[Mesa-dev] [PATCH 2/2] radeonsi: set no-signed-zeros-fp-math

[Marek Olšák]

From: Marek Olšák <marek.olsak at amd.com>

Recommended by Matt Arsenault.

46757 shaders in 28742 tests
Totals:
SGPRS: 2068851 -> 2066907 (-0.09 %)
VGPRS: 1604056 -> 1602676 (-0.09 %)
Spilled SGPRs: 1402 -> 1382 (-1.43 %)
Spilled VGPRs: 113 -> 113 (0.00 %)
Private memory VGPRs: 1332 -> 1332 (0.00 %)
Scratch size: 3224 -> 3188 (-1.12 %) dwords per thread
Code Size: 58815520 -> 58716788 (-0.17 %) bytes
LDS: 1162 -> 1162 (0.00 %) blocks
Max Waves: 354616 -> 354905 (0.08 %)
Wait states: 0 -> 0 (0.00 %)

Totals from affected shaders:
SGPRS: 786452 -> 784508 (-0.25 %)
VGPRS: 530000 -> 528620 (-0.26 %)
Spilled SGPRs: 958 -> 938 (-2.09 %)
Spilled VGPRs: 85 -> 85 (0.00 %)
Private memory VGPRs: 636 -> 636 (0.00 %)
Scratch size: 1880 -> 1844 (-1.91 %) dwords per thread
Code Size: 26349936 -> 26251204 (-0.37 %) bytes
LDS: 304 -> 304 (0.00 %) blocks
Max Waves: 108962 -> 109251 (0.27 %)
Wait states: 0 -> 0 (0.00 %)

CTS not tested.
---
 src/gallium/drivers/radeonsi/si_shader.c            | 4 ++++
 src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c | 2 +-
 2 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index 5c5f2e6..3af01cc 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -5490,20 +5490,24 @@ static void si_create_function(struct si_shader_context *ctx,
 		 * allows the optimization passes to move loads and reduces
 		 * SGPR spilling significantly.
 		 */
 		if (LLVMGetTypeKind(LLVMTypeOf(P)) == LLVMPointerTypeKind) {
 			lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_BYVAL);
 			lp_add_attr_dereferenceable(P, UINT64_MAX);
 		} else
 			lp_add_function_attr(ctx->main_fn, i + 1, LP_FUNC_ATTR_INREG);
 	}
 
+	LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
+					   "no-signed-zeros-fp-math",
+					   "true");
+
 	if (ctx->screen->b.debug_flags & DBG_UNSAFE_MATH) {
 		/* These were copied from some LLVM test. */
 		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
 						   "less-precise-fpmad",
 						   "true");
 		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
 						   "no-infs-fp-math",
 						   "true");
 		LLVMAddTargetDependentFunctionAttr(ctx->main_fn,
 						   "no-nans-fp-math",
diff --git a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
index 68153af..7830175 100644
--- a/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
+++ b/src/gallium/drivers/radeonsi/si_shader_tgsi_setup.c
@@ -1269,21 +1269,21 @@ void si_llvm_context_init(struct si_shader_context *ctx,
 	ctx->type = info ? info->processor : -1;
 
 	ctx->gallivm.context = LLVMContextCreate();
 	ctx->gallivm.module = LLVMModuleCreateWithNameInContext("tgsi",
 						ctx->gallivm.context);
 	LLVMSetTarget(ctx->gallivm.module, "amdgcn--");
 
 	bool unsafe_fpmath = (sscreen->b.debug_flags & DBG_UNSAFE_MATH) != 0;
 	enum lp_builder_type builder_type =
 		unsafe_fpmath ? LP_BUILDER_UNSAFE_FP_MATH :
-				LP_BUILDER_DEFAULT;
+				LP_BUILDER_NO_SIGNED_ZEROS_FP_MATH;
 
 	ctx->gallivm.builder = lp_create_builder(ctx->gallivm.context,
 						 builder_type);
 
 	ac_llvm_context_init(&ctx->ac, ctx->gallivm.context);
 	ctx->ac.module = ctx->gallivm.module;
 	ctx->ac.builder = ctx->gallivm.builder;
 
 	struct lp_build_tgsi_context *bld_base = &ctx->bld_base;
 
-- 
2.7.4