svga: implement generic variable index remapping

The state tracker may generate shaders that use generic vs outputs /
fs inputs like:

DCL IN[0], GENERIC[0]
DCL IN[1], GENERIC[10]
DCL IN[2], GENERIC[11]

This patch remaps 0, 10, 11 to small integers like 1, 2, 3 so that we
stay inside the SVGA3D limit (8).

The remapping is done to both the vertex shader outputs and the
fragment shader inputs.  The same mapping must be used for a vs/fs
pair.

Note that 'union svga_compile_key' is now 'struct svga_compile_key'
because we needed to add the register remapping table.  The change in
size isn't really significant though (it's not a search key).

Also, add assertions when building up SVGA3D src/dst registers to we
don't try to store too large of value for the bitfield size.

Reviewed-by: José Fonseca <jfonseca@vmware.com>

src/gallium/drivers/svga/svga_context.h

@@ -36,6 +36,7 @@
 #include "tgsi/tgsi_scan.h"
 
 #include "svga_state.h"
+#include "svga_tgsi.h"
 #include "svga_hw_reg.h"
 #include "svga3d_shaderdefs.h"
 
@@ -66,6 +67,11 @@ struct svga_shader
 struct svga_fragment_shader
 {
    struct svga_shader base;
+
+   /** Mask of which generic varying variables are read by this shader */
+   unsigned generic_inputs;
+   /** Table mapping original TGSI generic indexes to low integers */
+   int8_t generic_remap_table[MAX_GENERIC_VARYING];
 };
 
 struct svga_vertex_shader

src/gallium/drivers/svga/svga_pipe_fs.c

@@ -59,6 +59,10 @@ svga_create_fs_state(struct pipe_context *pipe,
 
    fs->base.id = svga->debug.shader_id++;
    
+   fs->generic_inputs = svga_get_generic_inputs_mask(&fs->base.info);
+
+   svga_remap_generics(fs->generic_inputs, fs->generic_remap_table);
+
    if (SVGA_DEBUG & DEBUG_TGSI || 0) {
       debug_printf("%s id: %u, inputs: %u, outputs: %u\n",
                    __FUNCTION__, fs->base.id,

src/gallium/drivers/svga/svga_state_fs.c

@@ -116,6 +116,7 @@ fail:
  */
 static enum pipe_error
 make_fs_key(const struct svga_context *svga,
+            struct svga_fragment_shader *fs,
             struct svga_fs_compile_key *key)
 {
    int i;
@@ -215,7 +216,7 @@ emit_hw_fs(struct svga_context *svga, unsigned dirty)
     * SVGA_NEW_NEED_SWTNL
     * SVGA_NEW_SAMPLER
     */
-   ret = make_fs_key( svga, &key );
+   ret = make_fs_key( svga, fs, &key );
    if (ret != PIPE_OK)
       return ret;
 

src/gallium/drivers/svga/svga_state_vs.c

@@ -107,7 +107,7 @@ fail:
    return ret;
 }
 
-/* SVGA_NEW_PRESCALE, SVGA_NEW_RAST, SVGA_NEW_ZERO_STRIDE
+/* SVGA_NEW_PRESCALE, SVGA_NEW_RAST, SVGA_NEW_ZERO_STRIDE, SVGA_NEW_FS
  */
 static void
 make_vs_key(struct svga_context *svga, struct svga_vs_compile_key *key)
@@ -119,6 +119,9 @@ make_vs_key(struct svga_context *svga, struct svga_vs_compile_key *key)
       svga->curr.zero_stride_vertex_elements;
    key->num_zero_stride_vertex_elements =
       svga->curr.num_zero_stride_vertex_elements;
+
+   /* SVGA_NEW_FS */
+   key->fs_generic_inputs = svga->curr.fs->generic_inputs;
 }
 
 
@@ -166,6 +169,7 @@ struct svga_tracked_state svga_hw_vs =
 {
    "vertex shader (hwtnl)",
    (SVGA_NEW_VS |
+    SVGA_NEW_FS |
     SVGA_NEW_PRESCALE |
     SVGA_NEW_NEED_SWTNL |
     SVGA_NEW_ZERO_STRIDE),

src/gallium/drivers/svga/svga_swtnl_state.c

@@ -31,7 +31,7 @@
 #include "svga_context.h"
 #include "svga_swtnl.h"
 #include "svga_state.h"
-
+#include "svga_tgsi.h"
 #include "svga_swtnl_private.h"
 
 
@@ -169,13 +169,15 @@ svga_swtnl_update_vdecl( struct svga_context *svga )
    nr_decls++;
 
    for (i = 0; i < fs->base.info.num_inputs; i++) {
-      unsigned name = fs->base.info.input_semantic_name[i];
-      unsigned index = fs->base.info.input_semantic_index[i];
-      src = draw_find_shader_output(draw, name, index);
-      vdecl[nr_decls].array.offset = offset;
-      vdecl[nr_decls].identity.usageIndex = fs->base.info.input_semantic_index[i];
+      const unsigned sem_name = fs->base.info.input_semantic_name[i];
+      const unsigned sem_index = fs->base.info.input_semantic_index[i];
 
-      switch (name) {
+      src = draw_find_shader_output(draw, sem_name, sem_index);
+
+      vdecl[nr_decls].array.offset = offset;
+      vdecl[nr_decls].identity.usageIndex = sem_index;
+
+      switch (sem_name) {
       case TGSI_SEMANTIC_COLOR:
          draw_emit_vertex_attr(vinfo, EMIT_4F, colorInterp, src);
          vdecl[nr_decls].identity.usage = SVGA3D_DECLUSAGE_COLOR;
@@ -187,7 +189,8 @@ svga_swtnl_update_vdecl( struct svga_context *svga )
          draw_emit_vertex_attr(vinfo, EMIT_4F, INTERP_PERSPECTIVE, src);
          vdecl[nr_decls].identity.usage = SVGA3D_DECLUSAGE_TEXCOORD;
          vdecl[nr_decls].identity.type = SVGA3D_DECLTYPE_FLOAT4;
-         vdecl[nr_decls].identity.usageIndex += 1;
+         vdecl[nr_decls].identity.usageIndex =
+            svga_remap_generic_index(fs->generic_remap_table, sem_index);
          offset += 16;
          nr_decls++;
          break;

src/gallium/drivers/svga/svga_tgsi.c

@@ -30,6 +30,7 @@
 #include "tgsi/tgsi_parse.h"
 #include "tgsi/tgsi_dump.h"
 #include "tgsi/tgsi_scan.h"
+#include "util/u_math.h"
 #include "util/u_memory.h"
 #include "util/u_bitmask.h"
 
@@ -156,7 +157,82 @@ static boolean svga_shader_emit_header( struct svga_shader_emitter *emit )
 }
 
 
+/**
+ * Use the shader info to generate a bitmask indicating which generic
+ * inputs are used by the shader.  A set bit indicates that GENERIC[i]
+ * is used.
+ */
+unsigned
+svga_get_generic_inputs_mask(const struct tgsi_shader_info *info)
+{
+   unsigned i, mask = 0x0;
 
+   for (i = 0; i < info->num_inputs; i++) {
+      if (info->input_semantic_name[i] == TGSI_SEMANTIC_GENERIC) {
+         unsigned j = info->input_semantic_index[i];
+         assert(j < sizeof(mask) * 8);
+         mask |= 1 << j;
+      }
+   }
+
+   return mask;
+}
+
+
+/**
+ * Given a mask of used generic variables (as returned by the above functions)
+ * fill in a table which maps those indexes to small integers.
+ * This table is used by the remap_generic_index() function in
+ * svga_tgsi_decl_sm30.c
+ * Example: if generics_mask = binary(1010) it means that GENERIC[1] and
+ * GENERIC[3] are used.  The remap_table will contain:
+ *   table[1] = 0;
+ *   table[3] = 1;
+ * The remaining table entries will be filled in with the next unused
+ * generic index (in this example, 2).
+ */
+void
+svga_remap_generics(unsigned generics_mask,
+                    int8_t remap_table[MAX_GENERIC_VARYING])
+{
+   /* Note texcoord[0] is reserved so start at 1 */
+   unsigned count = 1, i;
+
+   for (i = 0; i < MAX_GENERIC_VARYING; i++) {
+      remap_table[i] = -1;
+   }
+
+   /* for each bit set in generic_mask */
+   while (generics_mask) {
+      unsigned index = ffs(generics_mask) - 1;
+      remap_table[index] = count++;
+      generics_mask &= ~(1 << index);
+   }
+
+   for (i = 0; i < MAX_GENERIC_VARYING; i++) {
+      if (remap_table[i] == -1)
+         remap_table[i] = count;
+   }
+}
+
+
+/**
+ * Use the generic remap table to map a TGSI generic varying variable
+ * index to a small integer.
+ */
+int
+svga_remap_generic_index(const int8_t remap_table[MAX_GENERIC_VARYING],
+                         int generic_index)
+{
+   assert(generic_index < MAX_GENERIC_VARYING);
+
+   if (generic_index >= MAX_GENERIC_VARYING) {
+      /* just don't return a random/garbage value */
+      generic_index = MAX_GENERIC_VARYING - 1;
+   }
+
+   return remap_table[generic_index];
+}
 
 
 /* Parse TGSI shader and translate to SVGA/DX9 serialized
@@ -168,7 +244,7 @@ static boolean svga_shader_emit_header( struct svga_shader_emitter *emit )
  */
 static struct svga_shader_result *
 svga_tgsi_translate( const struct svga_shader *shader,
-                     union svga_compile_key key,
+                     struct svga_compile_key key,
                      unsigned unit )
 {
    struct svga_shader_result *result = NULL;
@@ -248,9 +324,13 @@ struct svga_shader_result *
 svga_translate_fragment_program( const struct svga_fragment_shader *fs,
                                  const struct svga_fs_compile_key *fkey )
 {
-   union svga_compile_key key;
+   struct svga_compile_key key;
+
    memcpy(&key.fkey, fkey, sizeof *fkey);
 
+   memcpy(key.generic_remap_table, fs->generic_remap_table,
+          sizeof(fs->generic_remap_table));
+
    return svga_tgsi_translate( &fs->base, 
                                key,
                                PIPE_SHADER_FRAGMENT );
@@ -260,9 +340,15 @@ struct svga_shader_result *
 svga_translate_vertex_program( const struct svga_vertex_shader *vs,
                                const struct svga_vs_compile_key *vkey )
 {
-   union svga_compile_key key;
+   struct svga_compile_key key;
+
    memcpy(&key.vkey, vkey, sizeof *vkey);
 
+   /* Note: we could alternately store the remap table in the vkey but
+    * that would make it larger.  We just regenerate it here instead.
+    */
+   svga_remap_generics(vkey->fs_generic_inputs, key.generic_remap_table);
+
    return svga_tgsi_translate( &vs->base, 
                                key,
                                PIPE_SHADER_VERTEX );

src/gallium/drivers/svga/svga_tgsi.h

@@ -30,6 +30,13 @@
 
 #include "svga_hw_reg.h"
 
+
+/**
+ * We use a 32-bit mask to keep track of the generic indexes.
+ */
+#define MAX_GENERIC_VARYING 32
+
+
 struct svga_fragment_shader;
 struct svga_vertex_shader;
 struct svga_shader;
@@ -39,6 +46,7 @@ struct tgsi_token;
 
 struct svga_vs_compile_key
 {
+   unsigned fs_generic_inputs;
    unsigned zero_stride_vertex_elements;
    unsigned need_prescale:1;
    unsigned allow_psiz:1;
@@ -67,9 +75,10 @@ struct svga_fs_compile_key
    } tex[PIPE_MAX_SAMPLERS];
 };
 
-union svga_compile_key {
+struct svga_compile_key {
    struct svga_vs_compile_key vkey;
    struct svga_fs_compile_key fkey;
+   int8_t generic_remap_table[MAX_GENERIC_VARYING];
 };
 
 struct svga_shader_result
@@ -78,7 +87,7 @@ struct svga_shader_result
 
    /* Parameters used to generate this compilation result:
     */
-   union svga_compile_key key;
+   struct svga_compile_key key;
 
    /* Compiled shader tokens:
     */
@@ -140,4 +149,18 @@ svga_translate_vertex_program( const struct svga_vertex_shader *fs,
 
 void svga_destroy_shader_result( struct svga_shader_result *result );
 
+unsigned
+svga_get_generic_inputs_mask(const struct tgsi_shader_info *info);
+
+unsigned
+svga_get_generic_outputs_mask(const struct tgsi_shader_info *info);
+
+void
+svga_remap_generics(unsigned generics_mask,
+                    int8_t remap_table[MAX_GENERIC_VARYING]);
+
+int
+svga_remap_generic_index(const int8_t remap_table[MAX_GENERIC_VARYING],
+                         int generic_index);
+
 #endif

src/gallium/drivers/svga/svga_tgsi_decl_sm30.c

@@ -30,7 +30,9 @@
 
 #include "svga_tgsi_emit.h"
 
-static boolean translate_vs_ps_semantic( struct tgsi_declaration_semantic semantic,
+
+static boolean translate_vs_ps_semantic( struct svga_shader_emitter *emit,
+                                         struct tgsi_declaration_semantic semantic,
                                          unsigned *usage,
                                          unsigned *idx )
 {
@@ -58,7 +60,8 @@ static boolean translate_vs_ps_semantic( struct tgsi_declaration_semantic semant
       *usage = SVGA3D_DECLUSAGE_PSIZE;
       break;
    case TGSI_SEMANTIC_GENERIC:   
-      *idx = semantic.Index + 1; /* texcoord[0] is reserved for fog & position */
+      *idx = svga_remap_generic_index(emit->key.generic_remap_table,
+                                      semantic.Index);
       *usage = SVGA3D_DECLUSAGE_TEXCOORD;
       break;
    case TGSI_SEMANTIC_NORMAL:    
@@ -84,6 +87,10 @@ static boolean emit_decl( struct svga_shader_emitter *emit,
    SVGA3DOpDclArgs dcl;
    SVGA3dShaderInstToken opcode;
 
+   /* check values against bitfield sizes */
+   assert(index < 16);
+   assert(usage <= SVGA3D_DECLUSAGE_MAX);
+
    opcode = inst_token( SVGA3DOP_DCL );
    dcl.values[0] = 0;
    dcl.values[1] = 0;
@@ -181,7 +188,7 @@ static boolean ps30_input( struct svga_shader_emitter *emit,
    else if (emit->key.fkey.light_twoside &&
             (semantic.Name == TGSI_SEMANTIC_COLOR)) {
 
-      if (!translate_vs_ps_semantic( semantic, &usage, &index ))
+      if (!translate_vs_ps_semantic( emit, semantic, &usage, &index ))
          return FALSE;
 
       emit->internal_color_idx[emit->internal_color_count] = idx;
@@ -195,7 +202,7 @@ static boolean ps30_input( struct svga_shader_emitter *emit,
          return FALSE;
 
       semantic.Name = TGSI_SEMANTIC_BCOLOR;
-      if (!translate_vs_ps_semantic( semantic, &usage, &index ))
+      if (!translate_vs_ps_semantic( emit, semantic, &usage, &index ))
          return FALSE;
 
       if (emit->ps30_input_count >= SVGA3D_INPUTREG_MAX)
@@ -234,7 +241,7 @@ static boolean ps30_input( struct svga_shader_emitter *emit,
    }
    else {
 
-      if (!translate_vs_ps_semantic( semantic, &usage, &index ))
+      if (!translate_vs_ps_semantic( emit, semantic, &usage, &index ))
          return FALSE;
 
       if (emit->ps30_input_count >= SVGA3D_INPUTREG_MAX)
@@ -398,7 +405,7 @@ static boolean vs30_output( struct svga_shader_emitter *emit,
    dcl.values[0] = 0;
    dcl.values[1] = 0;
 
-   if (!translate_vs_ps_semantic( semantic, &usage, &index ))
+   if (!translate_vs_ps_semantic( emit, semantic, &usage, &index ))
       return FALSE;
 
    if (emit->vs30_output_count >= SVGA3D_OUTPUTREG_MAX)

src/gallium/drivers/svga/svga_tgsi_emit.h

@@ -54,7 +54,7 @@ struct svga_shader_emitter
    char *buf;
    char *ptr;
 
-   union svga_compile_key key;
+   struct svga_compile_key key;
    struct tgsi_shader_info info;
    int unit;
 
@@ -262,6 +262,11 @@ dst_register( unsigned file,
 {
    SVGA3dShaderDestToken dest;
 
+   /* check values against bitfield sizes */
+   assert(number < (1 << 11));
+   assert((file >> 3) < 4);
+   assert((file & 0x7) < 8);
+
    dest.value = 0;
    dest.num = number;
    dest.type_upper = file >> 3;
@@ -291,6 +296,11 @@ src_token( unsigned file, int number )
 {
    SVGA3dShaderSrcToken src;
 
+   /* check values against bitfield sizes */
+   assert(number < (1 << 11));
+   assert((file >> 3) < 4);
+   assert((file & 0x7) < 8);
+
    src.value = 0;
    src.num = number;
    src.type_upper = file >> 3;