[NFC][SLP] Add test case for issue #69670.

[valerydmit]

Test exposes issue with delayed gather emission, which may lead to generating an instruction which does not dominate all users.

[alexey-bataev]

I think it is fixed already, no?

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Valery Dmitriev (valerydmit)

Changes

Test exposes issue with delayed gather emission, which may lead to generating an instruction which does not dominate all users.

---

Full diff: https://github.com/llvm/llvm-project/pull/70088.diff

1 Files Affected:

• (added) llvm/test/Transforms/SLPVectorizer/X86/delayed-gather-emission.ll (+59)

diff --git a/llvm/test/Transforms/SLPVectorizer/X86/delayed-gather-emission.ll b/llvm/test/Transforms/SLPVectorizer/X86/delayed-gather-emission.ll
new file mode 100644
index 000000000000000..8860cd4d8d141e3
--- /dev/null
+++ b/llvm/test/Transforms/SLPVectorizer/X86/delayed-gather-emission.ll
@@ -0,0 +1,59 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 3
+; RUN: opt -passes=slp-vectorizer -mtriple=x86_64 -mattr=+avx2 -S < %s | FileCheck %s
+
+; Test case for issue #69670.
+
+; #69392 uncovered a problem with delayed gather nodes emission, specifically
+; when the node is a PHI operand, which depends on another gathered node,
+; also an operand of a PHI (in another block).
+
+define void @test() {
+; CHECK-LABEL: define void @test(
+; CHECK-SAME: ) #[[ATTR0:[0-9]+]] {
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[DOTPRE_PRE:%.*]] = load float, ptr poison, align 4
+; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <2 x float> <float poison, float undef>, float [[DOTPRE_PRE]], i32 0
+; CHECK-NEXT:    br label [[BB1:%.*]]
+; CHECK:       bb1:
+; CHECK-NEXT:    [[TMP1:%.*]] = phi <2 x float> [ [[TMP0]], [[ENTRY:%.*]] ], [ [[TMP10:%.*]], [[BB2:%.*]] ]
+; CHECK-NEXT:    br label [[BB2]]
+; CHECK:       bb2:
+; CHECK-NEXT:    [[TMP2:%.*]] = phi <2 x float> [ [[TMP1]], [[BB1]] ], [ [[TMP9:%.*]], [[BB2]] ]
+; CHECK-NEXT:    [[I1:%.*]] = load float, ptr poison, align 4
+; CHECK-NEXT:    [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> <i32 1, i32 poison>
+; CHECK-NEXT:    [[TMP4:%.*]] = insertelement <2 x float> [[TMP3]], float [[I1]], i32 1
+; CHECK-NEXT:    [[TMP5:%.*]] = fdiv <2 x float> [[TMP2]], [[TMP4]]
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <2 x float> [[TMP5]], i32 0
+; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <2 x float> [[TMP5]], i32 1
+; CHECK-NEXT:    [[MUL:%.*]] = fmul float [[TMP6]], [[TMP7]]
+; CHECK-NEXT:    tail call void @foo(float [[MUL]])
+; CHECK-NEXT:    [[I2:%.*]] = load float, ptr poison, align 4
+; CHECK-NEXT:    [[TOBOOL:%.*]] = fcmp une float [[I2]], 0.000000e+00
+; CHECK-NEXT:    [[TMP8:%.*]] = insertelement <2 x float> poison, float [[I2]], i32 0
+; CHECK-NEXT:    [[TMP9]] = shufflevector <2 x float> [[TMP8]], <2 x float> [[TMP5]], <2 x i32> <i32 0, i32 2>
+; CHECK-NEXT:    [[TMP10]] = shufflevector <2 x float> [[TMP8]], <2 x float> [[TMP2]], <2 x i32> <i32 0, i32 3>
+; CHECK-NEXT:    br i1 [[TOBOOL]], label [[BB1]], label [[BB2]]
+;
+entry:
+  %.pre.pre = load float, ptr poison, align 4
+  br label %bb1
+
+bb1:                                              ; preds = %bb2, %entry
+  %.pre = phi float [ %.pre.pre, %entry ], [ %i2, %bb2 ]
+  %foxtrot.0 = phi float [ undef, %entry ], [ %gulf.0, %bb2 ]
+  br label %bb2
+
+bb2:                                              ; preds = %bb2, %bb1
+  %i = phi float [ %.pre, %bb1 ], [ %i2, %bb2 ]
+  %gulf.0 = phi float [ %foxtrot.0, %bb1 ], [ %div, %bb2 ]
+  %div = fdiv float %i, %gulf.0
+  %i1 = load float, ptr poison, align 4
+  %div1 = fdiv float %gulf.0, %i1
+  %mul = fmul float %div, %div1
+  tail call void @foo(float %mul)
+  %i2 = load float, ptr poison, align 4
+  %tobool = fcmp une float %i2, 0.000000e+00
+  br i1 %tobool, label %bb1, label %bb2
+}
+
+declare void @foo(float)

[alexey-bataev]

LG

[valerydmit]

I think it is fixed already, no?
I have a fix. The crash is only reproducible with #69392 (which was reverted). Will publish it soon along with the re-commit.