[Mesa-dev] [PATCH 3/3] ac/nir: set cache policy when loading/storing buffer images

[Samuel Pitoiset]

This was missing.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset at gmail.com>
---
 src/amd/common/ac_nir_to_llvm.c | 23 ++++++++++-------------
 1 file changed, 10 insertions(+), 13 deletions(-)

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index 736b726beb..eaf400e2fe 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -2334,9 +2334,12 @@ static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
 	LLVMValueRef res;
 	const nir_variable *var = get_image_variable(instr);
 	const struct glsl_type *type = var->type;
+	struct ac_image_args args = {};
 
 	type = glsl_without_array(type);
 
+	args.cache_policy = get_cache_policy(ctx, var->data.image.access, false);
+
 	const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
 	if (dim == GLSL_SAMPLER_DIM_BUF) {
 		unsigned mask = nir_ssa_def_components_read(&instr->dest.ssa);
@@ -2347,16 +2350,16 @@ static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
 		vindex = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
 						 ctx->ac.i32_0, "");
 
-		/* TODO: set "glc" and "can_speculate" when OpenGL needs it. */
+		/* TODO: set "can_speculate" when OpenGL needs it. */
 		res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
 						  ctx->ac.i32_0, num_channels,
-						  false, false);
+						  !!(args.cache_policy & ac_glc),
+						  false);
 		res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
 
 		res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);
 		res = ac_to_integer(&ctx->ac, res);
 	} else {
-		struct ac_image_args args = {};
 		args.opcode = ac_image_load;
 		get_image_coords(ctx, instr, &args);
 		args.resource = get_image_descriptor(ctx, instr, AC_DESC_IMAGE, false);
@@ -2364,8 +2367,6 @@ static LLVMValueRef visit_image_load(struct ac_nir_context *ctx,
 					    glsl_sampler_type_is_array(type));
 		args.dmask = 15;
 		args.attributes = AC_FUNC_ATTR_READONLY;
-		args.cache_policy =
-			get_cache_policy(ctx, var->data.image.access, false);
 
 		res = ac_build_image_opcode(&ctx->ac, &args);
 	}
@@ -2379,10 +2380,9 @@ static void visit_image_store(struct ac_nir_context *ctx,
 	const nir_variable *var = get_image_variable(instr);
 	const struct glsl_type *type = glsl_without_array(var->type);
 	const enum glsl_sampler_dim dim = glsl_get_sampler_dim(type);
-	LLVMValueRef glc = ctx->ac.i1false;
-	bool force_glc = ctx->ac.chip_class == SI;
-	if (force_glc)
-		glc = ctx->ac.i1true;
+	struct ac_image_args args = {};
+
+	args.cache_policy = get_cache_policy(ctx, var->data.image.access, true);
 
 	if (dim == GLSL_SAMPLER_DIM_BUF) {
 		LLVMValueRef rsrc = get_image_buffer_descriptor(ctx, instr, true);
@@ -2392,12 +2392,11 @@ static void visit_image_store(struct ac_nir_context *ctx,
 		params[2] = LLVMBuildExtractElement(ctx->ac.builder, get_src(ctx, instr->src[1]),
 						    ctx->ac.i32_0, ""); /* vindex */
 		params[3] = ctx->ac.i32_0; /* voffset */
-		params[4] = glc;  /* glc */
+		params[4] = LLVMConstInt(ctx->ac.i1, !!(args.cache_policy & ac_glc), 0);
 		params[5] = ctx->ac.i1false;  /* slc */
 		ac_build_intrinsic(&ctx->ac, "llvm.amdgcn.buffer.store.format.v4f32", ctx->ac.voidt,
 				   params, 6, 0);
 	} else {
-		struct ac_image_args args = {};
 		args.opcode = ac_image_store;
 		args.data[0] = ac_to_float(&ctx->ac, get_src(ctx, instr->src[3]));
 		get_image_coords(ctx, instr, &args);
@@ -2405,8 +2404,6 @@ static void visit_image_store(struct ac_nir_context *ctx,
 		args.dim = get_ac_image_dim(&ctx->ac, glsl_get_sampler_dim(type),
 					    glsl_sampler_type_is_array(type));
 		args.dmask = 15;
-		args.cache_policy =
-			get_cache_policy(ctx, var->data.image.access, true);
 
 		ac_build_image_opcode(&ctx->ac, &args);
 	}
-- 
2.19.1