InstCombine: Try to use exp10 intrinsic instead of libcall #92287

arsenm · 2024-05-15T16:04:46Z

Addresses old TODO about the exp10 intrinsic not existing.

llvmbot · 2024-05-15T16:05:19Z

@llvm/pr-subscribers-llvm-transforms

Author: Matt Arsenault (arsenm)

Changes

Addresses old TODO about the exp10 intrinsic not existing.

Full diff: https://github.com/llvm/llvm-project/pull/92287.diff

2 Files Affected:

(modified) llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp (+9-2)
(modified) llvm/test/Transforms/InstCombine/pow-1.ll (+131)

diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 9cb8e20b4806f..69b0f78530d7d 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -2134,12 +2134,19 @@ Value *LibCallSimplifier::replacePowWithExp(CallInst *Pow, IRBuilderBase &B) {
   }
 
   // pow(10.0, x) -> exp10(x)
-  // TODO: There is no exp10() intrinsic yet, but some day there shall be one.
   if (match(Base, m_SpecificFP(10.0)) &&
-      hasFloatFn(M, TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l))
+      hasFloatFn(M, TLI, Ty, LibFunc_exp10, LibFunc_exp10f, LibFunc_exp10l)) {
+
+    if (Pow->doesNotAccessMemory()) {
+      CallInst *NewExp10 =
+          B.CreateIntrinsic(Intrinsic::exp10, {Ty}, {Expo}, Pow, "exp10");
+      return copyFlags(*Pow, NewExp10);
+    }
+
     return copyFlags(*Pow, emitUnaryFloatFnCall(Expo, TLI, LibFunc_exp10,
                                                 LibFunc_exp10f, LibFunc_exp10l,
                                                 B, NoAttrs));
+  }
 
   // pow(x, y) -> exp2(log2(x) * y)
   if (Pow->hasApproxFunc() && Pow->hasNoNaNs() && BaseF->isFiniteNonZero() &&
diff --git a/llvm/test/Transforms/InstCombine/pow-1.ll b/llvm/test/Transforms/InstCombine/pow-1.ll
index 7a72902126f84..c52b97c38893b 100644
--- a/llvm/test/Transforms/InstCombine/pow-1.ll
+++ b/llvm/test/Transforms/InstCombine/pow-1.ll
@@ -893,3 +893,134 @@ define double @test_simplify19(double %x) {
   %retval = call double @pow(double 10.0, double %x)
   ret double %retval
 }
+
+define float @test_libcall_powf_10_f32_noerrno(float %x) {
+; CHECK-EXP10-LABEL: define float @test_libcall_powf_10_f32_noerrno(
+; CHECK-EXP10-SAME: float [[X:%.*]]) {
+; CHECK-EXP10-NEXT:    [[__EXP10F:%.*]] = call float @llvm.exp10.f32(float [[X]])
+; CHECK-EXP10-NEXT:    ret float [[__EXP10F]]
+;
+; CHECK-NO-EXP10-LABEL: define float @test_libcall_powf_10_f32_noerrno(
+; CHECK-NO-EXP10-SAME: float [[X:%.*]]) {
+; CHECK-NO-EXP10-NEXT:    [[RETVAL:%.*]] = call float @powf(float 1.000000e+01, float [[X]]) #[[ATTR3:[0-9]+]]
+; CHECK-NO-EXP10-NEXT:    ret float [[RETVAL]]
+;
+  %retval = call float @powf(float 10.0, float %x) memory(none)
+  ret float %retval
+}
+
+define double @test_libcall_pow_10_f64_noerrno(double %x) {
+; CHECK-EXP10-LABEL: define double @test_libcall_pow_10_f64_noerrno(
+; CHECK-EXP10-SAME: double [[X:%.*]]) {
+; CHECK-EXP10-NEXT:    [[__EXP10:%.*]] = call double @llvm.exp10.f64(double [[X]])
+; CHECK-EXP10-NEXT:    ret double [[__EXP10]]
+;
+; CHECK-NO-EXP10-LABEL: define double @test_libcall_pow_10_f64_noerrno(
+; CHECK-NO-EXP10-SAME: double [[X:%.*]]) {
+; CHECK-NO-EXP10-NEXT:    [[RETVAL:%.*]] = call double @pow(double 1.000000e+01, double [[X]]) #[[ATTR3]]
+; CHECK-NO-EXP10-NEXT:    ret double [[RETVAL]]
+;
+  %retval = call double @pow(double 10.0, double %x) memory(none)
+  ret double %retval
+}
+
+define half @test_pow_10_f16(half %x) {
+; CHECK-LABEL: define half @test_pow_10_f16(
+; CHECK-SAME: half [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call half @llvm.pow.f16(half 0xH4900, half [[X]])
+; CHECK-NEXT:    ret half [[RETVAL]]
+;
+  %retval = call half @llvm.pow.f16(half 10.0, half %x)
+  ret half %retval
+}
+
+define float @test_pow_10_f32(float %x) {
+; CHECK-EXP10-LABEL: define float @test_pow_10_f32(
+; CHECK-EXP10-SAME: float [[X:%.*]]) {
+; CHECK-EXP10-NEXT:    [[__EXP10F:%.*]] = call float @llvm.exp10.f32(float [[X]])
+; CHECK-EXP10-NEXT:    ret float [[__EXP10F]]
+;
+; CHECK-NO-EXP10-LABEL: define float @test_pow_10_f32(
+; CHECK-NO-EXP10-SAME: float [[X:%.*]]) {
+; CHECK-NO-EXP10-NEXT:    [[RETVAL:%.*]] = call float @llvm.pow.f32(float 1.000000e+01, float [[X]])
+; CHECK-NO-EXP10-NEXT:    ret float [[RETVAL]]
+;
+  %retval = call float @llvm.pow.f32(float 10.0, float %x)
+  ret float %retval
+}
+
+define double @test_pow_10_f64(double %x) {
+; CHECK-EXP10-LABEL: define double @test_pow_10_f64(
+; CHECK-EXP10-SAME: double [[X:%.*]]) {
+; CHECK-EXP10-NEXT:    [[__EXP10:%.*]] = call double @llvm.exp10.f64(double [[X]])
+; CHECK-EXP10-NEXT:    ret double [[__EXP10]]
+;
+; CHECK-NO-EXP10-LABEL: define double @test_pow_10_f64(
+; CHECK-NO-EXP10-SAME: double [[X:%.*]]) {
+; CHECK-NO-EXP10-NEXT:    [[RETVAL:%.*]] = call double @llvm.pow.f64(double 1.000000e+01, double [[X]])
+; CHECK-NO-EXP10-NEXT:    ret double [[RETVAL]]
+;
+  %retval = call double @llvm.pow.f64(double 10.0, double %x)
+  ret double %retval
+}
+
+define fp128 @test_pow_10_fp128(fp128 %x) {
+; CHECK-LABEL: define fp128 @test_pow_10_fp128(
+; CHECK-SAME: fp128 [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call fp128 @llvm.pow.f128(fp128 0xL00000000000000004002400000000000, fp128 [[X]])
+; CHECK-NEXT:    ret fp128 [[RETVAL]]
+;
+  %ten = fpext double 10.0 to fp128
+  %retval = call fp128 @llvm.pow.fp128(fp128 %ten, fp128 %x)
+  ret fp128 %retval
+}
+
+define bfloat @test_pow_10_bf16(bfloat %x) {
+; CHECK-LABEL: define bfloat @test_pow_10_bf16(
+; CHECK-SAME: bfloat [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call bfloat @llvm.pow.bf16(bfloat 0xR4120, bfloat [[X]])
+; CHECK-NEXT:    ret bfloat [[RETVAL]]
+;
+  %retval = call bfloat @llvm.pow.bf16(bfloat 10.0, bfloat %x)
+  ret bfloat %retval
+}
+
+define <2 x half> @test_pow_10_v2f16(<2 x half> %x) {
+; CHECK-LABEL: define <2 x half> @test_pow_10_v2f16(
+; CHECK-SAME: <2 x half> [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call <2 x half> @llvm.pow.v2f16(<2 x half> <half 0xH4900, half 0xH4900>, <2 x half> [[X]])
+; CHECK-NEXT:    ret <2 x half> [[RETVAL]]
+;
+  %retval = call <2 x half> @llvm.pow.v2f16(<2 x half> <half 10.0, half 10.0>, <2 x half> %x)
+  ret <2 x half> %retval
+}
+
+define <2 x float> @test_pow_10_v2f32(<2 x float> %x) {
+; CHECK-LABEL: define <2 x float> @test_pow_10_v2f32(
+; CHECK-SAME: <2 x float> [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call <2 x float> @llvm.pow.v2f32(<2 x float> <float 1.000000e+01, float 1.000000e+01>, <2 x float> [[X]])
+; CHECK-NEXT:    ret <2 x float> [[RETVAL]]
+;
+  %retval = call <2 x float> @llvm.pow.v2f32(<2 x float> <float 10.0, float 10.0>, <2 x float> %x)
+  ret <2 x float> %retval
+}
+
+define <2 x double> @test_pow_10_v2f64(<2 x double> %x) {
+; CHECK-LABEL: define <2 x double> @test_pow_10_v2f64(
+; CHECK-SAME: <2 x double> [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call <2 x double> @llvm.pow.v2f64(<2 x double> <double 1.000000e+01, double 1.000000e+01>, <2 x double> [[X]])
+; CHECK-NEXT:    ret <2 x double> [[RETVAL]]
+;
+  %retval = call <2 x double> @llvm.pow.v2f64(<2 x double> <double 10.0, double 10.0>, <2 x double> %x)
+  ret <2 x double> %retval
+}
+
+define <2 x bfloat> @test_pow_10_v2bf16(<2 x bfloat> %x) {
+; CHECK-LABEL: define <2 x bfloat> @test_pow_10_v2bf16(
+; CHECK-SAME: <2 x bfloat> [[X:%.*]]) {
+; CHECK-NEXT:    [[RETVAL:%.*]] = call <2 x bfloat> @llvm.pow.v2bf16(<2 x bfloat> <bfloat 0xR4120, bfloat 0xR4120>, <2 x bfloat> [[X]])
+; CHECK-NEXT:    ret <2 x bfloat> [[RETVAL]]
+;
+  %retval = call <2 x bfloat> @llvm.pow.v2bf16(<2 x bfloat> <bfloat 10.0, bfloat 10.0>, <2 x bfloat> %x)
+  ret <2 x bfloat> %retval
+}

nikic

LGTM. It would be nice to have a common helper to handle this recurring "intrinsic or libcall" pattern...

glandium · 2024-05-17T04:07:24Z

This broke mac and ios targets, with a missing exp10 symbol at link time.

arsenm · 2024-05-17T07:20:46Z

This broke mac and ios targets, with a missing exp10 symbol at link time.

Testcase? That suggest TargetLibraryInfo is incorrectly reporting exp10 as available

arsenm · 2024-05-17T07:27:52Z

This broke mac and ios targets, with a missing exp10 symbol at link time.

Testcase? That suggest TargetLibraryInfo is incorrectly reporting exp10 as available

This seems pretty well covered already:

llvm-project/llvm/lib/Analysis/TargetLibraryInfo.cpp

Line 537 in e0a293d

case Triple::MacOSX:

glandium · 2024-05-17T07:37:46Z

It's complaining about exp10 being missing, not __exp10, so I guess the information from there is not correctly used.

glandium · 2024-05-17T10:04:40Z

Here's a reproducer

long n, x;
void foo() { x = __builtin_powl(10, n); }

Build with clang -O1 -o libfoo.dylib -shared foo.c --target=aarch64-apple-darwin -isysroot /path/to/MacOSX.sdk/

glandium · 2024-05-17T10:05:25Z

Before the change, this compiled to use __exp10, now, it uses exp10.

It's really great that we have the same information duplicated in TargetLibraryInfo and RuntimeLibcalls which both assume everything by default. Should fix issue reported after llvm#92287

fhahn · 2024-05-20T12:51:07Z

As this is breaking builds on Darwin platforms I am planning to revert this soon to unbreak builds until #92520 lands

arsenm · 2024-05-20T12:57:05Z

As this is breaking builds on Darwin platforms I am planning to revert this soon to unbreak builds until #92520 lands

How about just review it instead

It's really great that we have the same information duplicated in TargetLibraryInfo and RuntimeLibcalls which both assume everything by default. Should fix issue reported after #92287

arsenm added 2 commits May 15, 2024 18:03

InstCombine: Add more pow(10, x) -> exp10 tests

57ae03e

InstCombine: Try to use exp10 intrinsic instead of libcall

565ef3f

Addresses old TODO about the exp10 intrinsic not existing.

arsenm added the llvm:instcombine label May 15, 2024

arsenm requested review from nikic, andykaylor, jcranmer-intel and efriedma-quic May 15, 2024 16:04

llvmbot added the llvm:transforms label May 15, 2024

nikic approved these changes May 16, 2024

View reviewed changes

arsenm merged commit ce1ce5d into llvm:main May 16, 2024
7 checks passed

arsenm deleted the instcombine-use-exp10-intrinsic branch May 16, 2024 07:09

arsenm mentioned this pull request May 17, 2024

CodeGen: Fix libcall names for exp10 on the various darwins #92520

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

InstCombine: Try to use exp10 intrinsic instead of libcall #92287

InstCombine: Try to use exp10 intrinsic instead of libcall #92287

arsenm commented May 15, 2024

llvmbot commented May 15, 2024

nikic left a comment

glandium commented May 17, 2024

arsenm commented May 17, 2024

arsenm commented May 17, 2024

glandium commented May 17, 2024

glandium commented May 17, 2024

glandium commented May 17, 2024

fhahn commented May 20, 2024

arsenm commented May 20, 2024

InstCombine: Try to use exp10 intrinsic instead of libcall #92287

InstCombine: Try to use exp10 intrinsic instead of libcall #92287

Conversation

arsenm commented May 15, 2024

llvmbot commented May 15, 2024

nikic left a comment

Choose a reason for hiding this comment

glandium commented May 17, 2024

arsenm commented May 17, 2024

arsenm commented May 17, 2024

glandium commented May 17, 2024

glandium commented May 17, 2024

glandium commented May 17, 2024

fhahn commented May 20, 2024

arsenm commented May 20, 2024