https://llvm.org/bugs/show_bug.cgi?id=25787
Bug ID: 25787
Summary: SLPVectorizer finds incorrect reduction value
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Loop Optimizer
Assignee: unassignedb...@nondot.org
Reporter: mattias.v.eriks...@ericsson.com
CC: llvm-bugs@lists.llvm.org
Classification: Unclassified
Created attachment 15424
--> https://llvm.org/bugs/attachment.cgi?id=15424&action=edit
reduced input file
The function getReductionValue can return a Value that is not dominated by the
phi-node. This can cause miscompiles.
I found the problem originally in an out-of-tree target with a huge test case.
I have added an assert in the code like this:
--- a/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -3937,7 +3937,8 @@ static bool PhiTypeSorterFunc(Value *V, Value *V2) {
///
/// \returns A candidate reduction value if possible, or \code nullptr
\endcode
/// if not possible.
-static Value *getReductionValue(PHINode *P, BasicBlock *ParentBB,
+static Value *getReductionValue(const DominatorTree *DT,
+ PHINode *P, BasicBlock *ParentBB,
LoopInfo *LI) {
Value *Rdx = nullptr;
@@ -3967,6 +3968,12 @@ static Value *getReductionValue(PHINode *P, BasicBlock
*ParentBB,
Rdx = P->getIncomingValue(1);
}
+ assert(!Rdx || !dyn_cast<Instruction>(Rdx) ||
+ DT->dominates(P->getParent(),
dyn_cast<Instruction>(Rdx)->getParent()));
+ // if (Rdx && dyn_cast<Instruction>(Rdx) &&
+ // !DT->dominates(P->getParent(),
dyn_cast<Instruction>(Rdx)->getParent()))
+ // return nullptr;
+
return Rdx;
}
@@ -4062,7 +4069,7 @@ bool SLPVectorizer::vectorizeChainsInBlock(BasicBlock
*BB, BoUpSLP &R) {
if (P->getNumIncomingValues() != 2)
return Changed;
- Value *Rdx = getReductionValue(P, BB, LI);
+ Value *Rdx = getReductionValue(DT, P, BB, LI);
// Check if this is a Binary Operator.
BinaryOperator *BI = dyn_cast_or_null<BinaryOperator>(Rdx);
And then reduced the failing test to something that still trigger the assert
(but it no longer miscompiles).
Look at the first phi-node in the dump below. The incoming value from the loop
latch is actually defined in bb2. Is this a valid phi-node? If it is, then
there must be a bug when SLP Vectorizer finds the reduction value "%_tmp279 =
load volatile i16, i16* null".
*** IR Dump Before SLP Vectorizer ***
define i16 @test_spill_stress(i16 %tnr.8.par) #0 {
bb2:
%_tmp279 = load volatile i16, i16* null, align 32768
br label %bb_usw4.outer.outer
bb_usw4.outer.outer.loopexit: ; preds = %bb_usw4
br label %bb_usw4.outer.outer
bb_usw4.outer.outer: ; preds =
%bb_usw4.outer.outer.loopexit, %bb2
%f3.351.0.ph.ph = phi i16 [ 0, %bb2 ], [ %_tmp279,
%bb_usw4.outer.outer.loopexit ]
br label %bb_usw4
bb_usw4: ; preds = %bb_usw4.backedge,
%bb_usw4.outer.outer
%i__0.358.0 = phi i16 [ 4, %bb_usw4.outer.outer ], [ %i__0.358.0.be,
%bb_usw4.backedge ]
switch i16 %i__0.358.0, label %bb15 [
i16 3, label %bb_usw4.outer.outer.loopexit
i16 5, label %bb_usw4.backedge
]
bb15: ; preds = %bb_usw4
%_tmp1143 = add i16 %i__0.358.0, 1
%_tmp1145 = icmp slt i16 %_tmp1143, 8
br i1 %_tmp1145, label %bb_usw4.backedge, label %bb16
bb_usw4.backedge: ; preds = %bb15, %bb_usw4
%i__0.358.0.be = phi i16 [ %_tmp1143, %bb15 ], [ 6, %bb_usw4 ]
br label %bb_usw4
bb16: ; preds = %bb15
%_tmp1161 = icmp eq i16 %f3.351.0.ph.ph, 11584
%_tmp1162 = zext i1 %_tmp1161 to i16
%_tmp1163 = tail call i16 @check_i(i16 998, i16 %_tmp1162, i16 1)
ret i16 2
}
With the added assert, the following command crashes for the attached file:
opt -S -mcpu=swift -mtriple=thumbv7-apple-ios -slp-vectorizer -O3
bugpoint-reduced-simplified.ll
--
You are receiving this mail because:
You are on the CC list for the bug.
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
