From 57a08b3343b2de8e0ab87ed0d823e07e28c7e878 Mon Sep 17 00:00:00 2001
From: Sanjay Patel <spatel@rotateright.com>
Date: Wed, 7 Nov 2018 16:15:01 +0000
Subject: [PATCH] [InstCombine] propagate FMF for fcmp+fabs folds

By morphing the instruction rather than deleting and creating a new one,
we retain fast-math-flags and potentially other metadata (profile info?).

llvm-svn: 346331
---
 .../InstCombine/InstCombineCompares.cpp       | 20 ++++++++++------
 llvm/test/Transforms/InstCombine/fcmp.ll      | 24 +++++++++----------
 2 files changed, 25 insertions(+), 19 deletions(-)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
index 7a56313000d8..2ba1174517ff 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -5291,6 +5291,12 @@ static Instruction *foldFabsWithFcmpZero(FCmpInst &I) {
       !match(I.getOperand(1), m_PosZeroFP()))
     return nullptr;
 
+  auto replacePredAndOp0 = [](FCmpInst *I, FCmpInst::Predicate P, Value *X) {
+    I->setPredicate(P);
+    I->setOperand(0, X);
+    return I;
+  };
+
   switch (I.getPredicate()) {
   case FCmpInst::FCMP_UGE:
   case FCmpInst::FCMP_OLT:
@@ -5300,29 +5306,29 @@ static Instruction *foldFabsWithFcmpZero(FCmpInst &I) {
 
   case FCmpInst::FCMP_OGT:
     // fabs(X) > 0.0 --> X != 0.0
-    return new FCmpInst(FCmpInst::FCMP_ONE, X, I.getOperand(1));
+    return replacePredAndOp0(&I, FCmpInst::FCMP_ONE, X);
 
   case FCmpInst::FCMP_UGT:
     // fabs(X) u> 0.0 --> X u!= 0.0
-    return new FCmpInst(FCmpInst::FCMP_UNE, X, I.getOperand(1));
+    return replacePredAndOp0(&I, FCmpInst::FCMP_UNE, X);
 
   case FCmpInst::FCMP_OLE:
     // fabs(X) <= 0.0 --> X == 0.0
-    return new FCmpInst(FCmpInst::FCMP_OEQ, X, I.getOperand(1));
+    return replacePredAndOp0(&I, FCmpInst::FCMP_OEQ, X);
 
   case FCmpInst::FCMP_ULE:
     // fabs(X) u<= 0.0 --> X u== 0.0
-    return new FCmpInst(FCmpInst::FCMP_UEQ, X, I.getOperand(1));
+    return replacePredAndOp0(&I, FCmpInst::FCMP_UEQ, X);
 
   case FCmpInst::FCMP_OGE:
     // fabs(X) >= 0.0 --> !isnan(X)
     assert(!I.hasNoNaNs() && "fcmp should have simplified");
-    return new FCmpInst(FCmpInst::FCMP_ORD, X, I.getOperand(1));
+    return replacePredAndOp0(&I, FCmpInst::FCMP_ORD, X);
 
   case FCmpInst::FCMP_ULT:
     // fabs(X) u< 0.0 --> isnan(X)
     assert(!I.hasNoNaNs() && "fcmp should have simplified");
-    return new FCmpInst(FCmpInst::FCMP_UNO, X, I.getOperand(1));
+    return replacePredAndOp0(&I, FCmpInst::FCMP_UNO, X);
 
   case FCmpInst::FCMP_OEQ:
   case FCmpInst::FCMP_UEQ:
@@ -5335,7 +5341,7 @@ static Instruction *foldFabsWithFcmpZero(FCmpInst &I) {
     // fabs(X) != 0.0 --> X != 0.0
     // isnan(fabs(X)) --> isnan(X)
     // !isnan(fabs(X) --> !isnan(X)
-    return new FCmpInst(I.getPredicate(), X, I.getOperand(1));
+    return replacePredAndOp0(&I, I.getPredicate(), X);
 
   default:
     return nullptr;
diff --git a/llvm/test/Transforms/InstCombine/fcmp.ll b/llvm/test/Transforms/InstCombine/fcmp.ll
index 15d9368f49ed..be7aedc7c60f 100644
--- a/llvm/test/Transforms/InstCombine/fcmp.ll
+++ b/llvm/test/Transforms/InstCombine/fcmp.ll
@@ -189,7 +189,7 @@ define i1 @fabs_olt(half %a) {
 
 define <2 x i1> @fabs_ole(<2 x float> %a) {
 ; CHECK-LABEL: @fabs_ole(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq <2 x float> [[A:%.*]], zeroinitializer
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ninf oeq <2 x float> [[A:%.*]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %call = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
@@ -199,7 +199,7 @@ define <2 x i1> @fabs_ole(<2 x float> %a) {
 
 define <2 x i1> @fabs_ule(<2 x float> %a) {
 ; CHECK-LABEL: @fabs_ule(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp ueq <2 x float> [[A:%.*]], zeroinitializer
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ninf arcp ueq <2 x float> [[A:%.*]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %call = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
@@ -209,7 +209,7 @@ define <2 x i1> @fabs_ule(<2 x float> %a) {
 
 define i1 @fabs_ogt(double %a) {
 ; CHECK-LABEL: @fabs_ogt(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp one double [[A:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp reassoc one double [[A:%.*]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call double @llvm.fabs.f64(double %a)
@@ -219,7 +219,7 @@ define i1 @fabs_ogt(double %a) {
 
 define i1 @fabs_ugt(double %a) {
 ; CHECK-LABEL: @fabs_ugt(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp une double [[A:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp reassoc ninf une double [[A:%.*]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call double @llvm.fabs.f64(double %a)
@@ -229,7 +229,7 @@ define i1 @fabs_ugt(double %a) {
 
 define i1 @fabs_oge(double %a) {
 ; CHECK-LABEL: @fabs_oge(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp ord double [[A:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp afn ord double [[A:%.*]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call double @llvm.fabs.f64(double %a)
@@ -239,7 +239,7 @@ define i1 @fabs_oge(double %a) {
 
 define i1 @fabs_ult(double %a) {
 ; CHECK-LABEL: @fabs_ult(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp uno double [[A:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp reassoc arcp uno double [[A:%.*]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call double @llvm.fabs.f64(double %a)
@@ -258,7 +258,7 @@ define <2 x i1> @fabs_ult_nnan(<2 x float> %a) {
 
 define i1 @fabs_une(half %a) {
 ; CHECK-LABEL: @fabs_une(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp une half [[A:%.*]], 0xH0000
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp ninf une half [[A:%.*]], 0xH0000
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call half @llvm.fabs.f16(half %a)
@@ -268,7 +268,7 @@ define i1 @fabs_une(half %a) {
 
 define i1 @fabs_oeq(double %a) {
 ; CHECK-LABEL: @fabs_oeq(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp oeq double [[A:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp reassoc ninf oeq double [[A:%.*]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call double @llvm.fabs.f64(double %a)
@@ -278,7 +278,7 @@ define i1 @fabs_oeq(double %a) {
 
 define i1 @fabs_one(double %a) {
 ; CHECK-LABEL: @fabs_one(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp one double [[A:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp fast one double [[A:%.*]], 0.000000e+00
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %call = call double @llvm.fabs.f64(double %a)
@@ -288,7 +288,7 @@ define i1 @fabs_one(double %a) {
 
 define <2 x i1> @fabs_ueq(<2 x float> %a) {
 ; CHECK-LABEL: @fabs_ueq(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp ueq <2 x float> [[A:%.*]], zeroinitializer
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp arcp ueq <2 x float> [[A:%.*]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %call = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
@@ -298,7 +298,7 @@ define <2 x i1> @fabs_ueq(<2 x float> %a) {
 
 define <2 x i1> @fabs_ord(<2 x float> %a) {
 ; CHECK-LABEL: @fabs_ord(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp ord <2 x float> [[A:%.*]], zeroinitializer
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp arcp ord <2 x float> [[A:%.*]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %call = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)
@@ -308,7 +308,7 @@ define <2 x i1> @fabs_ord(<2 x float> %a) {
 
 define <2 x i1> @fabs_uno(<2 x float> %a) {
 ; CHECK-LABEL: @fabs_uno(
-; CHECK-NEXT:    [[CMP:%.*]] = fcmp uno <2 x float> [[A:%.*]], zeroinitializer
+; CHECK-NEXT:    [[CMP:%.*]] = fcmp arcp uno <2 x float> [[A:%.*]], zeroinitializer
 ; CHECK-NEXT:    ret <2 x i1> [[CMP]]
 ;
   %call = call <2 x float> @llvm.fabs.v2f32(<2 x float> %a)