[Mesa-dev] [PATCH 05/13] radeonsi: use ordered compares for SSG and face selection

[Marek Olšák]

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

Ordered compares are what you have in C. Unordered compares are the result
of negating ordered compares (they return true if either argument is NaN).

That special NaN behavior is completely useless here, and unordered
compares produce horrible code with all stable LLVM versions.
(I think that has been fixed in LLVM git)
---
 src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c | 4 ++--
 src/gallium/drivers/radeonsi/si_shader.c            | 2 +-
 2 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
index c30a9d0..dce5b55 100644
--- a/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
+++ b/src/gallium/drivers/radeon/radeon_setup_tgsi_llvm.c
@@ -1107,9 +1107,9 @@ static void emit_ssg(
 		cmp = LLVMBuildICmp(builder, LLVMIntSGE, val, bld_base->int_bld.zero, "");
 		val = LLVMBuildSelect(builder, cmp, val, LLVMConstInt(bld_base->int_bld.elem_type, -1, true), "");
 	} else { // float SSG
-		cmp = LLVMBuildFCmp(builder, LLVMRealUGT, emit_data->args[0], bld_base->base.zero, "");
+		cmp = LLVMBuildFCmp(builder, LLVMRealOGT, emit_data->args[0], bld_base->base.zero, "");
 		val = LLVMBuildSelect(builder, cmp, bld_base->base.one, emit_data->args[0], "");
-		cmp = LLVMBuildFCmp(builder, LLVMRealUGE, val, bld_base->base.zero, "");
+		cmp = LLVMBuildFCmp(builder, LLVMRealOGE, val, bld_base->base.zero, "");
 		val = LLVMBuildSelect(builder, cmp, val, LLVMConstReal(bld_base->base.elem_type, -1), "");
 	}
 
diff --git a/src/gallium/drivers/radeonsi/si_shader.c b/src/gallium/drivers/radeonsi/si_shader.c
index ce59f0e..eb762c0 100644
--- a/src/gallium/drivers/radeonsi/si_shader.c
+++ b/src/gallium/drivers/radeonsi/si_shader.c
@@ -484,7 +484,7 @@ static void declare_input_fs(
 		face = LLVMGetParam(main_fn, SI_PARAM_FRONT_FACE);
 
 		is_face_positive = LLVMBuildFCmp(gallivm->builder,
-						 LLVMRealUGT, face,
+						 LLVMRealOGT, face,
 						 lp_build_const_float(gallivm, 0.0f),
 						 "");
 
-- 
2.1.0