nir: Add a load_global_constant intrinsic

This has the same semantics as load_global except the memory it reads is
known to be constant so load_global_constant intrinsics can be CSEd
rather than relying on more complex copy-propagation.

Reviewed-by: Karol Herbst <kherbst@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6379>

src/compiler/nir/nir_divergence_analysis.c

@@ -255,6 +255,7 @@ visit_intrinsic(nir_intrinsic_instr *instr, struct divergence_state *state)
    case nir_intrinsic_load_ssbo:
    case nir_intrinsic_load_shared:
    case nir_intrinsic_load_global:
+   case nir_intrinsic_load_global_constant:
    case nir_intrinsic_load_uniform:
    case nir_intrinsic_load_push_constant:
    case nir_intrinsic_load_constant:

src/compiler/nir/nir_intrinsics.py

@@ -768,6 +768,9 @@ load("constant", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET],
 # src[] = { address }.
 load("global", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])
 # src[] = { address }.
+load("global_constant", [1], [ACCESS, ALIGN_MUL, ALIGN_OFFSET],
+     [CAN_ELIMINATE, CAN_REORDER])
+# src[] = { address }.
 load("kernel_input", [1], [BASE, RANGE, ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE, CAN_REORDER])
 # src[] = { offset }.
 load("scratch", [1], [ALIGN_MUL, ALIGN_OFFSET], [CAN_ELIMINATE])

src/compiler/nir/nir_lower_io.c

@@ -1632,6 +1632,7 @@ nir_get_io_offset_src(nir_intrinsic_instr *instr)
    case nir_intrinsic_load_shared:
    case nir_intrinsic_load_uniform:
    case nir_intrinsic_load_global:
+   case nir_intrinsic_load_global_constant:
    case nir_intrinsic_load_scratch:
    case nir_intrinsic_load_fs_input_interp_deltas:
    case nir_intrinsic_shared_atomic_add:

src/compiler/nir/nir_lower_phis_to_scalar.c

@@ -103,6 +103,7 @@ is_phi_src_scalarizable(nir_phi_src *src,
       case nir_intrinsic_load_ubo:
       case nir_intrinsic_load_ssbo:
       case nir_intrinsic_load_global:
+      case nir_intrinsic_load_global_constant:
       case nir_intrinsic_load_input:
          return true;
       default:

src/compiler/nir/nir_opt_gcm.c

@@ -157,6 +157,7 @@ is_src_scalarizable(nir_src *src)
       case nir_intrinsic_load_ubo:
       case nir_intrinsic_load_ssbo:
       case nir_intrinsic_load_global:
+      case nir_intrinsic_load_global_constant:
       case nir_intrinsic_load_input:
          return true;
       default:

src/compiler/nir/nir_opt_shrink_vectors.c

@@ -123,6 +123,7 @@ opt_shrink_vectors_intrinsic(nir_builder *b, nir_intrinsic_instr *instr)
    case nir_intrinsic_load_push_constant:
    case nir_intrinsic_load_constant:
    case nir_intrinsic_load_global:
+   case nir_intrinsic_load_global_constant:
    case nir_intrinsic_load_kernel_input:
    case nir_intrinsic_load_scratch:
    case nir_intrinsic_store_output:

src/compiler/nir/nir_validate.c

@@ -598,6 +598,7 @@ validate_intrinsic_instr(nir_intrinsic_instr *instr, validate_state *state)
    case nir_intrinsic_load_ssbo:
    case nir_intrinsic_load_shared:
    case nir_intrinsic_load_global:
+   case nir_intrinsic_load_global_constant:
    case nir_intrinsic_load_scratch:
    case nir_intrinsic_load_constant:
       /* These memory load operations must have alignments */