AMDGPU: Implement hasBitPreservingFPLogic

[rampitec]

LGTM

[arsenm]

r315754

lib/Target/AMDGPU/SIISelLowering.h

unsigned getRegisterByName(const char* RegName, EVT VT,
                              SelectionDAG &DAG) const override;
 
   MachineBasicBlock *splitKillBlock(MachineInstr &MI,
                                     MachineBasicBlock *BB) const;
 
   MachineBasicBlock *
   EmitInstrWithCustomInserter(MachineInstr &MI,
                               MachineBasicBlock *BB) const override;
+
+  bool hasBitPreservingFPLogic(EVT VT) const override;
   bool enableAggressiveFMAFusion(EVT VT) const override;
   EVT getSetCCResultType(const DataLayout &DL, LLVMContext &Context,
                          EVT VT) const override;
   MVT getScalarShiftAmountTy(const DataLayout &, EVT) const override;
   bool isFMAFasterThanFMulAndFAdd(EVT VT) const override;
   SDValue LowerOperation(SDValue Op, SelectionDAG &DAG) const override;
   void ReplaceNodeResults(SDNode *N, SmallVectorImpl<SDValue> &Results,
                           SelectionDAG &DAG) const override;

lib/Target/AMDGPU/SIISelLowering.cpp

case AMDGPU::SI_TCRETURN_ISEL: {
     MI.eraseFromParent();
     return BB;
   }
   default:
     return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
   }
 }
 
+bool SITargetLowering::hasBitPreservingFPLogic(EVT VT) const {
+  return isTypeLegal(VT.getScalarType());
+}
+
 bool SITargetLowering::enableAggressiveFMAFusion(EVT VT) const {
   // This currently forces unfolding various combinations of fsub into fma with
   // free fneg'd operands. As long as we have fast FMA (controlled by
   // isFMAFasterThanFMulAndFAdd), we should perform these.
 
   // When fma is quarter rate, for f64 where add / sub are at best half rate,
   // most of these combines appear to be cycle neutral but save on instruction
   // count / code size.

test/CodeGen/AMDGPU/fabs.ll

 ; GCN: v_mul_f32_e64 v{{[0-9]+}}, |[[ABS_VALUE]]|, v{{[0-9]+}}
 define amdgpu_kernel void @fabs_fn_fold(float addrspace(1)* %out, float %in0, float %in1) {
   %fabs = call float @fabs(float %in0)
   %fmul = fmul float %fabs, %in1
   store float %fmul, float addrspace(1)* %out
   ret void
 }
 
-; GCN-LABEL: {{^}}fabs_fold:
+; FUNC-LABEL: {{^}}fabs_fold:
 ; SI: s_load_dword [[ABS_VALUE:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0xb
 ; VI: s_load_dword [[ABS_VALUE:s[0-9]+]], s[{{[0-9]+:[0-9]+}}], 0x2c
 ; GCN-NOT: and
 ; GCN: v_mul_f32_e64 v{{[0-9]+}}, |[[ABS_VALUE]]|, v{{[0-9]+}}
 define amdgpu_kernel void @fabs_fold(float addrspace(1)* %out, float %in0, float %in1) {
   %fabs = call float @llvm.fabs.f32(float %in0)
   %fmul = fmul float %fabs, %in1
   store float %fmul, float addrspace(1)* %out
   ret void
 }
 
+; Make sure we turn some integer operations back into fabs
+; FUNC-LABEL: {{^}}bitpreserve_fabs_f32:
+; GCN: v_add_f32_e64 v{{[0-9]+}}, |s{{[0-9]+}}|, 1.0
+define amdgpu_kernel void @bitpreserve_fabs_f32(float addrspace(1)* %out, float %in) {
+  %in.bc = bitcast float %in to i32
+  %int.abs = and i32 %in.bc, 2147483647
+  %bc = bitcast i32 %int.abs to float
+  %fadd = fadd float %bc, 1.0
+  store float %fadd, float addrspace(1)* %out
+  ret void
+}
+
 declare float @fabs(float) readnone
 declare float @llvm.fabs.f32(float) readnone
 declare <2 x float> @llvm.fabs.v2f32(<2 x float>) readnone
 declare <4 x float> @llvm.fabs.v4f32(<4 x float>) readnone

test/CodeGen/AMDGPU/fneg.ll

 ; GCN-NOT: xor
 ; GCN: v_mul_f32_e64 v{{[0-9]+}}, -[[NEG_VALUE]], [[NEG_VALUE]]
 define amdgpu_kernel void @fneg_fold_f32(float addrspace(1)* %out, float %in) {
   %fsub = fsub float -0.0, %in
   %fmul = fmul float %fsub, %in
   store float %fmul, float addrspace(1)* %out
   ret void
 }
+
+; Make sure we turn some integer operations back into fabs
+; FUNC-LABEL: {{^}}bitpreserve_fneg_f32:
+; GCN: v_mul_f32_e64 v{{[0-9]+}}, s{{[0-9]+}}, -4.0
+define amdgpu_kernel void @bitpreserve_fneg_f32(float addrspace(1)* %out, float %in) {
+  %in.bc = bitcast float %in to i32
+  %int.abs = xor i32 %in.bc, 2147483648
+  %bc = bitcast i32 %int.abs to float
+  %fadd = fmul float %bc, 4.0
+  store float %fadd, float addrspace(1)* %out
+  ret void
+}