[SimplifyLibCalls] Shrink sin, cos to sinf, cosf when allowed

[guy-david]

This optimization already exists, but for the libcall versions of these functions and not for their intrinsic form.
Solves #139044.

There are probably more opportunities for other intrinsics, because the switch-case in LibCallSimplifier::optimizeCall covers only pow, exp2, log, log2, log10, sqrt, memset, memcpy and memmove.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Guy David (guy-david)

Changes

This optimization already exists, but for the libcall versions of these functions and not for their intrinsic form.

---

Full diff: https://github.com/llvm/llvm-project/pull/139082.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp (+5)
• (added) llvm/test/Transforms/InstCombine/simplify-intrinsics.ll (+61)

diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index 941e787f91eff..94a79ad824370 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -4136,6 +4136,11 @@ Value *LibCallSimplifier::optimizeCall(CallInst *CI, IRBuilderBase &Builder) {
       return optimizeMemCpy(CI, Builder);
     case Intrinsic::memmove:
       return optimizeMemMove(CI, Builder);
+    case Intrinsic::sin:
+    case Intrinsic::cos:
+      if (UnsafeFPShrink)
+        return optimizeUnaryDoubleFP(CI, Builder, TLI, /*isPrecise=*/true);
+      return nullptr;
     default:
       return nullptr;
     }
diff --git a/llvm/test/Transforms/InstCombine/simplify-intrinsics.ll b/llvm/test/Transforms/InstCombine/simplify-intrinsics.ll
new file mode 100644
index 0000000000000..8d7f70b6d1a61
--- /dev/null
+++ b/llvm/test/Transforms/InstCombine/simplify-intrinsics.ll
@@ -0,0 +1,61 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt < %s -passes=instcombine -S                             | FileCheck %s --check-prefixes=ANY,NO-FLOAT-SHRINK
+; RUN: opt < %s -passes=instcombine -enable-double-float-shrink -S | FileCheck %s --check-prefixes=ANY,DO-FLOAT-SHRINK
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+
+declare double @llvm.cos.f64(double)
+declare float @llvm.cos.f32(float)
+
+declare double @llvm.sin.f64(double)
+declare float @llvm.sin.f32(float)
+
+; cos -> cosf
+
+; ANY-LABEL: define float @cos_no_fastmath(float %f) {
+; NO-FLOAT-SHRINK: call double @llvm.cos.f64
+; DO-FLOAT-SHRINK: %[[RES:[0-9]+]] = call float @llvm.cos.f32
+; DO-FLOAT-SHRINK: ret float %[[RES]]
+; ANY: }
+define float @cos_no_fastmath(float %f) {
+  %d = fpext float %f to double
+  %result = call double @llvm.cos.f64(double %d)
+  %truncated_result = fptrunc double %result to float
+  ret float %truncated_result
+}
+
+; ANY-LABEL: define float @cos_fastmath(float %f) {
+; ANY: %[[RES:[0-9]+]] = call fast float @llvm.cos.f32
+; ANY: ret float %[[RES]]
+; ANY: }
+define float @cos_fastmath(float %f) {
+  %d = fpext fast float %f to double
+  %result = call fast double @llvm.cos.f64(double %d)
+  %truncated_result = fptrunc fast double %result to float
+  ret float %truncated_result
+}
+
+; sin -> sinf
+
+; ANY-LABEL: define float @sin_no_fastmath(float %f) {
+; NO-FLOAT-SHRINK: call double @llvm.sin.f64
+; DO-FLOAT-SHRINK: %[[RES:[0-9]+]] = call float @llvm.sin.f32
+; DO-FLOAT-SHRINK: ret float %[[RES]]
+; ANY: }
+define float @sin_no_fastmath(float %f) {
+  %d = fpext float %f to double
+  %result = call double @llvm.sin.f64(double %d)
+  %truncated_result = fptrunc double %result to float
+  ret float %truncated_result
+}
+
+; ANY-LABEL: define float @sin_fastmath(float %f) {
+; ANY: %[[RES:[0-9]+]] = call fast float @llvm.sin.f32
+; ANY: ret float %[[RES]]
+; ANY: }
+define float @sin_fastmath(float %f) {
+  %d = fpext fast float %f to double
+  %result = call fast double @llvm.sin.f64(double %d)
+  %truncated_result = fptrunc fast double %result to float
+  ret float %truncated_result
+}

[fhahn]

is this just valid with all fast-math flags, or some subset?

[guy-david]

The existing optimization for libcalls requires all fast-math flags to be present.

[arsenm]

Probably should fix that, this shouldn't require all of them (in particular nsz or contract)

[arsenm]

LGTM but this can handle the vector case as well

[arsenm]

Also test vector case

[guy-david]

This doesn't work, I think because of this check:

llvm-project/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp

Line 1910 in 0577240

if (!CI->getType()->isDoubleTy() || !CalleeFn)

Let's build up on it in a subsequent PR.

[fhahn]

LGTM, thanks

Would be good to do this for all relevant FP functions