[SimplifyLibCalls] Add f128 type for merging sqrt into the power of exp

[lizhijin1024]

No description provided.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: None (lizhijin1024)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/112373.diff

2 Files Affected:

• (modified) llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp (+3)
• (modified) llvm/test/Transforms/InstCombine/sqrt.ll (+22)

diff --git a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
index db2acb9eed0938..aa7864ad4f6c0e 100644
--- a/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
+++ b/llvm/lib/Transforms/Utils/SimplifyLibCalls.cpp
@@ -2720,6 +2720,9 @@ Value *LibCallSimplifier::mergeSqrtToExp(CallInst *CI, IRBuilderBase &B) {
       ExpLb = LibFunc_exp;
       Exp2Lb = LibFunc_exp2;
       Exp10Lb = LibFunc_exp10;
+    } else if (CI->getType()->getScalarType()->isFP128Ty()) {
+      ExpLb = LibFunc_expl;
+      Exp2Lb = LibFunc_exp2l;
     } else
       return nullptr;
   } else
diff --git a/llvm/test/Transforms/InstCombine/sqrt.ll b/llvm/test/Transforms/InstCombine/sqrt.ll
index f72fe5a6a5817b..6168a61cf86cb0 100644
--- a/llvm/test/Transforms/InstCombine/sqrt.ll
+++ b/llvm/test/Transforms/InstCombine/sqrt.ll
@@ -201,6 +201,25 @@ define <2 x float> @sqrt_exp_vec(<2 x float> %x) {
   ret <2 x float> %res
 }
 
+define dso_local fp128 @sqrt_exp_long_double(fp128 %x, fp128 %y) {
+; CHECK-LABEL: @sqrt_exp_long_double(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[MERGED_SQRT:%.*]] = fmul fast fp128 [[X:%.*]], 0xL00000000000000003FFE000000000000
+; CHECK-NEXT:    [[TMP0:%.*]] = call fast fp128 @llvm.exp.f128(fp128 [[MERGED_SQRT]])
+; CHECK-NEXT:    [[MERGED_SQRT1:%.*]] = fmul fast fp128 [[Y:%.*]], 0xL00000000000000003FFE000000000000
+; CHECK-NEXT:    [[TMP1:%.*]] = call fast fp128 @llvm.exp2.f128(fp128 [[MERGED_SQRT1]])
+; CHECK-NEXT:    [[ADD:%.*]] = fadd fast fp128 [[TMP0]], [[TMP1]]
+; CHECK-NEXT:    ret fp128 [[ADD]]
+;
+entry:
+  %0 = call fast fp128 @llvm.exp.f128(fp128 %x)
+  %1 = call fast fp128 @llvm.sqrt.f128(fp128 %0)
+  %2 = call fast fp128 @llvm.exp2.f128(fp128 %y)
+  %3 = call fast fp128 @llvm.sqrt.f128(fp128 %2)
+  %add = fadd fast fp128 %1, %3
+  ret fp128 %add
+}
+
 declare i32 @foo(double)
 declare double @sqrt(double) readnone
 declare float @sqrtf(float)
@@ -212,3 +231,6 @@ declare double @exp2(double)
 declare double @exp10(double)
 declare <2 x float> @llvm.exp.v2f32(<2 x float>)
 declare <2 x float> @llvm.sqrt.v2f32(<2 x float>)
+declare fp128 @llvm.exp.f128(fp128)
+declare fp128 @llvm.sqrt.f128(fp128)
+declare fp128 @llvm.exp2.f128(fp128)

[arsenm]

Don't need dso_local

[arsenm]

Test the different cases in different functions.

Use minimal fast math flags, this looks like it only requires reassoc. Add an additional test which shows the preservation behavior of other fast math flags.

Also test the same situation with equivalent library calls instead of intrinsics (and the mixed intrinsic + libcall case)

[arsenm]

This intrinsic->intrinsic transform really shouldn't depend on the libcall availability...