AMDGPU: Fold fneg into fp_extend

[arsenm]

Move into target

[arsenm]

Add hasOneUse check

[nhaehnle]

LGTM

[arsenm]

r291777

lib/Target/AMDGPU/AMDGPUISelLowering.cpp

case ISD::FMAD: {
     else
       RHS = RHS.getOperand(0);
 
     SDValue Res = DAG.getNode(Opc, SL, VT, LHS, MHS, RHS);
     if (!N0.hasOneUse())
       DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
     return Res;
   }
+  case ISD::FP_EXTEND: {
+    SDValue CvtSrc = N0.getOperand(0);
+    if (CvtSrc.getOpcode() == ISD::FNEG) {
+      // (fneg (fp_extend (fneg x))) -> (fp_extend x)
+      return DAG.getNode(ISD::FP_EXTEND, SL, VT, CvtSrc.getOperand(0));
+    }
+
+    if (!N0.hasOneUse())
+      return SDValue();
+
+    // (fneg (fp_extend x)) -> (fp_extend (fneg x))
+    SDValue Neg = DAG.getNode(ISD::FNEG, SL, CvtSrc.getValueType(), CvtSrc);
+    return DAG.getNode(ISD::FP_EXTEND, SL, VT, Neg);
+  }
   default:
     return SDValue();
   }
 }
 
 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;

test/CodeGen/AMDGPU/fneg-combines.ll

define void @v_fneg_fmad_multi_use_fmad_f32(float addrspace(1)* %out, float addrspace(1)* %a.ptr, float addrspace(1)* %b.ptr, float addrspace(1)* %c.ptr) #0 {
   %fma = call float @llvm.fmuladd.f32(float %a, float %b, float %c)
   %fneg = fsub float -0.000000e+00, %fma
   %use1 = fmul float %fma, 4.0
   store volatile float %fneg, float addrspace(1)* %out
   store volatile float %use1, float addrspace(1)* %out
   ret void
 }
 
+; --------------------------------------------------------------------------------
+; fp_extend tests
+; --------------------------------------------------------------------------------
+
+; GCN-LABEL: {{^}}v_fneg_fp_extend_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_cvt_f64_f32_e64 [[RESULT:v\[[0-9]+:[0-9]+\]]], -[[A]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+define void @v_fneg_fp_extend_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fpext = fpext float %a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store double %fneg, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_fp_extend_fneg_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_cvt_f64_f32_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+define void @v_fneg_fp_extend_fneg_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fneg.a = fsub float -0.000000e+00, %a
+  %fpext = fpext float %fneg.a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store double %fneg, double addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_fp_extend_store_use_fneg_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN-DAG: v_cvt_f64_f32_e32 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[A]]
+; GCN-DAG: v_xor_b32_e32 [[FNEG_A:v[0-9]+]], 0x80000000, [[A]]
+; GCN: buffer_store_dwordx2 [[RESULT]]
+; GCN: buffer_store_dword [[FNEG_A]]
+define void @v_fneg_fp_extend_store_use_fneg_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fneg.a = fsub float -0.000000e+00, %a
+  %fpext = fpext float %fneg.a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store volatile double %fneg, double addrspace(1)* %out.gep
+  store volatile float %fneg.a, float addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_multi_use_fp_extend_fneg_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN-DAG: v_cvt_f64_f32_e32 v{{\[}}[[CVT_LO:[0-9]+]]:[[CVT_HI:[0-9]+]]{{\]}}, [[A]]
+; GCN-DAG: v_xor_b32_e32 v[[FNEG_A:[0-9]+]], 0x80000000, v[[CVT_HI]]
+; GCN: buffer_store_dwordx2 v{{\[[0-9]+}}:[[FNEG_A]]{{\]}}
+; GCN: buffer_store_dwordx2 v{{\[}}[[CVT_LO]]:[[CVT_HI]]{{\]}}
+define void @v_fneg_multi_use_fp_extend_fneg_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fpext = fpext float %a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  store volatile double %fneg, double addrspace(1)* %out.gep
+  store volatile double %fpext, double addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_multi_foldable_use_fp_extend_fneg_f32_to_f64:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN-DAG: v_cvt_f64_f32_e32 v{{\[}}[[CVT_LO:[0-9]+]]:[[CVT_HI:[0-9]+]]{{\]}}, [[A]]
+; GCN-DAG: v_xor_b32_e32 v[[FNEG_A:[0-9]+]], 0x80000000, v[[CVT_HI]]
+; GCN-DAG: v_mul_f64 [[MUL:v\[[0-9]+:[0-9]+\]]], v{{\[}}[[CVT_LO]]:[[CVT_HI]]{{\]}}, 4.0
+; GCN: buffer_store_dwordx2 v{{\[[0-9]+}}:[[FNEG_A]]{{\]}}
+; GCN: buffer_store_dwordx2 [[MUL]]
+define void @v_fneg_multi_foldable_use_fp_extend_fneg_f32_to_f64(double addrspace(1)* %out, float addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds float, float addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %fpext = fpext float %a to double
+  %fneg = fsub double -0.000000e+00, %fpext
+  %mul = fmul double %fpext, 4.0
+  store volatile double %fneg, double addrspace(1)* %out.gep
+  store volatile double %mul, double addrspace(1)* %out.gep
+  ret void
+}
+
+; FIXME: Source modifiers not folded for f16->f32
+; GCN-LABEL: {{^}}v_fneg_multi_use_fp_extend_fneg_f16_to_f32:
+define void @v_fneg_multi_use_fp_extend_fneg_f16_to_f32(float addrspace(1)* %out, half addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds half, half addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile half, half addrspace(1)* %a.gep
+  %fpext = fpext half %a to float
+  %fneg = fsub float -0.000000e+00, %fpext
+  store volatile float %fneg, float addrspace(1)* %out.gep
+  store volatile float %fpext, float addrspace(1)* %out.gep
+  ret void
+}
+
+; GCN-LABEL: {{^}}v_fneg_multi_foldable_use_fp_extend_fneg_f16_to_f32:
+define void @v_fneg_multi_foldable_use_fp_extend_fneg_f16_to_f32(float addrspace(1)* %out, half addrspace(1)* %a.ptr) #0 {
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %tid.ext = sext i32 %tid to i64
+  %a.gep = getelementptr inbounds half, half addrspace(1)* %a.ptr, i64 %tid.ext
+  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile half, half addrspace(1)* %a.gep
+  %fpext = fpext half %a to float
+  %fneg = fsub float -0.000000e+00, %fpext
+  %mul = fmul float %fpext, 4.0
+  store volatile float %fneg, float addrspace(1)* %out.gep
+  store volatile float %mul, float addrspace(1)* %out.gep
+  ret void
+}
 
 declare i32 @llvm.amdgcn.workitem.id.x() #1
 declare float @llvm.fma.f32(float, float, float) #1
 declare float @llvm.fmuladd.f32(float, float, float) #1
 
 attributes #0 = { nounwind }
 attributes #1 = { nounwind readnone }