From: Connor Abbott <cwabbott0@gmail.com>
Date: Tue, 4 Jun 2019 12:42:54 +0000 (+0200)
Subject: ac/nir: Set speculatable for buffer loads where allowed
X-Git-Tag: upstream/19.3.0~5335
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=53a7649e5de22d6d30dc43a129d60efdc5f3c7be;p=platform%2Fupstream%2Fmesa.git

ac/nir: Set speculatable for buffer loads where allowed

This brings the nir path in line with the TGSI path.

Totals from affected shaders:
SGPRS: 2984 -> 2984 (0.00 %)
VGPRS: 2792 -> 2652 (-5.01 %)
Spilled SGPRs: 0 -> 0 (0.00 %)
Spilled VGPRs: 0 -> 0 (0.00 %)
Private memory VGPRs: 0 -> 0 (0.00 %)
Scratch size: 0 -> 0 (0.00 %) dwords per thread
Code Size: 247380 -> 248072 (0.28 %) bytes
LDS: 0 -> 0 (0.00 %) blocks
Max Waves: 121 -> 132 (9.09 %)
Wait states: 0 -> 0 (0.00 %)

Most of the change came from DiRT: Showdown, and came from sinking SSBO
loads.

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
---

diff --git a/src/amd/common/ac_nir_to_llvm.c b/src/amd/common/ac_nir_to_llvm.c
index b73a0f5..dffaeed 100644
--- a/src/amd/common/ac_nir_to_llvm.c
+++ b/src/amd/common/ac_nir_to_llvm.c
@@ -1779,11 +1779,12 @@ static LLVMValueRef visit_load_buffer(struct ac_nir_context *ctx,
 							 cache_policy & ac_glc);
 		} else {
 			int num_channels = util_next_power_of_two(load_bytes) / 4;
+			bool can_speculate = access & ACCESS_CAN_REORDER;
 
 			ret = ac_build_buffer_load(&ctx->ac, rsrc, num_channels,
 						   vindex, offset, immoffset, 0,
 						   cache_policy & ac_glc, 0,
-						   false, false);
+						   can_speculate, false);
 		}
 
 		LLVMTypeRef byte_vec = LLVMVectorType(ctx->ac.i8, ac_get_type_size(LLVMTypeOf(ret)));
@@ -2489,11 +2490,11 @@ 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 "can_speculate" when OpenGL needs it. */
+		bool can_speculate = access & ACCESS_CAN_REORDER;
 		res = ac_build_buffer_load_format(&ctx->ac, rsrc, vindex,
 						  ctx->ac.i32_0, num_channels,
 						  !!(args.cache_policy & ac_glc),
-						  false);
+						  can_speculate);
 		res = ac_build_expand_to_vec4(&ctx->ac, res, num_channels);
 
 		res = ac_trim_vector(&ctx->ac, res, instr->dest.ssa.num_components);