[FPEnv] Don't transform FSUB(-0,X)->FNEG(X) in SelectionDAGBuilder.

This patch stops unconditionally transforming FSUB(-0,X) into an FNEG(X) while building the DAG. There is also one small change to handle the new FSUB(-0,X) similarly to FNEG(X) in the AMDGPU backend. Differential Revision: https://reviews.llvm.org/D84056
llvm · Aug 3, 2020 · 31c7a2f · 31c7a2f
1 parent 10851f9
commit 31c7a2f
Show file tree

Hide file tree

Showing 12 changed files with 155 additions and 162 deletions.
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -3005,20 +3005,6 @@ void SelectionDAGBuilder::visitUnreachable(const UnreachableInst &I) {
   DAG.setRoot(DAG.getNode(ISD::TRAP, getCurSDLoc(), MVT::Other, DAG.getRoot()));
 }
 
-void SelectionDAGBuilder::visitFSub(const User &I) {
-  // -0.0 - X --> fneg
-  Type *Ty = I.getType();
-  if (isa<Constant>(I.getOperand(0)) &&
-      I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
-    SDValue Op2 = getValue(I.getOperand(1));
-    setValue(&I, DAG.getNode(ISD::FNEG, getCurSDLoc(),
-                             Op2.getValueType(), Op2));
-    return;
-  }
-
-  visitBinary(I, ISD::FSUB);
-}
-
 void SelectionDAGBuilder::visitUnary(const User &I, unsigned Opcode) {
   SDNodeFlags Flags;
 

diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h b/llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h
@@ -692,7 +692,7 @@ class SelectionDAGBuilder {
   void visitAdd(const User &I)  { visitBinary(I, ISD::ADD); }
   void visitFAdd(const User &I) { visitBinary(I, ISD::FADD); }
   void visitSub(const User &I)  { visitBinary(I, ISD::SUB); }
-  void visitFSub(const User &I);
+  void visitFSub(const User &I) { visitBinary(I, ISD::FSUB); }
   void visitMul(const User &I)  { visitBinary(I, ISD::MUL); }
   void visitFMul(const User &I) { visitBinary(I, ISD::FMUL); }
   void visitURem(const User &I) { visitBinary(I, ISD::UREM); }

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -3795,8 +3795,15 @@ SDValue AMDGPUTargetLowering::performFNegCombine(SDNode *N,
     SDValue Res = DAG.getNode(AMDGPUISD::FMED3, SL, VT, Ops, N0->getFlags());
     if (Res.getOpcode() != AMDGPUISD::FMED3)
       return SDValue(); // Op got folded away.
-    if (!N0.hasOneUse())
-      DAG.ReplaceAllUsesWith(N0, DAG.getNode(ISD::FNEG, SL, VT, Res));
+
+    if (!N0.hasOneUse()) {
+      SDValue Neg = DAG.getNode(ISD::FNEG, SL, VT, Res);
+      DAG.ReplaceAllUsesWith(N0, Neg);
+
+      for (SDNode *U : Neg->uses())
+        DCI.AddToWorklist(U);
+    }
+
     return Res;
   }
   case ISD::FP_EXTEND:

diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
@@ -311,7 +311,7 @@ define amdgpu_kernel void @test_no_fold_canonicalize_fneg_value_f32(float addrsp
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
   %load = load float, float addrspace(1)* %gep, align 4
-  %v = fsub float -0.0, %load
+  %v = fneg float %load
   %canonicalized = tail call float @llvm.canonicalize.f32(float %v)
   store float %canonicalized, float addrspace(1)* %gep, align 4
   ret void
@@ -327,7 +327,7 @@ define amdgpu_kernel void @test_fold_canonicalize_fneg_value_f32(float addrspace
   %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %id
   %load = load float, float addrspace(1)* %gep, align 4
   %v0 = fadd float %load, 0.0
-  %v = fsub float -0.0, %v0
+  %v = fneg float %v0
   %canonicalized = tail call float @llvm.canonicalize.f32(float %v)
   store float %canonicalized, float addrspace(1)* %gep, align 4
   ret void

diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize.f16.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize.f16.ll
@@ -77,7 +77,7 @@ define amdgpu_kernel void @v_test_canonicalize_fabs_var_f16(half addrspace(1)* %
 define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f16(half addrspace(1)* %out) #1 {
   %val = load half, half addrspace(1)* %out
   %val.fabs = call half @llvm.fabs.f16(half %val)
-  %val.fabs.fneg = fsub half -0.0, %val.fabs
+  %val.fabs.fneg = fneg half %val.fabs
   %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs.fneg)
   store half %canonicalized, half addrspace(1)* %out
   ret void
@@ -91,7 +91,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f16(half addrspace(
 ; CI: v_mul_f32_e32 {{v[0-9]+}}, 1.0, {{v[0-9]+}}
 define amdgpu_kernel void @v_test_canonicalize_fneg_var_f16(half addrspace(1)* %out) #1 {
   %val = load half, half addrspace(1)* %out
-  %val.fneg = fsub half -0.0, %val
+  %val.fneg = fneg half %val
   %canonicalized = call half @llvm.canonicalize.f16(half %val.fneg)
   store half %canonicalized, half addrspace(1)* %out
   ret void
@@ -103,7 +103,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_var_f16(half addrspace(1)* %
 ; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_f16(half addrspace(1)* %out) #2 {
   %val = load half, half addrspace(1)* %out
-  %val.fneg = fsub half -0.0, %val
+  %val.fneg = fneg half %val
   %canonicalized = call half @llvm.canonicalize.f16(half %val.fneg)
   store half %canonicalized, half addrspace(1)* %out
   ret void
@@ -120,7 +120,7 @@ define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_var_f16(half ad
 define amdgpu_kernel void @v_test_no_denormals_canonicalize_fneg_fabs_var_f16(half addrspace(1)* %out) #2 {
   %val = load half, half addrspace(1)* %out
   %val.fabs = call half @llvm.fabs.f16(half %val)
-  %val.fabs.fneg = fsub half -0.0, %val.fabs
+  %val.fabs.fneg = fneg half %val.fabs
   %canonicalized = call half @llvm.canonicalize.f16(half %val.fabs.fneg)
   store half %canonicalized, half addrspace(1)* %out
   ret void
@@ -323,7 +323,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_v2f16(<2 x half> ad
   %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
   %val = load <2 x half>, <2 x half> addrspace(1)* %gep
   %val.fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %val)
-  %val.fabs.fneg = fsub <2 x half> <half -0.0, half -0.0>, %val.fabs
+  %val.fabs.fneg = fneg <2 x half> %val.fabs
   %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %val.fabs.fneg)
   store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
   ret void
@@ -340,7 +340,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_var_v2f16(<2 x half> addrspa
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr <2 x half>, <2 x half> addrspace(1)* %out, i32 %tid
   %val = load <2 x half>, <2 x half> addrspace(1)* %gep
-  %fneg.val = fsub <2 x half> <half -0.0, half -0.0>, %val
+  %fneg.val = fneg <2 x half> %val
   %canonicalized = call <2 x half> @llvm.canonicalize.v2f16(<2 x half> %fneg.val)
   store <2 x half> %canonicalized, <2 x half> addrspace(1)* %out
   ret void

diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize.ll
@@ -57,7 +57,7 @@ define amdgpu_kernel void @v_test_canonicalize_fabs_var_f32(float addrspace(1)*
 define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f32(float addrspace(1)* %out) #1 {
   %val = load float, float addrspace(1)* %out
   %val.fabs = call float @llvm.fabs.f32(float %val)
-  %val.fabs.fneg = fsub float -0.0, %val.fabs
+  %val.fabs.fneg = fneg float %val.fabs
   %canonicalized = call float @llvm.canonicalize.f32(float %val.fabs.fneg)
   store float %canonicalized, float addrspace(1)* %out
   ret void
@@ -69,7 +69,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f32(float addrspace
 ; GCN: {{flat|global}}_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_var_f32(float addrspace(1)* %out) #1 {
   %val = load float, float addrspace(1)* %out
-  %val.fneg = fsub float -0.0, %val
+  %val.fneg = fneg float %val
   %canonicalized = call float @llvm.canonicalize.f32(float %val.fneg)
   store float %canonicalized, float addrspace(1)* %out
   ret void
@@ -264,7 +264,7 @@ define amdgpu_kernel void @v_test_canonicalize_fabs_var_f64(double addrspace(1)*
 define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f64(double addrspace(1)* %out) #1 {
   %val = load double, double addrspace(1)* %out
   %val.fabs = call double @llvm.fabs.f64(double %val)
-  %val.fabs.fneg = fsub double -0.0, %val.fabs
+  %val.fabs.fneg = fneg double %val.fabs
   %canonicalized = call double @llvm.canonicalize.f64(double %val.fabs.fneg)
   store double %canonicalized, double addrspace(1)* %out
   ret void
@@ -275,7 +275,7 @@ define amdgpu_kernel void @v_test_canonicalize_fneg_fabs_var_f64(double addrspac
 ; GCN: {{flat|global}}_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[REG]]
 define amdgpu_kernel void @v_test_canonicalize_fneg_var_f64(double addrspace(1)* %out) #1 {
   %val = load double, double addrspace(1)* %out
-  %val.fneg = fsub double -0.0, %val
+  %val.fneg = fneg double %val
   %canonicalized = call double @llvm.canonicalize.f64(double %val.fneg)
   store double %canonicalized, double addrspace(1)* %out
   ret void

diff --git a/llvm/test/CodeGen/AMDGPU/fdiv32-to-rcp-folding.ll b/llvm/test/CodeGen/AMDGPU/fdiv32-to-rcp-folding.ll
@@ -311,7 +311,7 @@ define amdgpu_kernel void @div_v4_c_by_x_25ulp(<4 x float> addrspace(1)* %arg) {
 ; GCN:            global_store_dwordx4
 define amdgpu_kernel void @div_v4_c_by_minus_x_25ulp(<4 x float> addrspace(1)* %arg) {
   %load = load <4 x float>, <4 x float> addrspace(1)* %arg, align 16
-  %neg = fsub <4 x float> <float -0.000000e+00, float -0.000000e+00, float -0.000000e+00, float -0.000000e+00>, %load
+  %neg = fneg <4 x float> %load
   %div = fdiv <4 x float> <float 2.000000e+00, float 1.000000e+00, float -1.000000e+00, float -2.000000e+00>, %neg, !fpmath !0
   store <4 x float> %div, <4 x float> addrspace(1)* %arg, align 16
   ret void

diff --git a/llvm/test/CodeGen/AMDGPU/fmuladd.f16.ll b/llvm/test/CodeGen/AMDGPU/fmuladd.f16.ll
@@ -256,7 +256,7 @@ define amdgpu_kernel void @fmuladd_neg_2.0_neg_a_b_f16(half addrspace(1)* %out,
   %r1 = load volatile half, half addrspace(1)* %gep.0
   %r2 = load volatile half, half addrspace(1)* %gep.1
 
-  %r1.fneg = fsub half -0.000000e+00, %r1
+  %r1.fneg = fneg half %r1
 
   %r3 = tail call half @llvm.fmuladd.f16(half -2.0, half %r1.fneg, half %r2)
   store half %r3, half addrspace(1)* %gep.out
@@ -287,7 +287,7 @@ define amdgpu_kernel void @fmuladd_2.0_neg_a_b_f16(half addrspace(1)* %out, half
   %r1 = load volatile half, half addrspace(1)* %gep.0
   %r2 = load volatile half, half addrspace(1)* %gep.1
 
-  %r1.fneg = fsub half -0.000000e+00, %r1
+  %r1.fneg = fneg half %r1
 
   %r3 = tail call half @llvm.fmuladd.f16(half 2.0, half %r1.fneg, half %r2)
   store half %r3, half addrspace(1)* %gep.out
@@ -312,7 +312,7 @@ define amdgpu_kernel void @fmuladd_2.0_a_neg_b_f16(half addrspace(1)* %out, half
   %r1 = load volatile half, half addrspace(1)* %gep.0
   %r2 = load volatile half, half addrspace(1)* %gep.1
 
-  %r2.fneg = fsub half -0.000000e+00, %r2
+  %r2.fneg = fneg half %r2
 
   %r3 = tail call half @llvm.fmuladd.f16(half 2.0, half %r1, half %r2.fneg)
   store half %r3, half addrspace(1)* %gep.out
@@ -494,8 +494,8 @@ define amdgpu_kernel void @neg_neg_mad_f16(half addrspace(1)* noalias nocapture
   %a = load volatile half, half addrspace(1)* %gep0, align 2
   %b = load volatile half, half addrspace(1)* %gep1, align 2
   %c = load volatile half, half addrspace(1)* %gep2, align 2
-  %nega = fsub half -0.000000e+00, %a
-  %negb = fsub half -0.000000e+00, %b
+  %nega = fneg half %a
+  %negb = fneg half %b
   %mul = fmul half %nega, %negb
   %sub = fadd half %mul, %c
   store half %sub, half addrspace(1)* %outgep, align 2

diff --git a/llvm/test/CodeGen/AMDGPU/fmuladd.f32.ll b/llvm/test/CodeGen/AMDGPU/fmuladd.f32.ll
@@ -274,7 +274,7 @@ define amdgpu_kernel void @fmuladd_neg_2.0_neg_a_b_f32(float addrspace(1)* %out,
   %r1 = load volatile float, float addrspace(1)* %gep.0
   %r2 = load volatile float, float addrspace(1)* %gep.1
 
-  %r1.fneg = fsub float -0.000000e+00, %r1
+  %r1.fneg = fneg float %r1
 
   %r3 = tail call float @llvm.fmuladd.f32(float -2.0, float %r1.fneg, float %r2)
   store float %r3, float addrspace(1)* %gep.out
@@ -307,7 +307,7 @@ define amdgpu_kernel void @fmuladd_2.0_neg_a_b_f32(float addrspace(1)* %out, flo
   %r1 = load volatile float, float addrspace(1)* %gep.0
   %r2 = load volatile float, float addrspace(1)* %gep.1
 
-  %r1.fneg = fsub float -0.000000e+00, %r1
+  %r1.fneg = fneg float %r1
 
   %r3 = tail call float @llvm.fmuladd.f32(float 2.0, float %r1.fneg, float %r2)
   store float %r3, float addrspace(1)* %gep.out
@@ -339,7 +339,7 @@ define amdgpu_kernel void @fmuladd_2.0_a_neg_b_f32(float addrspace(1)* %out, flo
   %r1 = load volatile float, float addrspace(1)* %gep.0
   %r2 = load volatile float, float addrspace(1)* %gep.1
 
-  %r2.fneg = fsub float -0.000000e+00, %r2
+  %r2.fneg = fneg float %r2
 
   %r3 = tail call float @llvm.fmuladd.f32(float 2.0, float %r1, float %r2.fneg)
   store float %r3, float addrspace(1)* %gep.out
@@ -517,8 +517,8 @@ define amdgpu_kernel void @neg_neg_mad_f32(float addrspace(1)* noalias nocapture
   %a = load volatile float, float addrspace(1)* %gep0, align 4
   %b = load volatile float, float addrspace(1)* %gep1, align 4
   %c = load volatile float, float addrspace(1)* %gep2, align 4
-  %nega = fsub float -0.000000e+00, %a
-  %negb = fsub float -0.000000e+00, %b
+  %nega = fneg float %a
+  %negb = fneg float %b
   %mul = fmul float %nega, %negb
   %sub = fadd float %mul, %c
   store float %sub, float addrspace(1)* %outgep, align 4