[llvm-branch-commits] [llvm] 373180b - [SLP]Fix PR104422: Wrong value truncation

[Tobias Hieta via llvm-branch-commits]

Author: Alexey Bataev
Date: 2024-09-13T07:58:38+02:00
New Revision: 373180b440d04dc3cc0f6111b06684d18779d7c8

URL: https://github.com/llvm/llvm-project/commit/373180b440d04dc3cc0f6111b06684d18779d7c8
DIFF: https://github.com/llvm/llvm-project/commit/373180b440d04dc3cc0f6111b06684d18779d7c8.diff

LOG: [SLP]Fix PR104422: Wrong value truncation

The minbitwidth restrictions can be skipped only for immediate reduced
values, for other nodes still need to check if external users allow
bitwidth reduction.

Fixes https://github.com/llvm/llvm-project/issues/104422

(cherry picked from commit 56140a8258a3498cfcd9f0f05c182457d43cbfd2)

Added: 
    llvm/test/Transforms/SLPVectorizer/X86/operand-is-reduced-val.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 2f3d6b27378aee..ab2b96cdc42db8 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -15211,7 +15211,8 @@ bool BoUpSLP::collectValuesToDemote(
   if (any_of(E.Scalars, [&](Value *V) {
         return !all_of(V->users(), [=](User *U) {
           return getTreeEntry(U) ||
-                 (UserIgnoreList && UserIgnoreList->contains(U)) ||
+                 (E.Idx == 0 && UserIgnoreList &&
+                  UserIgnoreList->contains(U)) ||
                  (!isa<CmpInst>(U) && U->getType()->isSized() &&
                   !U->getType()->isScalableTy() &&
                   DL->getTypeSizeInBits(U->getType()) <= BitWidth);

diff  --git a/llvm/test/Transforms/SLPVectorizer/X86/operand-is-reduced-val.ll b/llvm/test/Transforms/SLPVectorizer/X86/operand-is-reduced-val.ll
new file mode 100644
index 00000000000000..5fcac3fbf3bafe
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/operand-is-reduced-val.ll
@@ -0,0 +1,49 @@
+; 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 < %s -slp-threshold=-10 | FileCheck %s
+
+define i64 @src(i32 %a) {
+; CHECK-LABEL: define i64 @src(
+; CHECK-SAME: i32 [[A:%.*]]) {
+; CHECK-NEXT:  [[ENTRY:.*:]]
+; CHECK-NEXT:    [[TMP17:%.*]] = sext i32 [[A]] to i64
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <4 x i32> poison, i32 [[A]], i32 0
+; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x i32> [[TMP1]], <4 x i32> poison, <4 x i32> zeroinitializer
+; CHECK-NEXT:    [[TMP3:%.*]] = sext <4 x i32> [[TMP2]] to <4 x i64>
+; CHECK-NEXT:    [[TMP4:%.*]] = add nsw <4 x i64> [[TMP3]], <i64 4294967297, i64 4294967297, i64 4294967297, i64 4294967297>
+; CHECK-NEXT:    [[TMP6:%.*]] = and <4 x i64> [[TMP4]], <i64 1, i64 1, i64 1, i64 1>
+; CHECK-NEXT:    [[TMP18:%.*]] = call i64 @llvm.vector.reduce.add.v4i64(<4 x i64> [[TMP6]])
+; CHECK-NEXT:    [[TMP16:%.*]] = call i64 @llvm.vector.reduce.add.v4i64(<4 x i64> [[TMP4]])
+; CHECK-NEXT:    [[TMP8:%.*]] = insertelement <2 x i64> poison, i64 [[TMP16]], i32 0
+; CHECK-NEXT:    [[TMP9:%.*]] = insertelement <2 x i64> [[TMP8]], i64 [[TMP18]], i32 1
+; CHECK-NEXT:    [[TMP10:%.*]] = insertelement <2 x i64> <i64 poison, i64 4294967297>, i64 [[TMP17]], i32 0
+; CHECK-NEXT:    [[TMP11:%.*]] = add <2 x i64> [[TMP9]], [[TMP10]]
+; CHECK-NEXT:    [[TMP12:%.*]] = extractelement <2 x i64> [[TMP11]], i32 0
+; CHECK-NEXT:    [[TMP13:%.*]] = extractelement <2 x i64> [[TMP11]], i32 1
+; CHECK-NEXT:    [[TMP21:%.*]] = add i64 [[TMP12]], [[TMP13]]
+; CHECK-NEXT:    ret i64 [[TMP21]]
+;
+entry:
+  %0 = sext i32 %a to i64
+  %1 = add nsw i64 %0, 4294967297
+  %2 = sext i32 %a to i64
+  %3 = add nsw i64 %2, 4294967297
+  %4 = add i64 %3, %1
+  %5 = and i64 %3, 1
+  %6 = add i64 %4, %5
+  %7 = sext i32 %a to i64
+  %8 = add nsw i64 %7, 4294967297
+  %9 = add i64 %8, %6
+  %10 = and i64 %8, 1
+  %11 = add i64 %9, %10
+  %12 = sext i32 %a to i64
+  %13 = add nsw i64 %12, 4294967297
+  %14 = add i64 %13, %11
+  %15 = and i64 %13, 1
+  %16 = add i64 %14, %15
+  %17 = sext i32 %a to i64
+  %18 = add nsw i64 %17, 4294967297
+  %19 = add i64 %18, %16
+  %20 = and i64 %18, 1
+  %21 = add i64 %19, %20
+  ret i64 %21
+}