nir: Use b2b opcodes for shared and constant memory

No shader-db changes on ICL with iris

Reviewed-by: Alyssa Rosenzweig <[email protected]>
Reviewed-by: Kenneth Graunke <[email protected]>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4338>

3 files changed, 24 insertions, 17 deletions

diff --git a/src/compiler/glsl/glsl_to_nir.cpp b/src/compiler/glsl/glsl_to_nir.cpp
index fe045eaca9b..1101b9d39a6 100644
--- a/src/compiler/glsl/glsl_to_nir.cpp
+++ b/src/compiler/glsl/glsl_to_nir.cpp
@@ -1549,7 +1549,7 @@ nir_visitor::visit(ir_call *ir)
 
          /* The value in shared memory is a 32-bit value */
          if (type->is_boolean())
-            ret = nir_i2b(&b, &instr->dest.ssa);
+            ret = nir_b2b1(&b, &instr->dest.ssa);
          break;
       }
       case nir_intrinsic_store_shared: {
@@ -1571,7 +1571,7 @@ nir_visitor::visit(ir_call *ir)
          nir_ssa_def *nir_val = evaluate_rvalue(val);
          /* The value in shared memory is a 32-bit value */
          if (val->type->is_boolean())
-            nir_val = nir_b2i32(&b, nir_val);
+            nir_val = nir_b2b32(&b, nir_val);
 
          instr->src[0] = nir_src_for_ssa(nir_val);
          instr->num_components = val->type->vector_elements;
diff --git a/src/compiler/nir/nir_lower_io.c b/src/compiler/nir/nir_lower_io.c
index aac1e0e257e..92d2a1f8ba0 100644
--- a/src/compiler/nir/nir_lower_io.c
+++ b/src/compiler/nir/nir_lower_io.c
@@ -938,8 +938,16 @@ build_explicit_io_load(nir_builder *b, nir_intrinsic_instr *intrin,
       result = &load->dest.ssa;
    }
 
-   if (intrin->dest.ssa.bit_size == 1)
-      result = nir_i2b(b, result);
+   if (intrin->dest.ssa.bit_size == 1) {
+      /* For shared, we can go ahead and use NIR's and/or the back-end's
+       * standard encoding for booleans rather than forcing a 0/1 boolean.
+       * This should save an instruction or two.
+       */
+      if (mode == nir_var_mem_shared)
+         result = nir_b2b1(b, result);
+      else
+         result = nir_i2b(b, result);
+   }
 
    return result;
 }
@@ -974,8 +982,16 @@ build_explicit_io_store(nir_builder *b, nir_intrinsic_instr *intrin,
    nir_intrinsic_instr *store = nir_intrinsic_instr_create(b->shader, op);
 
    if (value->bit_size == 1) {
-      /* TODO: Make the native bool bit_size an option. */
-      value = nir_b2i(b, value, 32);
+      /* For shared, we can go ahead and use NIR's and/or the back-end's
+       * standard encoding for booleans rather than forcing a 0/1 boolean.
+       * This should save an instruction or two.
+       *
+       * TODO: Make the native bool bit_size an option.
+       */
+      if (mode == nir_var_mem_shared)
+         value = nir_b2b32(b, value);
+      else
+         value = nir_b2i(b, value, 32);
    }
 
    store->src[0] = nir_src_for_ssa(value);
diff --git a/src/compiler/nir/nir_opt_large_constants.c b/src/compiler/nir/nir_opt_large_constants.c
index a3c002f5f4c..e32265e50ec 100644
--- a/src/compiler/nir/nir_opt_large_constants.c
+++ b/src/compiler/nir/nir_opt_large_constants.c
@@ -86,19 +86,10 @@ build_constant_load(nir_builder *b, nir_deref_instr *deref,
    nir_builder_instr_insert(b, &load->instr);
 
    if (load->dest.ssa.bit_size < 8) {
-      /* Booleans are special-cased to be 32-bit
-       *
-       * Ideally, for drivers that can handle 32-bit booleans, we wouldn't
-       * emit the i2b here.  However, at this point, the driver is likely to
-       * still have 1-bit booleans so we need to at least convert bit sizes.
-       * Unfortunately, we don't have a good way to annotate the load as
-       * loading a known boolean value so the optimizer isn't going to be
-       * able to get rid of the conversion.  Some day, we may solve that
-       * problem but not today.
-       */
+      /* Booleans are special-cased to be 32-bit */
       assert(glsl_type_is_boolean(deref->type));
       load->dest.ssa.bit_size = 32;
-      return nir_i2b(b, &load->dest.ssa);
+      return nir_b2b1(b, &load->dest.ssa);
    } else {
       return &load->dest.ssa;
    }