From 22ca3f886006e4500c43f8f0cf5a5c955aa310cc Mon Sep 17 00:00:00 2001 From: Matt Arsenault Date: Tue, 15 Jul 2014 23:50:10 +0000 Subject: R600/SI: Allow using f32 rcp / rsq when denormals not handled. These are precise enough to use for OpenCL unless denormals are handled. llvm-svn: 213107 --- llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll | 18 +++++++----------- 1 file changed, 7 insertions(+), 11 deletions(-) (limited to 'llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll') diff --git a/llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll b/llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll index 42910a99afd..8d5d66e149b 100644 --- a/llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll +++ b/llvm/test/CodeGen/R600/llvm.AMDGPU.rcp.ll @@ -1,5 +1,7 @@ -; RUN: llc -march=r600 -mcpu=SI -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI -check-prefix=FUNC %s -; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s +; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -enable-unsafe-fp-math -verify-machineinstrs < %s | FileCheck -check-prefix=SI-UNSAFE -check-prefix=SI -check-prefix=FUNC %s +; RUN: llc -march=r600 -mcpu=SI -mattr=-fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE -check-prefix=SI -check-prefix=FUNC %s + +; XUN: llc -march=r600 -mcpu=SI -mattr=+fp32-denormals -verify-machineinstrs < %s | FileCheck -check-prefix=SI-SAFE-SPDENORM -check-prefix=SI -check-prefix=FUNC %s declare float @llvm.AMDGPU.rcp.f32(float) nounwind readnone declare double @llvm.AMDGPU.rcp.f64(double) nounwind readnone @@ -25,15 +27,8 @@ define void @rcp_f64(double addrspace(1)* %out, double %src) nounwind { } ; FUNC-LABEL: @rcp_pat_f32 -; SI-UNSAFE-NOT: V_MUL_F32 -; SI-UNSAFE: V_RCP_F32_e32 -; SI-UNSAFE-NOT: V_MUL_F32 - -; Check for surrounding multiplies the correct divide has. -; SI-SAFE: V_MUL_F32 ; SI-SAFE: V_RCP_F32_e32 -; SI-SAFE: V_MUL_F32 - +; XSI-SAFE-SPDENORM-NOT: V_RCP_F32_e32 define void @rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind { %rcp = fdiv float 1.0, %src store float %rcp, float addrspace(1)* %out, align 4 @@ -60,7 +55,8 @@ define void @rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind { } ; FUNC-LABEL: @rsq_rcp_pat_f64 -; SI: V_RSQ_F64_e32 +; SI-UNSAFE: V_RSQ_F64_e32 +; SI-SAFE-NOT: V_RSQ_F64_e32 define void @rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind { %sqrt = call double @llvm.sqrt.f64(double %src) nounwind readnone %rcp = call double @llvm.AMDGPU.rcp.f64(double %sqrt) nounwind readnone -- cgit v1.2.3