From 4a54e3eed347ca2bcb840c52c553892101afd252 Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Thu, 31 Mar 2022 08:21:28 -0400
Subject: [PATCH] [x86] try to replace 0.0 in fcmp with negated operand

This inverts a fold recently added to IR with:
3491f2f4b033

We can put -bidirectional on the Alive2 examples to show that
the reverse transforms work:
https://alive2.llvm.org/ce/z/8iVQwB

The motivation for the IR change was to improve matching to
'fabs' in IR (see https://github.com/llvm/llvm-project/issues/38828 ),
but it regressed x86 codegen for 'not-quite-fabs' patterns like
(X > -X) ? X : -X.
Ie, when there is no fast-math (nsz), the cmp+select is not a proper
fabs operation, but it does map nicely to the unusual NAN semantics
of MINSS/MAXSS.

I drafted this as a target-independent fold, but it doesn't appear to
help any other targets and seems to cause regressions for SystemZ at
least.

Differential Revision: https://reviews.llvm.org/D122726
---
 llvm/lib/Target/X86/X86ISelLowering.cpp   | 10 ++++
 llvm/test/CodeGen/X86/fp-stack-compare.ll | 10 ++--
 llvm/test/CodeGen/X86/setcc-combine.ll    | 98 ++++++++++---------------------
 3 files changed, 45 insertions(+), 73 deletions(-)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 8521b0e..80fec0a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -51502,6 +51502,16 @@ static SDValue combineSetCC(SDNode *N, SelectionDAG &DAG,
       LHS.getValueType() == MVT::v4f32)
     return LowerVSETCC(SDValue(N, 0), Subtarget, DAG);
 
+  // X pred 0.0 --> X pred -X
+  // If the negation of X already exists, use it in the comparison. This removes
+  // the need to materialize 0.0 and allows matching to SSE's MIN/MAX
+  // instructions in patterns with a 'select' node.
+  if (isNullFPScalarOrVectorConst(RHS)) {
+    SDVTList FNegVT = DAG.getVTList(OpVT);
+    if (SDNode *FNeg = DAG.getNodeIfExists(ISD::FNEG, FNegVT, {LHS}))
+      return DAG.getSetCC(DL, VT, LHS, SDValue(FNeg, 0), CC);
+  }
+
   return SDValue();
 }
 
diff --git a/llvm/test/CodeGen/X86/fp-stack-compare.ll b/llvm/test/CodeGen/X86/fp-stack-compare.ll
index 83796b3..fabcbaf 100644
--- a/llvm/test/CodeGen/X86/fp-stack-compare.ll
+++ b/llvm/test/CodeGen/X86/fp-stack-compare.ll
@@ -8,13 +8,12 @@ define float @foo(float* %col) {
 ; BASE:       # %bb.0:
 ; BASE-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; BASE-NEXT:    flds (%eax)
-; BASE-NEXT:    fldz
-; BASE-NEXT:    fucomp %st(1)
+; BASE-NEXT:    fld %st(0)
+; BASE-NEXT:    fchs
+; BASE-NEXT:    fucom %st(1)
 ; BASE-NEXT:    fnstsw %ax
 ; BASE-NEXT:    # kill: def $ah killed $ah killed $ax
 ; BASE-NEXT:    sahf
-; BASE-NEXT:    fld %st(0)
-; BASE-NEXT:    fchs
 ; BASE-NEXT:    ja .LBB0_2
 ; BASE-NEXT:  # %bb.1:
 ; BASE-NEXT:    fstp %st(0)
@@ -28,10 +27,9 @@ define float @foo(float* %col) {
 ; CMOV:       # %bb.0:
 ; CMOV-NEXT:    movl {{[0-9]+}}(%esp), %eax
 ; CMOV-NEXT:    flds (%eax)
-; CMOV-NEXT:    fldz
-; CMOV-NEXT:    fucompi %st(1), %st
 ; CMOV-NEXT:    fld %st(0)
 ; CMOV-NEXT:    fchs
+; CMOV-NEXT:    fucomi %st(1), %st
 ; CMOV-NEXT:    fxch %st(1)
 ; CMOV-NEXT:    fcmovnbe %st(1), %st
 ; CMOV-NEXT:    fstp %st(1)
diff --git a/llvm/test/CodeGen/X86/setcc-combine.ll b/llvm/test/CodeGen/X86/setcc-combine.ll
index 33d58ca..44994ab 100644
--- a/llvm/test/CodeGen/X86/setcc-combine.ll
+++ b/llvm/test/CodeGen/X86/setcc-combine.ll
@@ -364,12 +364,7 @@ define float @olt(float %x) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movaps {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    xorps %xmm0, %xmm1
-; CHECK-NEXT:    xorps %xmm2, %xmm2
-; CHECK-NEXT:    movaps %xmm0, %xmm3
-; CHECK-NEXT:    cmpltss %xmm2, %xmm3
-; CHECK-NEXT:    andps %xmm3, %xmm0
-; CHECK-NEXT:    andnps %xmm1, %xmm3
-; CHECK-NEXT:    orps %xmm3, %xmm0
+; CHECK-NEXT:    minss %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %cmp = fcmp olt float %x, 0.0
   %neg = fneg float %x
@@ -382,11 +377,7 @@ define double @ogt(double %x) {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    movapd {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0]
 ; CHECK-NEXT:    xorpd %xmm0, %xmm1
-; CHECK-NEXT:    xorpd %xmm2, %xmm2
-; CHECK-NEXT:    cmpltsd %xmm0, %xmm2
-; CHECK-NEXT:    andpd %xmm2, %xmm0
-; CHECK-NEXT:    andnpd %xmm1, %xmm2
-; CHECK-NEXT:    orpd %xmm2, %xmm0
+; CHECK-NEXT:    maxsd %xmm1, %xmm0
 ; CHECK-NEXT:    retq
   %neg = fneg double %x
   %cmp = fcmp ogt double %x, 0.0
@@ -395,28 +386,13 @@ define double @ogt(double %x) {
 }
 
 define <4 x float> @olt_swap(<4 x float> %x) {
-; SSE2-LABEL: olt_swap:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    xorps %xmm1, %xmm1
-; SSE2-NEXT:    movaps %xmm0, %xmm2
-; SSE2-NEXT:    cmpltps %xmm1, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm1
-; SSE2-NEXT:    andnps %xmm0, %xmm1
-; SSE2-NEXT:    xorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    andps %xmm2, %xmm0
-; SSE2-NEXT:    orps %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE41-LABEL: olt_swap:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movaps %xmm0, %xmm1
-; SSE41-NEXT:    xorps %xmm2, %xmm2
-; SSE41-NEXT:    cmpltps %xmm2, %xmm0
-; SSE41-NEXT:    movaps {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
-; SSE41-NEXT:    xorps %xmm1, %xmm2
-; SSE41-NEXT:    blendvps %xmm0, %xmm2, %xmm1
-; SSE41-NEXT:    movaps %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; CHECK-LABEL: olt_swap:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movaps {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; CHECK-NEXT:    xorps %xmm0, %xmm1
+; CHECK-NEXT:    maxps %xmm0, %xmm1
+; CHECK-NEXT:    movaps %xmm1, %xmm0
+; CHECK-NEXT:    retq
   %cmp = fcmp olt <4 x float> %x, zeroinitializer
   %neg = fneg <4 x float> %x
   %r = select <4 x i1> %cmp, <4 x float> %neg, <4 x float> %x
@@ -424,28 +400,13 @@ define <4 x float> @olt_swap(<4 x float> %x) {
 }
 
 define <2 x double> @ogt_swap(<2 x double> %x) {
-; SSE2-LABEL: ogt_swap:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movapd {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0]
-; SSE2-NEXT:    xorpd %xmm0, %xmm2
-; SSE2-NEXT:    xorpd %xmm1, %xmm1
-; SSE2-NEXT:    cmpltpd %xmm0, %xmm1
-; SSE2-NEXT:    andpd %xmm1, %xmm2
-; SSE2-NEXT:    andnpd %xmm0, %xmm1
-; SSE2-NEXT:    orpd %xmm2, %xmm1
-; SSE2-NEXT:    movapd %xmm1, %xmm0
-; SSE2-NEXT:    retq
-;
-; SSE41-LABEL: ogt_swap:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movapd %xmm0, %xmm1
-; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0]
-; SSE41-NEXT:    xorpd %xmm0, %xmm2
-; SSE41-NEXT:    xorpd %xmm0, %xmm0
-; SSE41-NEXT:    cmpltpd %xmm1, %xmm0
-; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
-; SSE41-NEXT:    movapd %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; CHECK-LABEL: ogt_swap:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movapd {{.*#+}} xmm1 = [-0.0E+0,-0.0E+0]
+; CHECK-NEXT:    xorpd %xmm0, %xmm1
+; CHECK-NEXT:    minpd %xmm0, %xmm1
+; CHECK-NEXT:    movapd %xmm1, %xmm0
+; CHECK-NEXT:    retq
   %neg = fneg <2 x double> %x
   %cmp = fcmp ogt <2 x double> %x, zeroinitializer
   %r = select <2 x i1> %cmp, <2 x double> %neg, <2 x double> %x
@@ -455,23 +416,22 @@ define <2 x double> @ogt_swap(<2 x double> %x) {
 define <4 x float> @ole(<4 x float> %x) {
 ; SSE2-LABEL: ole:
 ; SSE2:       # %bb.0:
-; SSE2-NEXT:    xorps %xmm1, %xmm1
-; SSE2-NEXT:    movaps %xmm0, %xmm2
-; SSE2-NEXT:    cmpleps %xmm1, %xmm2
-; SSE2-NEXT:    movaps %xmm2, %xmm1
+; SSE2-NEXT:    movaps {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
+; SSE2-NEXT:    xorps %xmm0, %xmm2
+; SSE2-NEXT:    movaps %xmm0, %xmm1
+; SSE2-NEXT:    cmpleps %xmm2, %xmm1
+; SSE2-NEXT:    andps %xmm1, %xmm2
 ; SSE2-NEXT:    andnps %xmm0, %xmm1
-; SSE2-NEXT:    xorps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
-; SSE2-NEXT:    andps %xmm2, %xmm0
-; SSE2-NEXT:    orps %xmm1, %xmm0
+; SSE2-NEXT:    orps %xmm2, %xmm1
+; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: ole:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movaps %xmm0, %xmm1
-; SSE41-NEXT:    xorps %xmm2, %xmm2
-; SSE41-NEXT:    cmpleps %xmm2, %xmm0
 ; SSE41-NEXT:    movaps {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0,-0.0E+0,-0.0E+0]
-; SSE41-NEXT:    xorps %xmm1, %xmm2
+; SSE41-NEXT:    xorps %xmm0, %xmm2
+; SSE41-NEXT:    cmpleps %xmm2, %xmm0
 ; SSE41-NEXT:    blendvps %xmm0, %xmm2, %xmm1
 ; SSE41-NEXT:    movaps %xmm1, %xmm0
 ; SSE41-NEXT:    retq
@@ -486,7 +446,7 @@ define <2 x double> @oge(<2 x double> %x) {
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movapd {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0]
 ; SSE2-NEXT:    xorpd %xmm0, %xmm2
-; SSE2-NEXT:    xorpd %xmm1, %xmm1
+; SSE2-NEXT:    movapd %xmm2, %xmm1
 ; SSE2-NEXT:    cmplepd %xmm0, %xmm1
 ; SSE2-NEXT:    andpd %xmm1, %xmm2
 ; SSE2-NEXT:    andnpd %xmm0, %xmm1
@@ -499,7 +459,7 @@ define <2 x double> @oge(<2 x double> %x) {
 ; SSE41-NEXT:    movapd %xmm0, %xmm1
 ; SSE41-NEXT:    movapd {{.*#+}} xmm2 = [-0.0E+0,-0.0E+0]
 ; SSE41-NEXT:    xorpd %xmm0, %xmm2
-; SSE41-NEXT:    xorpd %xmm0, %xmm0
+; SSE41-NEXT:    movapd %xmm2, %xmm0
 ; SSE41-NEXT:    cmplepd %xmm1, %xmm0
 ; SSE41-NEXT:    blendvpd %xmm0, %xmm2, %xmm1
 ; SSE41-NEXT:    movapd %xmm1, %xmm0
@@ -510,6 +470,8 @@ define <2 x double> @oge(<2 x double> %x) {
   ret <2 x double> %r
 }
 
+; negative test - don't create an fneg to replace 0.0 operand
+
 define double @ogt_no_fneg(double %x, double %y) {
 ; CHECK-LABEL: ogt_no_fneg:
 ; CHECK:       # %bb.0:
@@ -524,6 +486,8 @@ define double @ogt_no_fneg(double %x, double %y) {
   ret double %r
 }
 
+; negative test - can't change the setcc for non-zero constant
+
 define double @ogt_no_zero(double %x) {
 ; CHECK-LABEL: ogt_no_zero:
 ; CHECK:       # %bb.0:
-- 
2.7.4