R600/SI: Match rsq instructions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@210226 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 3a4723ebb8490a8af71372a652117e241a9a76e2..1b6c35d6951d21a7e531c3727ac19a90175f3f2d 100644 (file)
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1017,12 +1017,16 @@ defm V_RSQ_LEGACY_F32 : VOP1_32 <
   0x0000002d, "V_RSQ_LEGACY_F32",
   [(set f32:$dst, (int_AMDGPU_rsq f32:$src0))]
 >;
-defm V_RSQ_F32 : VOP1_32 <0x0000002e, "V_RSQ_F32", []>;
+defm V_RSQ_F32 : VOP1_32 <0x0000002e, "V_RSQ_F32",
+  [(set f32:$dst, (fdiv FP_ONE, (fsqrt f32:$src0)))]
+>;
 defm V_RCP_F64 : VOP1_64 <0x0000002f, "V_RCP_F64",
   [(set f64:$dst, (fdiv FP_ONE, f64:$src0))]
 >;
 defm V_RCP_CLAMP_F64 : VOP1_64 <0x00000030, "V_RCP_CLAMP_F64", []>;
-defm V_RSQ_F64 : VOP1_64 <0x00000031, "V_RSQ_F64", []>;
+defm V_RSQ_F64 : VOP1_64 <0x00000031, "V_RSQ_F64",
+  [(set f64:$dst, (fdiv FP_ONE, (fsqrt f64:$src0)))]
+>;
 defm V_RSQ_CLAMP_F64 : VOP1_64 <0x00000032, "V_RSQ_CLAMP_F64", []>;
 defm V_SQRT_F32 : VOP1_32 <0x00000033, "V_SQRT_F32",
   [(set f32:$dst, (fsqrt f32:$src0))]

diff --git a/test/CodeGen/R600/rsq.ll b/test/CodeGen/R600/rsq.ll
new file mode 100644 (file)
index 0000000..87c0570
--- /dev/null
+++ b/test/CodeGen/R600/rsq.ll
@@ -0,0 +1,26 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare float @llvm.sqrt.f32(float) nounwind readnone
+declare double @llvm.sqrt.f64(double) nounwind readnone
+
+; SI-LABEL: @rsq_f32
+; SI: V_RSQ_F32_e32
+; SI: S_ENDPGM
+define void @rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
+  %val = load float addrspace(1)* %in, align 4
+  %sqrt = call float @llvm.sqrt.f32(float %val) nounwind readnone
+  %div = fdiv float 1.0, %sqrt
+  store float %div, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @rsq_f64
+; SI: V_RSQ_F64_e32
+; SI: S_ENDPGM
+define void @rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
+  %val = load double addrspace(1)* %in, align 4
+  %sqrt = call double @llvm.sqrt.f64(double %val) nounwind readnone
+  %div = fdiv double 1.0, %sqrt
+  store double %div, double addrspace(1)* %out, align 4
+  ret void
+}