[llvm] [InstCombine] Skip foldFBinOpOfIntCastsFromSign for vector ops (PR #162804)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-llvm-transforms

Author: Nikita Popov (nikic)

<details>
<summary>Changes</summary>

Converting a vector float op into a vector int op may be non-profitable, especially for targets where the float op for a given type is legal, but the integer op is not.

We could of course also try to address this via a reverse transform in the backend, but I don't think it's worth the bother, given that vectors were never the intended use case for this transform in the first place.

Fixes https://github.com/llvm/llvm-project/issues/162749.

---
Full diff: https://github.com/llvm/llvm-project/pull/162804.diff


3 Files Affected:

- (modified) llvm/lib/Transforms/InstCombine/InstructionCombining.cpp (+5) 
- (modified) llvm/test/Transforms/InstCombine/add-sitofp.ll (+5-2) 
- (modified) llvm/test/Transforms/InstCombine/binop-itofp.ll (+21-1) 


``````````diff
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index d56a1af49ef32..b012a8fd47bec 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -1690,6 +1690,11 @@ Instruction *InstCombinerImpl::foldFBinOpOfIntCastsFromSign(
 //    2) (fp_binop ({s|u}itofp x), FpC)
 //        -> ({s|u}itofp (int_binop x, (fpto{s|u}i FpC)))
 Instruction *InstCombinerImpl::foldFBinOpOfIntCasts(BinaryOperator &BO) {
+  // Don't perform the fold on vectors, as the integer operation may be much
+  // more expensive than the float operation in that case.
+  if (BO.getType()->isVectorTy())
+    return nullptr;
+
   std::array<Value *, 2> IntOps = {nullptr, nullptr};
   Constant *Op1FpC = nullptr;
   // Check for:
diff --git a/llvm/test/Transforms/InstCombine/add-sitofp.ll b/llvm/test/Transforms/InstCombine/add-sitofp.ll
index fae1365dfa853..e1d39fdb8b871 100644
--- a/llvm/test/Transforms/InstCombine/add-sitofp.ll
+++ b/llvm/test/Transforms/InstCombine/add-sitofp.ll
@@ -99,12 +99,15 @@ define float @test_3(i32 %a, i32 %b) {
   ret float %p
 }
 
+; Don't perform the fold on vector operations, as the integer op may be
+; much more expensive than the float op in that case.
 define <4 x double> @test_4(<4 x i32> %a, <4 x i32> %b) {
 ; CHECK-LABEL: @test_4(
 ; CHECK-NEXT:    [[A_AND:%.*]] = and <4 x i32> [[A:%.*]], splat (i32 1073741823)
 ; CHECK-NEXT:    [[B_AND:%.*]] = and <4 x i32> [[B:%.*]], splat (i32 1073741823)
-; CHECK-NEXT:    [[TMP1:%.*]] = add nuw nsw <4 x i32> [[A_AND]], [[B_AND]]
-; CHECK-NEXT:    [[RES:%.*]] = uitofp nneg <4 x i32> [[TMP1]] to <4 x double>
+; CHECK-NEXT:    [[A_AND_FP:%.*]] = uitofp nneg <4 x i32> [[A_AND]] to <4 x double>
+; CHECK-NEXT:    [[B_AND_FP:%.*]] = uitofp nneg <4 x i32> [[B_AND]] to <4 x double>
+; CHECK-NEXT:    [[RES:%.*]] = fadd <4 x double> [[A_AND_FP]], [[B_AND_FP]]
 ; CHECK-NEXT:    ret <4 x double> [[RES]]
 ;
   ; Drop two highest bits to guarantee that %a + %b doesn't overflow
diff --git a/llvm/test/Transforms/InstCombine/binop-itofp.ll b/llvm/test/Transforms/InstCombine/binop-itofp.ll
index 702bbbbf7d176..57184ea54583a 100644
--- a/llvm/test/Transforms/InstCombine/binop-itofp.ll
+++ b/llvm/test/Transforms/InstCombine/binop-itofp.ll
@@ -1063,6 +1063,25 @@ define float @negzero_check_on_constant_for_si_fmul(i1 %c, i1 %.b, ptr %g_2345)
   ret float %mul3.i.i
 }
 
+; Don't perform the fold on vector operations, as the integer op may be
+; much more expensive than the float op in that case.
+define <2 x half> @test_ui_ui_i8_mul_vec(<2 x i8> noundef %x_in, <2 x i8> noundef %y_in) {
+; CHECK-LABEL: @test_ui_ui_i8_mul_vec(
+; CHECK-NEXT:    [[X:%.*]] = and <2 x i8> [[X_IN:%.*]], splat (i8 15)
+; CHECK-NEXT:    [[Y:%.*]] = and <2 x i8> [[Y_IN:%.*]], splat (i8 15)
+; CHECK-NEXT:    [[XF:%.*]] = uitofp nneg <2 x i8> [[X]] to <2 x half>
+; CHECK-NEXT:    [[YF:%.*]] = uitofp nneg <2 x i8> [[Y]] to <2 x half>
+; CHECK-NEXT:    [[R:%.*]] = fmul <2 x half> [[XF]], [[YF]]
+; CHECK-NEXT:    ret <2 x half> [[R]]
+;
+  %x = and <2 x i8> %x_in, splat (i8 15)
+  %y = and <2 x i8> %y_in, splat (i8 15)
+  %xf = uitofp <2 x i8> %x to <2 x half>
+  %yf = uitofp <2 x i8> %y to <2 x half>
+  %r = fmul <2 x half> %xf, %yf
+  ret <2 x half> %r
+}
+
 define <2 x float> @nonzero_check_on_constant_for_si_fmul_vec_w_poison(i1 %c, i1 %.b, ptr %g_2345) {
 ; CHECK-LABEL: @nonzero_check_on_constant_for_si_fmul_vec_w_poison(
 ; CHECK-NEXT:    [[SEL:%.*]] = select i1 [[C:%.*]], i32 65529, i32 53264
@@ -1091,8 +1110,9 @@ define <2 x float> @nonzero_check_on_constant_for_si_fmul_nz_vec_w_poison(i1 %c,
 ; CHECK-NEXT:    [[CONV_I_V:%.*]] = insertelement <2 x i16> poison, i16 [[CONV_I_S]], i64 0
 ; CHECK-NEXT:    [[CONV_I:%.*]] = shufflevector <2 x i16> [[CONV_I_V]], <2 x i16> poison, <2 x i32> zeroinitializer
 ; CHECK-NEXT:    [[MUL3_I_I:%.*]] = sitofp <2 x i16> [[CONV_I]] to <2 x float>
+; CHECK-NEXT:    [[MUL3_I_I1:%.*]] = fmul <2 x float> [[MUL3_I_I]], <float poison, float 1.000000e+00>
 ; CHECK-NEXT:    store i32 [[SEL]], ptr [[G_2345:%.*]], align 4
-; CHECK-NEXT:    ret <2 x float> [[MUL3_I_I]]
+; CHECK-NEXT:    ret <2 x float> [[MUL3_I_I1]]
 ;
   %sel = select i1 %c, i32 65529, i32 53264
   %conv.i.s = trunc i32 %sel to i16

``````````

</details>


https://github.com/llvm/llvm-project/pull/162804