[llvm] d0ea176 - [SLP]Do not consider SExt/ZExt profitable for demotion, if the user is a bitcast to float

[Alexey Bataev via llvm-commits]

Author: Alexey Bataev
Date: 2025-09-08T07:59:01-07:00
New Revision: d0ea176ccea8bdcd60aacc2753734294b9eec71f

URL: https://github.com/llvm/llvm-project/commit/d0ea176ccea8bdcd60aacc2753734294b9eec71f
DIFF: https://github.com/llvm/llvm-project/commit/d0ea176ccea8bdcd60aacc2753734294b9eec71f.diff

LOG: [SLP]Do not consider SExt/ZExt profitable for demotion, if the user is a bitcast to float

If the user node of the SExt/ZExt node is a bitcast to a float point
type, the node itself should not be considered legal to demote, since
still the casting is required to match the size of the float point type.

Fixes #157277

Added: 
    llvm/test/Transforms/SLPVectorizer/X86/parent-bitcast-with-fp.ll

Modified: 
    llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index 4e90c1d5b9e1f..1cfcd3ffbd664 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -21889,6 +21889,10 @@ bool BoUpSLP::collectValuesToDemote(
     return TryProcessInstruction(BitWidth);
   case Instruction::ZExt:
   case Instruction::SExt:
+    if (E.UserTreeIndex.UserTE && E.UserTreeIndex.UserTE->hasState() &&
+        E.UserTreeIndex.UserTE->getOpcode() == Instruction::BitCast &&
+        E.UserTreeIndex.UserTE->getMainOp()->getType()->isFPOrFPVectorTy())
+      return false;
     IsProfitableToDemote = true;
     return TryProcessInstruction(BitWidth);
 

diff  --git a/llvm/test/Transforms/SLPVectorizer/X86/parent-bitcast-with-fp.ll b/llvm/test/Transforms/SLPVectorizer/X86/parent-bitcast-with-fp.ll
new file mode 100644
index 0000000000000..a2f6b900d73cd
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/parent-bitcast-with-fp.ll
@@ -0,0 +1,36 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -S --passes=slp-vectorizer -mtriple=x86_64-unknown-linux-gnu < %s | FileCheck %s
+
+define i1 @test(i32 %0) {
+; CHECK-LABEL: define i1 @test(
+; CHECK-SAME: i32 [[TMP0:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[CONV22_I_I:%.*]] = sext i32 [[TMP0]] to i64
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i64 [[CONV22_I_I]] to double
+; CHECK-NEXT:    [[TMP2:%.*]] = fadd double [[TMP1]], 0.000000e+00
+; CHECK-NEXT:    [[ADD_I_I_I:%.*]] = select i1 false, double 0.000000e+00, double [[TMP2]]
+; CHECK-NEXT:    [[TMP3:%.*]] = bitcast double [[ADD_I_I_I]] to i64
+; CHECK-NEXT:    [[CMP3998_I_I:%.*]] = icmp ne i64 [[TMP3]], [[CONV22_I_I]]
+; CHECK-NEXT:    [[CONV22_1_I_I:%.*]] = sext i32 0 to i64
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast i64 [[CONV22_1_I_I]] to double
+; CHECK-NEXT:    [[TMP5:%.*]] = fadd double [[TMP4]], 0.000000e+00
+; CHECK-NEXT:    [[ADD_I_1_I_I:%.*]] = select i1 false, double 0.000000e+00, double [[TMP5]]
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast double [[ADD_I_1_I_I]] to i64
+; CHECK-NEXT:    [[CMP3998_1_I_I:%.*]] = icmp ne i64 [[TMP6]], [[CONV22_1_I_I]]
+; CHECK-NEXT:    ret i1 [[CMP3998_1_I_I]]
+;
+entry:
+  %conv22.i.i = sext i32 %0 to i64
+  %1 = bitcast i64 %conv22.i.i to double
+  %2 = fadd double %1, 0.000000e+00
+  %add.i.i.i = select i1 false, double 0.000000e+00, double %2
+  %3 = bitcast double %add.i.i.i to i64
+  %cmp3998.i.i = icmp ne i64 %3, %conv22.i.i
+  %conv22.1.i.i = sext i32 0 to i64
+  %4 = bitcast i64 %conv22.1.i.i to double
+  %5 = fadd double %4, 0.000000e+00
+  %add.i.1.i.i = select i1 false, double 0.000000e+00, double %5
+  %6 = bitcast double %add.i.1.i.i to i64
+  %cmp3998.1.i.i = icmp ne i64 %6, %conv22.1.i.i
+  ret i1 %cmp3998.1.i.i
+}