AMDGPU: fdiv -1, x -> rcp -x

[nhaehnle]

lib/Target/AMDGPU/SIISelLowering.cpp

2450–2451

Comment: -1.0 / x -> rcp (fneg x)

test/CodeGen/AMDGPU/rsq.ll

79

SI-SAFE (E missing)

110

dito

[arsenm]

Fix comment and check lines

[nhaehnle]

The comment still says 1.0 instead of -1.0. With that fixed, LGTM.

[arsenm]

r277535

lib/Target/AMDGPU/SIISelLowering.cpp

SDValue SITargetLowering::lowerFastUnsafeFDIV(SDValue Op,
                                               SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue LHS = Op.getOperand(0);
   SDValue RHS = Op.getOperand(1);
   EVT VT = Op.getValueType();
   bool Unsafe = DAG.getTarget().Options.UnsafeFPMath;
 
   if (const ConstantFPSDNode *CLHS = dyn_cast<ConstantFPSDNode>(LHS)) {
-    if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals())) &&
-        CLHS->isExactlyValue(1.0)) {
+    if ((Unsafe || (VT == MVT::f32 && !Subtarget->hasFP32Denormals()))) {
+
+      if (CLHS->isExactlyValue(1.0)) {
         // v_rcp_f32 and v_rsq_f32 do not support denormals, and according to
         // the CI documentation has a worst case error of 1 ulp.
         // OpenCL requires <= 2.5 ulp for 1.0 / x, so it should always be OK to
         // use it as long as we aren't trying to use denormals.
 
         // 1.0 / sqrt(x) -> rsq(x)
         //
         // XXX - Is UnsafeFPMath sufficient to do this for f64? The maximum ULP
         // error seems really high at 2^29 ULP.
         if (RHS.getOpcode() == ISD::FSQRT)
           return DAG.getNode(AMDGPUISD::RSQ, SL, VT, RHS.getOperand(0));
 
         // 1.0 / x -> rcp(x)
         return DAG.getNode(AMDGPUISD::RCP, SL, VT, RHS);
       }
+
+      // Same as for 1.0, but expand the sign out of the constant.
+      if (CLHS->isExactlyValue(-1.0)) {
+        // 1.0 / x -> rcp (fabs x)
+        SDValue FNegRHS = DAG.getNode(ISD::FNEG, SL, VT, RHS);
+        return DAG.getNode(AMDGPUISD::RCP, SL, VT, FNegRHS);
+      }
+    }
   }
 
   const SDNodeFlags *Flags = Op->getFlags();
 
   if (Unsafe || Flags->hasAllowReciprocal()) {
     // Turn into multiply by the reciprocal.
     // x / y -> x * (1.0 / y)
     SDNodeFlags Flags;

test/CodeGen/AMDGPU/rcp-pattern.ll

 ; EG: RECIP_IEEE
 define void @rcp_fabs_pat_f32(float addrspace(1)* %out, float %src) #0 {
   %src.fabs = call float @llvm.fabs.f32(float %src)
   %rcp = fdiv float 1.0, %src.fabs
   store float %rcp, float addrspace(1)* %out, align 4
   ret void
 }
 
-; FIXME: fneg folded into constant 1
+; FUNC-LABEL: {{^}}neg_rcp_pat_f32:
+; GCN: s_load_dword [[SRC:s[0-9]+]]
+; GCN: v_rcp_f32_e64 [[RCP:v[0-9]+]], -[[SRC]]
+; GCN: buffer_store_dword [[RCP]]
+
+; EG: RECIP_IEEE
+define void @neg_rcp_pat_f32(float addrspace(1)* %out, float %src) #0 {
+  %rcp = fdiv float -1.0, %src
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
 ; FUNC-LABEL: {{^}}rcp_fabs_fneg_pat_f32:
+; GCN: s_load_dword [[SRC:s[0-9]+]]
+; GCN: v_rcp_f32_e64 [[RCP:v[0-9]+]], -|[[SRC]]|
+; GCN: buffer_store_dword [[RCP]]
 define void @rcp_fabs_fneg_pat_f32(float addrspace(1)* %out, float %src) #0 {
   %src.fabs = call float @llvm.fabs.f32(float %src)
   %src.fabs.fneg = fsub float -0.0, %src.fabs
   %rcp = fdiv float 1.0, %src.fabs.fneg
   store float %rcp, float addrspace(1)* %out, align 4
   ret void
 }
 
+; FUNC-LABEL: {{^}}rcp_fabs_fneg_pat_multi_use_f32:
+; GCN: s_load_dword [[SRC:s[0-9]+]]
+; GCN: v_rcp_f32_e64 [[RCP:v[0-9]+]], -|[[SRC]]|
+; GCN: buffer_store_dword [[RCP]]
+
+; GCN: v_mul_f32_e64 [[MUL:v[0-9]+]], [[SRC]], -|[[SRC]]|
+; GCN: buffer_store_dword [[MUL]]
+define void @rcp_fabs_fneg_pat_multi_use_f32(float addrspace(1)* %out, float %src) #0 {
+  %src.fabs = call float @llvm.fabs.f32(float %src)
+  %src.fabs.fneg = fsub float -0.0, %src.fabs
+  %rcp = fdiv float 1.0, %src.fabs.fneg
+  store volatile float %rcp, float addrspace(1)* %out, align 4
+
+  %other = fmul float %src, %src.fabs.fneg
+  store volatile float %other, float addrspace(1)* %out, align 4
+  ret void
+}
+
 
 declare float @llvm.fabs.f32(float) #1
+declare float @llvm.sqrt.f32(float) #1
 
 attributes #0 = { nounwind "unsafe-fp-math"="false" }
 attributes #1 = { nounwind readnone }
 attributes #2 = { nounwind "unsafe-fp-math"="true" }
 
 !0 = !{float 2.500000e+00}

test/CodeGen/AMDGPU/rsq.ll

define void @rsqrt_fmul(float addrspace(1)* %out, float addrspace(1)* %in) {
   %c = load volatile float, float addrspace(1)* %gep.2
 
   %x = call float @llvm.sqrt.f32(float %a)
   %y = fmul float %x, %b
   %z = fdiv float %c, %y
   store float %z, float addrspace(1)* %out.gep
   ret void
 }
+
+; SI-LABEL: {{^}}neg_rsq_f32:
+; SI-SAFE: v_sqrt_f32_e32 [[SQRT:v[0-9]+]], v{{[0-9]+}}
+; SI-SAFE: v_rcp_f32_e64 [[RSQ:v[0-9]+]], -[[SQRT]]
+; SI-SAF: buffer_store_dword [[RSQ]]
+
+; SI-UNSAFE: v_rsq_f32_e32 [[RSQ:v[0-9]+]], v{{[0-9]+}}
+; SI-UNSAFE: v_xor_b32_e32 [[NEG_RSQ:v[0-9]+]], 0x80000000, [[RSQ]]
+; SI-UNSAFE: buffer_store_dword [[NEG_RSQ]]
+define void @neg_rsq_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
+  %val = load float, float addrspace(1)* %in, align 4
+  %sqrt = call float @llvm.sqrt.f32(float %val)
+  %div = fdiv float -1.0, %sqrt
+  store float %div, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}neg_rsq_f64:
+; SI-SAFE: v_sqrt_f64_e32
+; SI-SAFE: v_div_scale_f64
+
+; SI-UNSAFE: v_sqrt_f64_e32 [[SQRT:v\[[0-9]+:[0-9]+\]]], v{{\[[0-9]+:[0-9]+\]}}
+; SI-UNSAFE: v_rcp_f64_e64 [[RCP:v\[[0-9]+:[0-9]+\]]], -[[SQRT]]
+; SI-UNSAFE: buffer_store_dwordx2 [[RCP]]
+define void @neg_rsq_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
+  %val = load double, double addrspace(1)* %in, align 4
+  %sqrt = call double @llvm.sqrt.f64(double %val)
+  %div = fdiv double -1.0, %sqrt
+  store double %div, double addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}neg_rsq_neg_f32:
+; SI-SAFE: v_sqrt_f32_e64 [[SQRT:v[0-9]+]], -v{{[0-9]+}}
+; SI-SAFE: v_rcp_f32_e64 [[RSQ:v[0-9]+]], -[[SQRT]]
+; SI-SAF: buffer_store_dword [[RSQ]]
+
+; SI-UNSAFE: v_rsq_f32_e64 [[RSQ:v[0-9]+]], -v{{[0-9]+}}
+; SI-UNSAFE: v_xor_b32_e32 [[NEG_RSQ:v[0-9]+]], 0x80000000, [[RSQ]]
+; SI-UNSAFE: buffer_store_dword [[NEG_RSQ]]
+define void @neg_rsq_neg_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
+  %val = load float, float addrspace(1)* %in, align 4
+  %val.fneg = fsub float -0.0, %val
+  %sqrt = call float @llvm.sqrt.f32(float %val.fneg)
+  %div = fdiv float -1.0, %sqrt
+  store float %div, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}neg_rsq_neg_f64:
+; SI-SAFE: v_sqrt_f64_e64 v{{\[[0-9]+:[0-9]+\]}}, -v{{\[[0-9]+:[0-9]+\]}}
+; SI-SAFE: v_div_scale_f64
+
+; SI-UNSAFE: v_sqrt_f64_e64 [[SQRT:v\[[0-9]+:[0-9]+\]]], -v{{\[[0-9]+:[0-9]+\]}}
+; SI-UNSAFE: v_rcp_f64_e64 [[RCP:v\[[0-9]+:[0-9]+\]]], -[[SQRT]]
+; SI-UNSAFE: buffer_store_dwordx2 [[RCP]]
+define void @neg_rsq_neg_f64(double addrspace(1)* noalias %out, double addrspace(1)* noalias %in) nounwind {
+  %val = load double, double addrspace(1)* %in, align 4
+  %val.fneg = fsub double -0.0, %val
+  %sqrt = call double @llvm.sqrt.f64(double %val.fneg)
+  %div = fdiv double -1.0, %sqrt
+  store double %div, double addrspace(1)* %out, align 4
+  ret void
+}