nir: add d2i, d2u, d2b opcodes

[mesa.git] / src / compiler / nir / nir_opt_constant_folding.c

diff --git a/src/compiler/nir/nir_opt_constant_folding.c b/src/compiler/nir/nir_opt_constant_folding.c
index 20b31a8110cd4298e9efebfbea32c5242e4e8272..caa4231b1884d3728639c135e595897ec3f646e4 100644 (file)
--- a/src/compiler/nir/nir_opt_constant_folding.c
+++ b/src/compiler/nir/nir_opt_constant_folding.c
@@ -46,10 +46,28 @@ constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
    if (!instr->dest.dest.is_ssa)
       return false;
 
+   /* In the case that any outputs/inputs have unsized types, then we need to
+    * guess the bit-size. In this case, the validator ensures that all
+    * bit-sizes match so we can just take the bit-size from first
+    * output/input with an unsized type. If all the outputs/inputs are sized
+    * then we don't need to guess the bit-size at all because the code we
+    * generate for constant opcodes in this case already knows the sizes of
+    * the types involved and does not need the provided bit-size for anything
+    * (although it still requires to receive a valid bit-size).
+    */
+   unsigned bit_size = 0;
+   if (!nir_alu_type_get_type_size(nir_op_infos[instr->op].output_type))
+      bit_size = instr->dest.dest.ssa.bit_size;
+
    for (unsigned i = 0; i < nir_op_infos[instr->op].num_inputs; i++) {
       if (!instr->src[i].src.is_ssa)
          return false;
 
+      if (bit_size == 0 &&
+          !nir_alu_type_get_type_size(nir_op_infos[instr->op].input_sizes[i])) {
+         bit_size = instr->src[i].src.ssa->bit_size;
+      }
+
       nir_instr *src_instr = instr->src[i].src.ssa->parent_instr;
 
       if (src_instr->type != nir_instr_type_load_const)
@@ -58,23 +76,30 @@ constant_fold_alu_instr(nir_alu_instr *instr, void *mem_ctx)
 
       for (unsigned j = 0; j < nir_ssa_alu_instr_src_components(instr, i);
            j++) {
-         src[i].u[j] = load_const->value.u[instr->src[i].swizzle[j]];
+         if (load_const->def.bit_size == 64)
+            src[i].u64[j] = load_const->value.u64[instr->src[i].swizzle[j]];
+         else
+            src[i].u32[j] = load_const->value.u32[instr->src[i].swizzle[j]];
       }
 
       /* We shouldn't have any source modifiers in the optimization loop. */
       assert(!instr->src[i].abs && !instr->src[i].negate);
    }
 
+   if (bit_size == 0)
+      bit_size = 32;
+
    /* We shouldn't have any saturate modifiers in the optimization loop. */
    assert(!instr->dest.saturate);
 
    nir_const_value dest =
       nir_eval_const_opcode(instr->op, instr->dest.dest.ssa.num_components,
-                            src);
+                            bit_size, src);
 
    nir_load_const_instr *new_instr =
       nir_load_const_instr_create(mem_ctx,
-                                  instr->dest.dest.ssa.num_components);
+                                  instr->dest.dest.ssa.num_components,
+                                  instr->dest.dest.ssa.bit_size);
 
    new_instr->value = dest;
 
@@ -106,7 +131,7 @@ constant_fold_deref(nir_instr *instr, nir_deref_var *deref)
          nir_load_const_instr *indirect =
             nir_instr_as_load_const(arr->indirect.ssa->parent_instr);
 
-         arr->base_offset += indirect->value.u[0];
+         arr->base_offset += indirect->value.u32[0];
 
          /* Clear out the source */
          nir_instr_rewrite_src(instr, &arr->indirect, nir_src_for_ssa(NULL));
@@ -136,10 +161,15 @@ constant_fold_intrinsic_instr(nir_intrinsic_instr *instr)
 static bool
 constant_fold_tex_instr(nir_tex_instr *instr)
 {
+   bool progress = false;
+
    if (instr->texture)
-      return constant_fold_deref(&instr->instr, instr->texture);
-   else
-      return false;
+      progress |= constant_fold_deref(&instr->instr, instr->texture);
+
+   if (instr->sampler)
+      progress |= constant_fold_deref(&instr->instr, instr->sampler);
+
+   return progress;
 }
 
 static bool