1 ; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s 2 3 declare float @llvm.AMDGPU.rcp.f32(float) nounwind readnone 4 declare double @llvm.AMDGPU.rcp.f64(double) nounwind readnone 5 6 7 declare float @llvm.sqrt.f32(float) nounwind readnone 8 declare double @llvm.sqrt.f64(double) nounwind readnone 9 10 ; FUNC-LABEL: @rcp_f32 11 ; SI: V_RCP_F32_e32 12 define void @rcp_f32(float addrspace(1)* %out, float %src) nounwind { 13 %rcp = call float @llvm.AMDGPU.rcp.f32(float %src) nounwind readnone 14 store float %rcp, float addrspace(1)* %out, align 4 15 ret void 16 } 17 18 ; FUNC-LABEL: @rcp_f64 19 ; SI: V_RCP_F64_e32 20 define void @rcp_f64(double addrspace(1)* %out, double %src) nounwind { 21 %rcp = call double @llvm.AMDGPU.rcp.f64(double %src) nounwind readnone 22 store double %rcp, double addrspace(1)* %out, align 8 23 ret void 24 } 25 26 ; FUNC-LABEL: @rcp_pat_f32 27 ; SI: V_RCP_F32_e32 28 define void @rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind { 29 %rcp = fdiv float 1.0, %src 30 store float %rcp, float addrspace(1)* %out, align 4 31 ret void 32 } 33 34 ; FUNC-LABEL: @rcp_pat_f64 35 ; SI: V_RCP_F64_e32 36 define void @rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind { 37 %rcp = fdiv double 1.0, %src 38 store double %rcp, double addrspace(1)* %out, align 8 39 ret void 40 } 41 42 ; FUNC-LABEL: @rsq_rcp_pat_f32 43 ; SI: V_RSQ_F32_e32 44 define void @rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind { 45 %sqrt = call float @llvm.sqrt.f32(float %src) nounwind readnone 46 %rcp = call float @llvm.AMDGPU.rcp.f32(float %sqrt) nounwind readnone 47 store float %rcp, float addrspace(1)* %out, align 4 48 ret void 49 } 50 51 ; FUNC-LABEL: @rsq_rcp_pat_f64 52 ; SI: V_RSQ_F64_e32 53 define void @rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind { 54 %sqrt = call double @llvm.sqrt.f64(double %src) nounwind readnone 55 %rcp = call double @llvm.AMDGPU.rcp.f64(double %sqrt) nounwind readnone 56 store double %rcp, double addrspace(1)* %out, align 8 57 ret void 58 } 59
