tgsi_dump(shader->base.tokens, 0);
+   }
+
+   /*
+    * Generate the function prototype. Any change here must be reflected in
+    * lp_jit.h's lp_jit_frag_func function pointer type, and vice-versa.
+    */
+
+   snprintf(func_name, sizeof(func_name), "fs%u_variant%u_linear",
+            shader->no, variant->no);
+
+   ret_type = pint8t;
+   arg_types[0] = variant->jit_linear_context_ptr_type; /* context */
+   arg_types[1] = int32t;                              /* x */
+   arg_types[2] = int32t;                              /* y */
+   arg_types[3] = int32t;                              /* width */
+
+   func_type = LLVMFunctionType(ret_type, arg_types, ARRAY_SIZE(arg_types), 0);
+
+   function = LLVMAddFunction(gallivm->module, func_name, func_type);
+   LLVMSetFunctionCallConv(function, LLVMCCallConv);
+
+   variant->linear_function = function;
+
+   /* XXX: need to propagate noalias down into color param now we are
+    * passing a pointer-to-pointer?
+    */
+   for (i = 0; i < ARRAY_SIZE(arg_types); ++i) {
+      if (LLVMGetTypeKind(arg_types[i]) == LLVMPointerTypeKind) {
+         lp_add_function_attr(function, i + 1, LP_FUNC_ATTR_NOALIAS);
+      }
+   }
+
+   context_ptr  = LLVMGetParam(function, 0);
+   x            = LLVMGetParam(function, 1);
+   y            = LLVMGetParam(function, 2);
+   width        = LLVMGetParam(function, 3);
+
+   lp_build_name(context_ptr, "context");
+   lp_build_name(x, "x");
+   lp_build_name(y, "y");
+   lp_build_name(width, "width");
+
+   /*
+    * Function body
+    */
+
+   block = LLVMAppendBasicBlockInContext(gallivm->context, function, "entry");
+   builder = gallivm->builder;
+
+   LLVMPositionBuilderAtEnd(builder, block);
+
+   lp_build_context_init(&bld, gallivm, fs_type);
+
+   /*
+    * Get context data
+    */
+
+   consts_ptr = lp_jit_linear_context_constants(gallivm, context_ptr);
+   interpolators_ptr = lp_jit_linear_context_inputs(gallivm, context_ptr);
+   samplers_ptr = lp_jit_linear_context_tex(gallivm, context_ptr);
+
+   color0_ptr = lp_jit_linear_context_color0(gallivm, context_ptr);
+   color0_ptr = LLVMBuildLoad(builder, color0_ptr, "");
+   color0_ptr = LLVMBuildBitCast(builder, color0_ptr, LLVMPointerType(bld.vec_type, 0), "");
+
+   blend_color = lp_jit_linear_context_blend_color(gallivm, context_ptr);
+   blend_color = LLVMBuildLoad(builder, blend_color, "");
+   blend_color = lp_build_broadcast(gallivm, LLVMVectorType(int32t, 4), blend_color);
+   blend_color = LLVMBuildBitCast(builder, blend_color, LLVMVectorType(int8t, 16), "");
+
+   alpha_ref = lp_jit_linear_context_alpha_ref(gallivm, context_ptr);
+   alpha_ref = LLVMBuildLoad(builder, alpha_ref, "");
+
+   /*
+    * Invoke the input interpolators
+    */
+
+   for (attrib = 0; attrib < shader->info.base.num_inputs; ++attrib) {
+      LLVMValueRef index;
+      LLVMValueRef elem;
+      LLVMValueRef fetch_ptr;
+      LLVMValueRef inputs_ptr;
+
+      assert(attrib < LP_MAX_LINEAR_INPUTS);
+      if (attrib >= LP_MAX_LINEAR_INPUTS) {
+         break;
+      }
+
+      index = LLVMConstInt(int32t, attrib, 0);
+
+      elem = lp_build_array_get(bld.gallivm, interpolators_ptr, index);
+      assert(LLVMGetTypeKind(LLVMTypeOf(elem)) == LLVMPointerTypeKind);
+
+      fetch_ptr = lp_build_pointer_get(builder, elem,
+                                       LLVMConstInt(int32t, 0, 0));
+      assert(LLVMGetTypeKind(LLVMTypeOf(fetch_ptr)) == LLVMPointerTypeKind);
+
+      /* Pointer to a row of interpolated inputs */
+      elem = LLVMBuildBitCast(builder, elem, pint8t, "");
+      inputs_ptr = LLVMBuildCall(builder, fetch_ptr, &elem, 1, "");
+      assert(LLVMGetTypeKind(LLVMTypeOf(inputs_ptr)) == LLVMPointerTypeKind);
+
+      /* Mark the function read-only so that LLVM can optimize it away */
+      lp_add_function_attr(inputs_ptr, -1, LP_FUNC_ATTR_READONLY);
+      lp_add_function_attr(inputs_ptr, -1, LP_FUNC_ATTR_NOUNWIND);
+
+      lp_build_name(inputs_ptr, "input%u_ptr", attrib);
+
+      inputs_ptrs[attrib] = inputs_ptr;
+   }
+
+   /*
+    * Invoke and hook up the texture samplers.
+    */
+
+   memset(&sampler, 0, sizeof sampler);
+   sampler.base.emit_fetch_texel = &emit_fetch_texel_linear;
+
+   for (attrib = 0; attrib < shader->info.num_texs; ++attrib) {
+      LLVMValueRef index;
+      LLVMValueRef elem;
+      LLVMValueRef fetch_ptr;
+      LLVMValueRef texels_ptr;
+
+      assert(attrib < LP_MAX_LINEAR_TEXTURES);
+      if (attrib >= LP_MAX_LINEAR_TEXTURES) {
+         break;
+      }
+
+      index = LLVMConstInt(int32t, attrib, 0);
+
+      elem = lp_build_array_get(bld.gallivm, samplers_ptr, index);
+      assert(LLVMGetTypeKind(LLVMTypeOf(elem)) == LLVMPointerTypeKind);
+
+      fetch_ptr = lp_build_pointer_get(builder, elem, LLVMConstInt(int32t, 0, 0));
+      assert(LLVMGetTypeKind(LLVMTypeOf(fetch_ptr)) == LLVMPointerTypeKind);
+
+      /* Pointer to a row of texels */
+      elem = LLVMBuildBitCast(builder, elem, pint8t, "");
+      texels_ptr = LLVMBuildCall(builder, fetch_ptr, &elem, 1, "");
+      assert(LLVMGetTypeKind(LLVMTypeOf(texels_ptr)) == LLVMPointerTypeKind);
+
+      /* Mark the function read-only so that LLVM can optimize it away */
+      lp_add_function_attr(texels_ptr, -1, LP_FUNC_ATTR_READONLY);
+      lp_add_function_attr(texels_ptr, -1, LP_FUNC_ATTR_NOUNWIND);
+
+      lp_build_name(texels_ptr, "tex%u_ptr", attrib);
+
+      sampler.texels_ptrs[attrib] = texels_ptr;
+   }
+
+   excess = LLVMBuildAnd(builder, width, LLVMConstInt(int32t, 3, 0), "");
+   width = LLVMBuildLShr(builder, width, LLVMConstInt(int32t, 2, 0), "");
+
+   /* Loop over blocks of 4 pixels */
+   lp_build_for_loop_begin(&loop, gallivm, LLVMConstInt(int32t, 0, 0), LLVMIntULT, width, LLVMConstInt(int32t, 1, 0));
+   {
+      LLVMValueRef value;
+      sampler.counter = loop.counter;
+
+      /* Read 4 pixels */
+      value = lp_build_pointer_get_unaligned(builder, color0_ptr, loop.counter, 4);
+
+      /* Perform fragment shader body */
+      value = llvm_fragment_body(&bld, shader, variant, &sampler, inputs_ptrs, consts_ptr, blend_color, alpha_ref, fs_type, value);
+
+      /* Write 4 pixels */
+      lp_build_pointer_set_unaligned(builder, color0_ptr, loop.counter, value, 4);
+   }
+   lp_build_for_loop_end(&loop);
+
+   /* Compute the edge pixels (width % 4) */
+   lp_build_if(&ifstate, gallivm, LLVMBuildICmp(builder, LLVMIntNE, excess, LLVMConstInt(int32t, 0, 0), ""));
+   {
+      struct lp_build_loop_state loop_read, loop_write;
+      LLVMValueRef buf, elem, result, pixel_ptr;
+      LLVMValueRef buf_ptr = lp_build_alloca(gallivm, pixelt, "");
+
+      sampler.counter = width;
+
+      /* Get the i32* pixel pointer from the <i16x8>* element pointer */
+      pixel_ptr = LLVMBuildGEP(gallivm->builder, color0_ptr, &width, 1, "");
+      pixel_ptr = LLVMBuildBitCast(gallivm->builder, pixel_ptr, LLVMPointerType(int32t, 0), "");
+
+      /* Copy individual pixels from memory to local buffer */
+      lp_build_loop_begin(&loop_read, gallivm, LLVMConstInt(int32t, 0, 0));
+      {
+         elem = lp_build_pointer_get(gallivm->builder, pixel_ptr, loop_read.counter);
+
+         buf = LLVMBuildLoad(gallivm->builder, buf_ptr, "");
+         buf = LLVMBuildInsertElement(builder, buf, elem, loop_read.counter, "");
+         LLVMBuildStore(builder, buf, buf_ptr);
+      }
+      lp_build_loop_end_cond(&loop_read, excess, LLVMConstInt(int32t, 1, 0), LLVMIntUGE);
+
+      /* Perform fragment shader body */
+      buf = LLVMBuildLoad(gallivm->builder, buf_ptr, "");
+      buf = LLVMBuildBitCast(builder, buf, bld.vec_type, "");
+
+      result = llvm_fragment_body(&bld, shader, variant, &sampler, inputs_ptrs, consts_ptr, blend_color, alpha_ref, fs_type, buf);
+      result = LLVMBuildBitCast(builder, result, pixelt, "");
+
+      /* Write individual pixels from local buffer to the memory */
+      lp_build_loop_begin(&loop_write, gallivm, LLVMConstInt(int32t, 0, 0));
+      {
+         elem = LLVMBuildExtractElement(builder, result, loop_write.counter, "");
+
+         lp_build_pointer_set(gallivm->builder, pixel_ptr, loop_write.counter, elem);
+      }
+      lp_build_loop_end_cond(&loop_write, excess, LLVMConstInt(int32t, 1, 0), LLVMIntUGE);
+   }
+   lp_build_endif(&ifstate);
+
+
+   color0_ptr = LLVMBuildBitCast(builder, color0_ptr, pint8t, "");
+
+   LLVMBuildRet(builder, color0_ptr);
+
+   gallivm_verify_function(gallivm, function);
+}
+
+
diff --git a/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c b/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c
index c9d0e17e27e..2b1d65af1c3 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_rasterizer.c
@@ -32,7 +32,7 @@
 #include "lp_setup.h"
 #include "draw/draw_context.h"
 
-struct lp_rast_state {
+struct lp_rasterizer_state {
    struct pipe_rasterizer_state lp_state;
    struct pipe_rasterizer_state draw_state;
 };
@@ -63,7 +63,7 @@ llvmpipe_create_rasterizer_state(struct pipe_context *pipe,
    /* Partition rasterizer state into what we want the draw module to
     * handle, and what we'll look after ourselves.
     */
-   struct lp_rast_state *state = MALLOC_STRUCT(lp_rast_state);
+   struct lp_rasterizer_state *state = MALLOC_STRUCT(lp_rasterizer_state);
    if (!state)
       return NULL;
 
@@ -102,8 +102,8 @@ static void
 llvmpipe_bind_rasterizer_state(struct pipe_context *pipe, void *handle)
 {
    struct llvmpipe_context *llvmpipe = llvmpipe_context(pipe);
-   const struct lp_rast_state *state =
-      (const struct lp_rast_state *) handle;
+   const struct lp_rasterizer_state *state =
+      (const struct lp_rasterizer_state *) handle;
 
    if (state) {
       llvmpipe->rasterizer = &state->lp_state;
@@ -125,6 +125,7 @@ llvmpipe_bind_rasterizer_state(struct pipe_context *pipe, void *handle)
                               state->lp_state.line_rectangular);
       lp_setup_set_point_state( llvmpipe->setup,
                                state->lp_state.point_size,
+                               state->lp_state.point_tri_clip,
                                state->lp_state.point_size_per_vertex,
                                state->lp_state.sprite_coord_enable,
                                state->lp_state.sprite_coord_mode,
diff --git a/src/gallium/drivers/llvmpipe/lp_state_sampler.c b/src/gallium/drivers/llvmpipe/lp_state_sampler.c
index 9303e3bd7ec..5332bb049c2 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_sampler.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_sampler.c
@@ -178,8 +178,12 @@ llvmpipe_set_sampler_views(struct pipe_context *pipe,
    }
    else if (shader == PIPE_SHADER_COMPUTE) {
       llvmpipe->cs_dirty |= LP_CSNEW_SAMPLER_VIEW;
-   } else {
+   }
+   else if (shader == PIPE_SHADER_FRAGMENT) {
       llvmpipe->dirty |= LP_NEW_SAMPLER_VIEW;
+      lp_setup_set_fragment_sampler_views(llvmpipe->setup,
+                                          llvmpipe->num_sampler_views[PIPE_SHADER_FRAGMENT],
+                                          llvmpipe->sampler_views[PIPE_SHADER_FRAGMENT]);
    }
 }
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_setup.c b/src/gallium/drivers/llvmpipe/lp_state_setup.c
index c737850be22..e092179cf80 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_setup.c
+++ b/src/gallium/drivers/llvmpipe/lp_state_setup.c
@@ -71,6 +71,7 @@ struct lp_setup_args
    LLVMValueRef a0;
    LLVMValueRef dadx;
    LLVMValueRef dady;
+   LLVMValueRef key;
 
    /* Derived:
     */
@@ -708,7 +709,7 @@ generate_setup_variant(struct lp_setup_variant_key *key,
    char func_name[64];
    LLVMTypeRef vec4f_type;
    LLVMTypeRef func_type;
-   LLVMTypeRef arg_types[7];
+   LLVMTypeRef arg_types[8];
    LLVMBasicBlockRef block;
    LLVMBuilderRef builder;
    int64_t t0 = 0, t1;
@@ -752,6 +753,7 @@ generate_setup_variant(struct lp_setup_variant_key *key,
    arg_types[4] = LLVMPointerType(vec4f_type, 0);	/* a0, aligned */
    arg_types[5] = LLVMPointerType(vec4f_type, 0);	/* dadx, aligned */
    arg_types[6] = LLVMPointerType(vec4f_type, 0);	/* dady, aligned */
+   arg_types[7] = LLVMPointerType(vec4f_type, 0);	/* key (placeholder) */
 
    func_type = LLVMFunctionType(LLVMVoidTypeInContext(gallivm->context),
                                 arg_types, ARRAY_SIZE(arg_types), 0);
@@ -769,6 +771,7 @@ generate_setup_variant(struct lp_setup_variant_key *key,
    args.a0       = LLVMGetParam(variant->function, 4);
    args.dadx     = LLVMGetParam(variant->function, 5);
    args.dady     = LLVMGetParam(variant->function, 6);
+   args.key      = LLVMGetParam(variant->function, 7);
 
    lp_build_name(args.v0, "in_v0");
    lp_build_name(args.v1, "in_v1");
@@ -777,6 +780,7 @@ generate_setup_variant(struct lp_setup_variant_key *key,
    lp_build_name(args.a0, "out_a0");
    lp_build_name(args.dadx, "out_dadx");
    lp_build_name(args.dady, "out_dady");
+   lp_build_name(args.key, "key");
 
    /*
     * Function body
@@ -864,6 +868,7 @@ lp_make_setup_variant_key(struct llvmpipe_context *lp,
 
    key->pgon_offset_scale = lp->rasterizer->offset_scale;
    key->pgon_offset_clamp = lp->rasterizer->offset_clamp;
+   key->uses_constant_interp = 0;
    key->pad = 0;
    memcpy(key->inputs, fs->inputs, key->num_inputs * sizeof key->inputs[0]);
    for (i = 0; i < key->num_inputs; i++) {
@@ -873,8 +878,10 @@ lp_make_setup_variant_key(struct llvmpipe_context *lp,
          else
             key->inputs[i].interp = LP_INTERP_PERSPECTIVE;
       }
+      if (key->inputs[i].interp == LP_INTERP_CONSTANT) {
+         key->uses_constant_interp = 1;
+      }
    }
-
 }
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_state_setup.h b/src/gallium/drivers/llvmpipe/lp_state_setup.h
index 18e0ea8f08e..a92a9e7f288 100644
--- a/src/gallium/drivers/llvmpipe/lp_state_setup.h
+++ b/src/gallium/drivers/llvmpipe/lp_state_setup.h
@@ -25,6 +25,7 @@ struct lp_setup_variant_key {
    unsigned pixel_center_half:1;
    unsigned twoside:1;
    unsigned floating_point_depth:1;
+   unsigned uses_constant_interp:1;
    unsigned multisample:1;
    unsigned pad:3;
 
@@ -42,7 +43,8 @@ typedef void (*lp_jit_setup_triangle)( const float (*v0)[4],
 				       boolean front_facing,
 				       float (*a0)[4],
 				       float (*dadx)[4],
-				       float (*dady)[4] );
+				       float (*dady)[4],
+                                       const struct lp_setup_variant_key *key );
 
 
 
diff --git a/src/gallium/drivers/llvmpipe/lp_texture.h b/src/gallium/drivers/llvmpipe/lp_texture.h
index bda52814165..d1ba22b2685 100644
--- a/src/gallium/drivers/llvmpipe/lp_texture.h
+++ b/src/gallium/drivers/llvmpipe/lp_texture.h
@@ -109,8 +109,6 @@ struct llvmpipe_resource
 struct llvmpipe_transfer
 {
    struct pipe_transfer base;
-
-   unsigned long offset;
 };
 
 
diff --git a/src/gallium/drivers/llvmpipe/meson.build b/src/gallium/drivers/llvmpipe/meson.build
index 777830ace92..661eb1e8a2e 100644
--- a/src/gallium/drivers/llvmpipe/meson.build
+++ b/src/gallium/drivers/llvmpipe/meson.build
@@ -44,6 +44,10 @@ files_llvmpipe = files(
   'lp_jit.c',
   'lp_jit.h',
   'lp_limits.h',
+  'lp_linear.c',
+  'lp_linear_fastpath.c',
+  'lp_linear_interp.c',
+  'lp_linear_sampler.c',
   'lp_memory.c',
   'lp_memory.h',
   'lp_perf.c',
@@ -54,7 +58,10 @@ files_llvmpipe = files(
   'lp_rast.c',
   'lp_rast_debug.c',
   'lp_rast.h',
+  'lp_rast_linear.c',
+  'lp_rast_linear_fallback.c',
   'lp_rast_priv.h',
+  'lp_rast_rect.c',
   'lp_rast_tri.c',
   'lp_rast_tri_tmp.h',
   'lp_scene.c',
@@ -64,10 +71,12 @@ files_llvmpipe = files(
   'lp_screen.c',
   'lp_screen.h',
   'lp_setup.c',
+  'lp_setup_analysis.c',
   'lp_setup_context.h',
   'lp_setup.h',
   'lp_setup_line.c',
   'lp_setup_point.c',
+  'lp_setup_rect.c',
   'lp_setup_tri.c',
   'lp_setup_vbuf.c',
   'lp_state_blend.c',
@@ -77,6 +86,10 @@ files_llvmpipe = files(
   'lp_state_cs.h',
   'lp_state_fs.c',
   'lp_state_fs.h',
+  'lp_state_fs_analysis.c',
+  'lp_state_fs_fastpath.c',
+  'lp_state_fs_linear.c',
+  'lp_state_fs_linear_llvm.c',
   'lp_state_gs.c',
   'lp_state.h',
   'lp_state_rasterizer.c',