[InstCombine] Handle more even/odd math functions

[agentcooper]

At the moment PR adds support only for erf function.

Does it make sense to create a list of known even/odd functions and call the newly introduced LibCallSimplifier::optimizeSymmetric at the point where optimizeTrigReflections is now called?

Fixes #77220.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

@llvm/pr-subscribers-llvm-analysis

Author: Artem Tyurin (agentcooper)

Changes

At the moment PR adds support only for erf function.

Does it make sense to create a list of known even/odd functions and call the newly introduced LibCallSimplifier::optimizeSymmetric at the point where optimizeTrigReflections is now called?

Fixes #77220.

---

Full diff: https://github.com/llvm/llvm-project/pull/81324.diff

4 Files Affected:

• (modified) llvm/include/llvm/Analysis/TargetLibraryInfo.def (+15)
• (modified) llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h (+1)
• (modified) llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp (+19)
• (added) llvm/test/Transforms/InstCombine/math-odd-even-parity.ll (+17)

diff --git a/llvm/include/llvm/Analysis/TargetLibraryInfo.def b/llvm/include/llvm/Analysis/TargetLibraryInfo.def
index 6bd922eed89e15..a2942b538fad77 100644
--- a/llvm/include/llvm/Analysis/TargetLibraryInfo.def
+++ b/llvm/include/llvm/Analysis/TargetLibraryInfo.def
@@ -1837,6 +1837,21 @@ TLI_DEFINE_ENUM_INTERNAL(pow)
 TLI_DEFINE_STRING_INTERNAL("pow")
 TLI_DEFINE_SIG_INTERNAL(Dbl, Dbl, Dbl)
 
+/// double erf(double x);
+TLI_DEFINE_ENUM_INTERNAL(erf)
+TLI_DEFINE_STRING_INTERNAL("erf")
+TLI_DEFINE_SIG_INTERNAL(Dbl, Dbl)
+
+/// float erf(float x);
+TLI_DEFINE_ENUM_INTERNAL(erff)
+TLI_DEFINE_STRING_INTERNAL("erff")
+TLI_DEFINE_SIG_INTERNAL(Flt, Flt)
+
+/// long double cbrtl(long double x);
+TLI_DEFINE_ENUM_INTERNAL(erfl)
+TLI_DEFINE_STRING_INTERNAL("erfl")
+TLI_DEFINE_SIG_INTERNAL(LDbl, LDbl)
+
 /// float powf(float x, float y);
 TLI_DEFINE_ENUM_INTERNAL(powf)
 TLI_DEFINE_STRING_INTERNAL("powf")
diff --git a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
index 1b6b525b19caef..05a788e4ea30f4 100644
--- a/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
+++ b/llvm/include/llvm/Transforms/Utils/SimplifyLibCalls.h
@@ -204,6 +204,7 @@ class LibCallSimplifier {
   Value *mergeSqrtToExp(CallInst *CI, IRBuilderBase &B);
   Value *optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B);
   Value *optimizeTrigInversionPairs(CallInst *CI, IRBuilderBase &B);
+  Value *optimizeSymmetric(CallInst *CI, bool isEven, IRBuilderBase &B);
   // Wrapper for all floating point library call optimizations
   Value *optimizeFloatingPointLibCall(CallInst *CI, LibFunc Func,
                                       IRBuilderBase &B);
diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 26a34aa99e1b87..12551dc8f2befc 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -2797,6 +2797,21 @@ static bool insertSinCosCall(IRBuilderBase &B, Function *OrigCallee, Value *Arg,
   return true;
 }
 
+Value *LibCallSimplifier::optimizeSymmetric(CallInst *CI, bool IsEven,
+                                            IRBuilderBase &B) {
+  Value *X;
+  if (match(CI->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X))))) {
+    auto *CallInst = copyFlags(*CI, B.CreateCall(CI->getCalledFunction(), {X}));
+    if (IsEven) {
+      // Even function: f(-x) = f(x)
+      return CallInst;
+    }
+    // Odd function: f(-x) = -f(x)
+    return B.CreateFNeg(CallInst);
+  }
+  return nullptr;
+}
+
 Value *LibCallSimplifier::optimizeSinCosPi(CallInst *CI, bool IsSin, IRBuilderBase &B) {
   // Make sure the prototype is as expected, otherwise the rest of the
   // function is probably invalid and likely to abort.
@@ -3779,6 +3794,10 @@ Value *LibCallSimplifier::optimizeFloatingPointLibCall(CallInst *CI,
   case LibFunc_cabsf:
   case LibFunc_cabsl:
     return optimizeCAbs(CI, Builder);
+  case LibFunc_erf:
+  case LibFunc_erff:
+  case LibFunc_erfl:
+    return optimizeSymmetric(CI, /*IsEven*/ false, Builder);
   default:
     return nullptr;
   }
diff --git a/llvm/test/Transforms/InstCombine/math-odd-even-parity.ll b/llvm/test/Transforms/InstCombine/math-odd-even-parity.ll
new file mode 100644
index 00000000000000..2372ff3c97966b
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/math-odd-even-parity.ll
@@ -0,0 +1,17 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
+; RUN: opt < %s -passes=instcombine -S | FileCheck %s
+
+declare double @erf(double)
+
+; Check odd parity: -erf(-x) == erf(x)
+define double @test_erf1(double %x) {
+; CHECK-LABEL: define double @test_erf1(
+; CHECK-SAME: double [[X:%.*]]) {
+; CHECK-NEXT:    [[TMP1:%.*]] = call double @erf(double [[X]])
+; CHECK-NEXT:    ret double [[TMP1]]
+;
+  %neg_x = fneg double %x
+  %res = call double @erf(double %neg_x)
+  %neg_res = fneg double %res
+  ret double %neg_res
+}

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[agentcooper]

@pogo59 Could you please help with updating llvm/test/tools/llvm-tli-checker/ps4-tli-check.yaml? It is unclear to me how it should be updated when adding new math.h functions.

[Krishna-13-cyber]

@pogo59 Could you please help with updating llvm/test/tools/llvm-tli-checker/ps4-tli-check.yaml? It is unclear to me how it should be updated when adding new math.h functions.

Hi @agentcooper,
You can just set the added functions as unavailable for PS4/PS5 in TargetLibraryInfo.cpp which will help you pass all the checks.

    +TLI.setUnavailable(LibFunc_erf);
    +TLI.setUnavailable(LibFunc_erff);
    +TLI.setUnavailable(LibFunc_erfl);

[dtcxzyw]

LGTM. Thanks!
Please wait for additional approval from other reviewers.

[dtcxzyw]

I think f(fabs(x)) --> f(x) and f(copysign(x, y)) --> f(x) also hold for even functions.

[pogo59]

You can just set the added functions as unavailable for PS4/PS5 in TargetLibraryInfo.cpp which will help you pass all the checks.

That will work for now. I'll add an internal task to handle these properly on our side. Thanks!

[agentcooper]

@arsenm @jcranmer-intel Could you please take a look?

[agentcooper]

@jcranmer-intel @dtcxzyw Could you please land this? I don't have commit access. The Linux build is green on CI and failure in the Windows build seems to be unrelated to this PR.

[dtcxzyw]

Thanks for your contribution!