nir: optimize atomic isub if supported

Foz-DB Navi48:
Totals from 1 (0.00% of 80287) affected shaders:
Instrs: 1641 -> 1637 (-0.24%)
CodeSize: 8472 -> 8456 (-0.19%)
Latency: 19132 -> 19131 (-0.01%)
InvThroughput: 9566 -> 9565 (-0.01%)
Copies: 126 -> 125 (-0.79%)
VALU: 565 -> 563 (-0.35%)
SALU: 439 -> 438 (-0.23%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/37702>

src/amd/common/nir/ac_nir.c

@@ -90,6 +90,7 @@ void ac_nir_set_options(struct radeon_info *info, bool use_llvm,
    options->has_shfr32 = true;
    options->has_mul24_relaxed = true;
    options->has_f2e4m3fn_satfn = !use_llvm && info->gfx_level >= GFX12;
+   options->has_atomic_isub = true;
    options->lower_int64_options = nir_lower_imul64 | nir_lower_imul_high64 | nir_lower_imul_2x32_64 | nir_lower_divmod64 |
                                   nir_lower_minmax64 | nir_lower_iabs64 | nir_lower_iadd_sat64 | nir_lower_conv64 |
                                   nir_lower_bitfield_extract64;

src/compiler/nir/nir_opt_intrinsics.c

@@ -324,6 +324,26 @@ try_opt_exclusive_scan_to_inclusive(nir_builder *b, nir_intrinsic_instr *intrin)
    return true;
 }
 
+static bool
+try_opt_atomic_isub(nir_builder *b, nir_intrinsic_instr *intrin,
+                    const struct nir_shader_compiler_options *options,
+                    unsigned data_idx)
+{
+   if (nir_intrinsic_atomic_op(intrin) != nir_atomic_op_iadd || !options->has_atomic_isub)
+      return false;
+
+   nir_scalar data = nir_scalar_resolved(intrin->src[data_idx].ssa, 0);
+
+   if (!nir_scalar_is_alu(data) || nir_scalar_alu_op(data) != nir_op_ineg)
+      return false;
+
+   data = nir_scalar_chase_alu_src(data, 0);
+
+   nir_src_rewrite(&intrin->src[data_idx], nir_mov_scalar(b, data));
+   nir_intrinsic_set_atomic_op(intrin, nir_atomic_op_isub);
+   return true;
+}
+
 static bool
 opt_intrinsics_intrin(nir_builder *b, nir_intrinsic_instr *intrin,
                       const struct nir_shader_compiler_options *options)
@@ -379,6 +399,17 @@ opt_intrinsics_intrin(nir_builder *b, nir_intrinsic_instr *intrin,
    }
    case nir_intrinsic_exclusive_scan:
       return try_opt_exclusive_scan_to_inclusive(b, intrin);
+   case nir_intrinsic_shared_atomic:
+   case nir_intrinsic_global_atomic:
+   case nir_intrinsic_global_atomic_amd:
+   case nir_intrinsic_deref_atomic:
+      return try_opt_atomic_isub(b, intrin, options, 1);
+   case nir_intrinsic_ssbo_atomic:
+      return try_opt_atomic_isub(b, intrin, options, 2);
+   case nir_intrinsic_image_deref_atomic:
+   case nir_intrinsic_image_atomic:
+   case nir_intrinsic_bindless_image_atomic:
+      return try_opt_atomic_isub(b, intrin, options, 3);
    default:
       return false;
    }

src/compiler/nir/nir_shader_compiler_options.h

@@ -649,6 +649,9 @@ typedef struct nir_shader_compiler_options {
    /** Backend supports f2i32_rtne opcode. */
    bool has_f2i32_rtne;
 
+   /** Backend supports atomic isub. */
+   bool has_atomic_isub;
+
    /**
     * Is this the Intel vec4 backend?
     *