AMDGPU: Address todo for handling 1/(2 pi)

llvm-svn: 339814

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

@@ -3449,9 +3449,27 @@ SDValue AMDGPUTargetLowering::performSelectCombine(SDNode *N,
   return performCtlz_CttzCombine(SDLoc(N), Cond, True, False, DCI);
 }
 
-static bool isConstantFPZero(SDValue N) {
-  if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N))
-    return C->isZero() && !C->isNegative();
+static bool isInv2Pi(const APFloat &APF) {
+  static const APFloat KF16(APFloat::IEEEhalf(), APInt(16, 0x3118));
+  static const APFloat KF32(APFloat::IEEEsingle(), APInt(32, 0x3e22f983));
+  static const APFloat KF64(APFloat::IEEEdouble(), APInt(64, 0x3fc45f306dc9c882));
+
+  return APF.bitwiseIsEqual(KF16) ||
+         APF.bitwiseIsEqual(KF32) ||
+         APF.bitwiseIsEqual(KF64);
+}
+
+// 0 and 1.0 / (0.5 * pi) do not have inline immmediates, so there is an
+// additional cost to negate them.
+bool AMDGPUTargetLowering::isConstantCostlierToNegate(SDValue N) const {
+  if (const ConstantFPSDNode *C = isConstOrConstSplatFP(N)) {
+    if (C->isZero() && !C->isNegative())
+      return true;
+
+    if (Subtarget->hasInv2PiInlineImm() && isInv2Pi(C->getValueAPF()))
+      return true;
+  }
+
   return false;
 }
 
@@ -3577,9 +3595,8 @@ SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
     SDValue RHS = N0.getOperand(1);
 
     // 0 doesn't have a negated inline immediate.
-    // TODO: Shouldn't fold 1/2pi either, and should be generalized to other
-    // operations.
-    if (isConstantFPZero(RHS))
+    // TODO: This constant check should be generalized to other operations.
+    if (isConstantCostlierToNegate(RHS))
       return SDValue();
 
     SDValue NegLHS = DAG.getNode(ISD::FNEG, SL, VT, LHS);

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h

@@ -95,6 +95,8 @@ protected:
   SDValue performCtlz_CttzCombine(const SDLoc &SL, SDValue Cond, SDValue LHS,
                              SDValue RHS, DAGCombinerInfo &DCI) const;
   SDValue performSelectCombine(SDNode *N, DAGCombinerInfo &DCI) const;
+
+  bool isConstantCostlierToNegate(SDValue N) const;
   SDValue performFNegCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performFAbsCombine(SDNode *N, DAGCombinerInfo &DCI) const;
   SDValue performRcpCombine(SDNode *N, DAGCombinerInfo &DCI) const;

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp

@@ -136,6 +136,7 @@ AMDGPUSubtarget::AMDGPUSubtarget(const Triple &TT,
   HasVOP3PInsts(false),
   HasMulI24(true),
   HasMulU24(true),
+  HasInv2PiInlineImm(false),
   HasFminFmaxLegacy(true),
   EnablePromoteAlloca(false),
   LocalMemorySize(0),
@@ -190,7 +191,6 @@ GCNSubtarget::GCNSubtarget(const Triple &TT, StringRef GPU, StringRef FS,
     HasVGPRIndexMode(false),
     HasScalarStores(false),
     HasScalarAtomics(false),
-    HasInv2PiInlineImm(false),
     HasSDWAOmod(false),
     HasSDWAScalar(false),
     HasSDWASdst(false),

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

@@ -72,6 +72,7 @@ protected:
   bool HasVOP3PInsts;
   bool HasMulI24;
   bool HasMulU24;
+  bool HasInv2PiInlineImm;
   bool HasFminFmaxLegacy;
   bool EnablePromoteAlloca;
   int LocalMemorySize;
@@ -170,6 +171,10 @@ public:
     return HasMulU24;
   }
 
+  bool hasInv2PiInlineImm() const {
+    return HasInv2PiInlineImm;
+  }
+
   bool hasFminFmaxLegacy() const {
     return HasFminFmaxLegacy;
   }
@@ -347,7 +352,6 @@ protected:
   bool HasVGPRIndexMode;
   bool HasScalarStores;
   bool HasScalarAtomics;
-  bool HasInv2PiInlineImm;
   bool HasSDWAOmod;
   bool HasSDWAScalar;
   bool HasSDWASdst;
@@ -782,9 +786,6 @@ public:
     return HasScalarAtomics;
   }
 
-  bool hasInv2PiInlineImm() const {
-    return HasInv2PiInlineImm;
-  }
 
   bool hasDPP() const {
     return HasDPP;

llvm/test/CodeGen/AMDGPU/fneg-combines.si.ll

@@ -0,0 +1,28 @@
+; RUN: llc -march=amdgcn -mcpu=tahiti -start-after=sink -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GCN-SAFE -check-prefix=SI -check-prefix=FUNC %s
+; RUN: llc -enable-no-signed-zeros-fp-math -march=amdgcn -mcpu=tahiti -start-after=sink -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=GCN-NSZ -check-prefix=SI -check-prefix=FUNC %s
+
+; --------------------------------------------------------------------------------
+; rcp_legacy tests
+; --------------------------------------------------------------------------------
+
+; GCN-LABEL: {{^}}v_fneg_rcp_legacy_f32:
+; GCN: {{buffer|flat}}_load_dword [[A:v[0-9]+]]
+; GCN: v_rcp_legacy_f32_e64 [[RESULT:v[0-9]+]], -[[A]]
+; GCN: {{buffer|flat}}_store_dword [[RESULT]]
+define amdgpu_kernel void @v_fneg_rcp_legacy_f32(float 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 float, float addrspace(1)* %out, i64 %tid.ext
+  %a = load volatile float, float addrspace(1)* %a.gep
+  %rcp = call float @llvm.amdgcn.rcp.legacy(float %a)
+  %fneg = fsub float -0.000000e+00, %rcp
+  store float %fneg, float addrspace(1)* %out.gep
+  ret void
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x() #1
+declare float @llvm.amdgcn.rcp.legacy(float) #1
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }