[llvm] [InstCombine] Fold fmod to frem if we know it does not set errno. (PR #107912)
David Green via llvm-commits
llvm-commits at lists.llvm.org
Thu Sep 12 13:14:28 PDT 2024
https://github.com/davemgreen updated https://github.com/llvm/llvm-project/pull/107912
>From ce2375c453fa7bf761b329a37daf38f29186808d Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Thu, 12 Sep 2024 19:17:06 +0100
Subject: [PATCH 1/2] [InstCombine] Pass DomTree and DomTreeCacheto
LibCallSimplifier
This allows any combines to pick up Known states from dominating conditions.
---
.../llvm/Transforms/Utils/SimplifyLibCalls.h | 7 +++-
.../InstCombine/InstCombineCalls.cpp | 4 +--
.../lib/Transforms/Utils/SimplifyLibCalls.cpp | 14 ++++----
llvm/test/Transforms/InstCombine/pow-1.ll | 32 +++++++++++++------
4 files changed, 37 insertions(+), 20 deletions(-)
diff --git a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
index 43b5c9250a8908..2e7a0ec29ed999 100644
--- a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
+++ b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
@@ -22,6 +22,8 @@ class AssumptionCache;
class StringRef;
class Value;
class CallInst;
+class DominatorTree;
+class DomConditionCache;
class DataLayout;
class Instruction;
class IRBuilderBase;
@@ -103,6 +105,8 @@ class LibCallSimplifier {
FortifiedLibCallSimplifier FortifiedSimplifier;
const DataLayout &DL;
const TargetLibraryInfo *TLI;
+ DominatorTree *DT;
+ DomConditionCache *DC;
AssumptionCache *AC;
OptimizationRemarkEmitter &ORE;
BlockFrequencyInfo *BFI;
@@ -136,7 +140,8 @@ class LibCallSimplifier {
public:
LibCallSimplifier(
- const DataLayout &DL, const TargetLibraryInfo *TLI, AssumptionCache *AC,
+ const DataLayout &DL, const TargetLibraryInfo *TLI, DominatorTree *DT,
+ DomConditionCache *DC, AssumptionCache *AC,
OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
ProfileSummaryInfo *PSI,
function_ref<void(Instruction *, Value *)> Replacer =
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index eb94e894b57b06..61011d55227e7b 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -3711,8 +3711,8 @@ Instruction *InstCombinerImpl::tryOptimizeCall(CallInst *CI) {
auto InstCombineErase = [this](Instruction *I) {
eraseInstFromFunction(*I);
};
- LibCallSimplifier Simplifier(DL, &TLI, &AC, ORE, BFI, PSI, InstCombineRAUW,
- InstCombineErase);
+ LibCallSimplifier Simplifier(DL, &TLI, &DT, &DC, &AC, ORE, BFI, PSI,
+ InstCombineRAUW, InstCombineErase);
if (Value *With = Simplifier.optimizeCall(CI, Builder)) {
++NumSimplified;
return CI->use_empty() ? CI : replaceInstUsesWith(*CI, With);
diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 1e6dc88ed93532..917f81863cf673 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -2278,8 +2278,8 @@ Value *LibCallSimplifier::replacePowWithSqrt(CallInst *Pow, IRBuilderBase &B) {
// pow(-Inf, 0.5) is optionally required to have a result of +Inf (not setting
// errno), but sqrt(-Inf) is required by various standards to set errno.
if (!Pow->doesNotAccessMemory() && !Pow->hasNoInfs() &&
- !isKnownNeverInfinity(Base, 0,
- SimplifyQuery(DL, TLI, /*DT=*/nullptr, AC, Pow)))
+ !isKnownNeverInfinity(
+ Base, 0, SimplifyQuery(DL, TLI, DT, AC, Pow, true, true, DC)))
return nullptr;
Sqrt = getSqrtCall(Base, AttributeList(), Pow->doesNotAccessMemory(), Mod, B,
@@ -4162,13 +4162,13 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) {
}
LibCallSimplifier::LibCallSimplifier(
- const DataLayout &DL, const TargetLibraryInfo *TLI, AssumptionCache *AC,
- OptimizationRemarkEmitter &ORE, BlockFrequencyInfo *BFI,
- ProfileSummaryInfo *PSI,
+ const DataLayout &DL, const TargetLibraryInfo *TLI, DominatorTree *DT,
+ DomConditionCache *DC, AssumptionCache *AC, OptimizationRemarkEmitter &ORE,
+ BlockFrequencyInfo *BFI, ProfileSummaryInfo *PSI,
function_ref<void(Instruction *, Value *)> Replacer,
function_ref<void(Instruction *)> Eraser)
- : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), AC(AC), ORE(ORE), BFI(BFI),
- PSI(PSI), Replacer(Replacer), Eraser(Eraser) {}
+ : FortifiedSimplifier(TLI), DL(DL), TLI(TLI), DT(DT), DC(DC), AC(AC),
+ ORE(ORE), BFI(BFI), PSI(PSI), Replacer(Replacer), Eraser(Eraser) {}
void LibCallSimplifier::replaceAllUsesWith(Instruction *I, Value *With) {
// Indirect through the replacer used in this instance.
diff --git a/llvm/test/Transforms/InstCombine/pow-1.ll b/llvm/test/Transforms/InstCombine/pow-1.ll
index f4bbd3eb46bf79..91e909aa25c7e3 100644
--- a/llvm/test/Transforms/InstCombine/pow-1.ll
+++ b/llvm/test/Transforms/InstCombine/pow-1.ll
@@ -863,16 +863,28 @@ define double @pow_libcall_half_no_FMF(double %x) {
}
define double @pow_libcall_half_fromdomcondition(double %x) {
-; CHECK-LABEL: define double @pow_libcall_half_fromdomcondition(
-; CHECK-SAME: double [[X:%.*]]) {
-; CHECK-NEXT: [[A:%.*]] = call double @llvm.fabs.f64(double [[X]])
-; CHECK-NEXT: [[C:%.*]] = fcmp oeq double [[A]], 0x7FF0000000000000
-; CHECK-NEXT: br i1 [[C]], label [[THEN:%.*]], label [[ELSE:%.*]]
-; CHECK: then:
-; CHECK-NEXT: ret double 0.000000e+00
-; CHECK: else:
-; CHECK-NEXT: [[RETVAL:%.*]] = call double @pow(double [[X]], double 5.000000e-01)
-; CHECK-NEXT: ret double [[RETVAL]]
+; LIB-LABEL: define double @pow_libcall_half_fromdomcondition(
+; LIB-SAME: double [[X:%.*]]) {
+; LIB-NEXT: [[A:%.*]] = call double @llvm.fabs.f64(double [[X]])
+; LIB-NEXT: [[C:%.*]] = fcmp oeq double [[A]], 0x7FF0000000000000
+; LIB-NEXT: br i1 [[C]], label [[THEN:%.*]], label [[ELSE:%.*]]
+; LIB: then:
+; LIB-NEXT: ret double 0.000000e+00
+; LIB: else:
+; LIB-NEXT: [[SQRT:%.*]] = call double @sqrt(double [[X]])
+; LIB-NEXT: [[ABS:%.*]] = call double @llvm.fabs.f64(double [[SQRT]])
+; LIB-NEXT: ret double [[ABS]]
+;
+; NOLIB-LABEL: define double @pow_libcall_half_fromdomcondition(
+; NOLIB-SAME: double [[X:%.*]]) {
+; NOLIB-NEXT: [[A:%.*]] = call double @llvm.fabs.f64(double [[X]])
+; NOLIB-NEXT: [[C:%.*]] = fcmp oeq double [[A]], 0x7FF0000000000000
+; NOLIB-NEXT: br i1 [[C]], label [[THEN:%.*]], label [[ELSE:%.*]]
+; NOLIB: then:
+; NOLIB-NEXT: ret double 0.000000e+00
+; NOLIB: else:
+; NOLIB-NEXT: [[RETVAL:%.*]] = call double @pow(double [[X]], double 5.000000e-01)
+; NOLIB-NEXT: ret double [[RETVAL]]
;
%a = call double @llvm.fabs.f64(double %x)
%c = fcmp oeq double %a, 0x7FF0000000000000
>From 6bd9b29a88e3c6d15f0ba93ae86669154031d40a Mon Sep 17 00:00:00 2001
From: David Green <david.green at arm.com>
Date: Thu, 12 Sep 2024 21:14:10 +0100
Subject: [PATCH 2/2] [InstCombine] Fold fmod to frem if we know it does not
set errno.
fmod will be folded to frem in clang under -fno-math-errno and can be constant
folded in llvm if the operands are known. It can be relatively common to have
fp code that handles special values before doing some calculation:
```
if (isnan(f))
return handlenan;
if (isinf(f))
return handleinf;
..
fmod(f, 2.0)
```
This patch enables the folding of fmod to frem in InstCombine (via
LibCallSimplifier) if the first parameter is not inf and the second is not
zero. Other combinations do not set errno.
---
.../llvm/Transforms/Utils/SimplifyLibCalls.h | 1 +
.../lib/Transforms/Utils/SimplifyLibCalls.cpp | 25 +++++++++++++++++++
llvm/test/Transforms/InstCombine/fmod.ll | 10 ++++----
3 files changed, 31 insertions(+), 5 deletions(-)
diff --git a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
index 2e7a0ec29ed999..2d3d2ada6183a7 100644
--- a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
+++ b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
@@ -206,6 +206,7 @@ class LibCallSimplifier {
Value *optimizeFMinFMax(CallInst *CI, IRBuilderBase &B);
Value *optimizeLog(CallInst *CI, IRBuilderBase &B);
Value *optimizeSqrt(CallInst *CI, IRBuilderBase &B);
+ Value *optimizeFMod(CallInst *CI, IRBuilderBase &B);
Value *mergeSqrtToExp(CallInst *CI, IRBuilderBase &B);
Value *optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B);
Value *optimizeTrigInversionPairs(CallInst *CI, IRBuilderBase &B);
diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 917f81863cf673..b9f20218bc89c5 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -2796,6 +2796,27 @@ Value *LibCallSimplifier::optimizeSqrt(CallInst *CI, IRBuilderBase &B) {
return copyFlags(*CI, FabsCall);
}
+Value *LibCallSimplifier::optimizeFMod(CallInst *CI, IRBuilderBase &B) {
+ SimplifyQuery SQ(DL, TLI, DT, AC, CI, true, true, DC);
+
+ // fmod(x,y) can set errno if y == 0 or x == +/-inf. Otherwise we can change
+ // it to a frem instruction.
+ KnownFPClass Known0 = computeKnownFPClass(CI->getOperand(0), fcInf,
+ /*Depth=*/0, SQ);
+ if (Known0.isKnownNeverInfinity()) {
+ KnownFPClass Known1 = computeKnownFPClass(CI->getOperand(1), fcZero,
+ /*Depth=*/0, SQ);
+ Function *F = CI->getParent()->getParent();
+ if (Known1.isKnownNeverLogicalZero(*F, CI->getType())) {
+ Value *FRem = B.CreateFRemFMF(CI->getOperand(0), CI->getOperand(1), CI);
+ substituteInParent(CI, FRem);
+ return nullptr;
+ }
+ }
+
+ return nullptr;
+}
+
Value *LibCallSimplifier::optimizeTrigInversionPairs(CallInst *CI,
IRBuilderBase &B) {
Module *M = CI->getModule();
@@ -3945,6 +3966,10 @@ Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
case LibFunc_sqrt:
case LibFunc_sqrtl:
return optimizeSqrt(CI, Builder);
+ case LibFunc_fmod:
+ case LibFunc_fmodf:
+ case LibFunc_fmodl:
+ return optimizeFMod(CI, Builder);
case LibFunc_logf:
case LibFunc_log:
case LibFunc_logl:
diff --git a/llvm/test/Transforms/InstCombine/fmod.ll b/llvm/test/Transforms/InstCombine/fmod.ll
index c021d27e95fa5e..b9ad1571090578 100644
--- a/llvm/test/Transforms/InstCombine/fmod.ll
+++ b/llvm/test/Transforms/InstCombine/fmod.ll
@@ -9,7 +9,7 @@ define float @test_inf_const(float %f) {
; CHECK-NEXT: [[ISINF:%.*]] = fcmp oeq float [[ABS]], 0x7FF0000000000000
; CHECK-NEXT: br i1 [[ISINF]], label [[RETURN:%.*]], label [[IF_END:%.*]]
; CHECK: if.end:
-; CHECK-NEXT: [[CALL:%.*]] = tail call float @fmodf(float [[F]], float 2.000000e+00)
+; CHECK-NEXT: [[CALL:%.*]] = frem float [[F]], 2.000000e+00
; CHECK-NEXT: ret float [[CALL]]
; CHECK: return:
; CHECK-NEXT: ret float 0.000000e+00
@@ -34,7 +34,7 @@ define float @test_const_zero(float %f) {
; CHECK-NEXT: [[ISZERO:%.*]] = fcmp oeq float [[F]], 0.000000e+00
; CHECK-NEXT: br i1 [[ISZERO]], label [[RETURN:%.*]], label [[IF_END:%.*]]
; CHECK: if.end:
-; CHECK-NEXT: [[CALL:%.*]] = tail call float @fmodf(float 2.000000e+00, float [[F]])
+; CHECK-NEXT: [[CALL:%.*]] = frem float 2.000000e+00, [[F]]
; CHECK-NEXT: ret float [[CALL]]
; CHECK: return:
; CHECK-NEXT: ret float 0.000000e+00
@@ -67,7 +67,7 @@ define float @test_noinf_nozero(float nofpclass(inf) %f, float nofpclass(zero) %
; CHECK-LABEL: define float @test_noinf_nozero(
; CHECK-SAME: float nofpclass(inf) [[F:%.*]], float nofpclass(zero) [[G:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[CALL:%.*]] = tail call nnan float @fmodf(float [[F]], float [[G]])
+; CHECK-NEXT: [[CALL:%.*]] = frem nnan float [[F]], [[G]]
; CHECK-NEXT: ret float [[CALL]]
;
entry:
@@ -79,7 +79,7 @@ define double @test_double(double nofpclass(inf) %f, double nofpclass(zero) %g)
; CHECK-LABEL: define double @test_double(
; CHECK-SAME: double nofpclass(inf) [[F:%.*]], double nofpclass(zero) [[G:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[CALL:%.*]] = tail call double @fmod(double [[F]], double [[G]])
+; CHECK-NEXT: [[CALL:%.*]] = frem double [[F]], [[G]]
; CHECK-NEXT: ret double [[CALL]]
;
entry:
@@ -91,7 +91,7 @@ define fp128 @test_fp128(fp128 nofpclass(inf) %f, fp128 nofpclass(zero) %g) {
; CHECK-LABEL: define fp128 @test_fp128(
; CHECK-SAME: fp128 nofpclass(inf) [[F:%.*]], fp128 nofpclass(zero) [[G:%.*]]) {
; CHECK-NEXT: entry:
-; CHECK-NEXT: [[CALL:%.*]] = tail call fp128 @fmodl(fp128 [[F]], fp128 [[G]])
+; CHECK-NEXT: [[CALL:%.*]] = frem fp128 [[F]], [[G]]
; CHECK-NEXT: ret fp128 [[CALL]]
;
entry:
More information about the llvm-commits
mailing list