[llvm] r347934 - [SCEV] Guard movement of insertion point for loop-invariants
via llvm-commits
llvm-commits at lists.llvm.org
Wed Dec 5 15:59:45 PST 2018
> OK, in that case I will keep working on a reduced test case. Once I have something I can share, I will forward to you.
>
> Sorry about the revert, and thanks for your help.
Sorry for the trouble in the first place. Thanks for looking further! It’s very much appreciated.
-Warren
From: David Jones <dlj at google.com>
Sent: Wednesday, December 5, 2018 3:17 PM
To: Ristow, Warren <warren.ristow at sony.com>
Cc: llvm-commits at lists.llvm.org; Reid Kleckner <rnk at google.com>
Subject: Re: [llvm] r347934 - [SCEV] Guard movement of insertion point for loop-invariants
On Wed, Dec 5, 2018 at 2:56 PM <warren.ristow at sony.com<mailto:warren.ristow at sony.com>> wrote:
Hi David,
[+ Reid]
> ...
> I suspect your change is exposing a pre-existing bug; but unfortunately, it does very
> cleanly resolve to exactly r347934, and it occurs with multiple targets.
>
> Could you advise on a fix? Unfortunately, I'm not familiar with this code, so the best
> option I can offer is a revert. (If you won't able to author a timely fix, please let
> me know so I can submit the revert for you.)
Sorry to hear about that. I definitely won't be able to author a timely fix, so please do go ahead and revert it.
OK... sorry for bringing out the big hammer. :-/
FYI: Committed revision 348426.
As it happens, while looking at the history of this SCEV div-by-0 area, in prep for a follow-up patch:
https://reviews.llvm.org/D55232
I noticed that an equivalent change was made about a year ago (r289412). Reid found it caused a problem in instcombine when ASan was enabled. This issue you reported sounds very similar. Reid reverted that year-old commit at r289453:
‘’’
Reverts r289412. It caused an OOB PHI operand access in instcombine when
ASan is enabled. Reduction in progress.
‘’’
> I don't feel I have a useful reproducer to share... between creduce and bugpoint, I've
> gotten one target down to 68k of .bc (which disassembles to about 180k of .ll). (And it
> would take a lot of effort to redact before I can share it.)
Possibly Reid has access to a reduced test-case from back then (or can create one now)?
I'm not familiar with the code where the instcombine failure is happening. Without a test-case, I feel my hands are tied. So a reduced test-case from either you or Reid would be greatly appreciated!
OK, in that case I will keep working on a reduced test case. Once I have something I can share, I will forward to you.
Sorry about the revert, and thanks for your help.
--dlj
Thanks,
-Warren
From: David Jones <dlj at google.com<mailto:dlj at google.com>>
Sent: Wednesday, December 5, 2018 2:25 PM
To: Ristow, Warren <warren.ristow at sony.com<mailto:warren.ristow at sony.com>>
Cc: llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>
Subject: Re: [llvm] r347934 - [SCEV] Guard movement of insertion point for loop-invariants
Hi Warren --
I've observed several Clang crashes on SEGV, which bisect to this change.
The crashes generally look something like this:
$ opt -O2 -o /dev/null bugpoint-reduced-simplified.ll
Stack dump:
0. Program arguments: opt -O2 -o /dev/null bugpoint-reduced-simplified.ll
1. Running pass 'CallGraph Pass Manager' on module 'bugpoint-reduced-simplified.ll'.
2. Running pass 'Combine redundant instructions' on function '@_Z[redacted]'
#0 0x0000558aa2221628 llvm::sys::RunSignalHandlers() (opt+0x1f3b628)
#1 0x0000558aa2223746 SignalHandler(int) (opt+0x1f3d746)
#2 0x00007f5def60a9a0 __restore_rt (/usr/grte/v4/lib64/libpthread.so.0+0xf9a0)
#3 0x0000558aa1bbec7f llvm::InstCombiner::visitPHINode(llvm::PHINode&) (opt+0x18d8c7f)
#4 0x0000558aa1b8db24 llvm::InstCombiner::run() (opt+0x18a7b24)
#5 0x0000558aa1b8f504 combineInstructionsOverFunction(llvm::Function&, llvm::InstCombineWorklist&, llvm::AAResults*, llvm::AssumptionCache&, llvm::TargetLibraryInfo&, llvm::DominatorTree&, llvm::OptimizationRemarkEmitter&, bool, llvm::LoopInfo*) (opt+0x18a9504)
#6 0x0000558aa1b8f8af llvm::InstructionCombiningPass::runOnFunction(llvm::Function&) (opt+0x18a98af)
#7 0x0000558aa207d2fc llvm::FPPassManager::runOnFunction(llvm::Function&) (opt+0x1d972fc)
#8 0x0000558aa1f062e0 (anonymous namespace)::CGPassManager::runOnModule(llvm::Module&) (opt+0x1c202e0)
#9 0x0000558aa207da85 llvm::legacy::PassManagerImpl::run(llvm::Module&) (opt+0x1d97a85)
#10 0x0000558aa0a8900b main (opt+0x7a300b)
#11 0x00007f5def378bbd __libc_start_main (/usr/grte/v4/lib64/libc.so.6+0x38bbd)
#12 0x0000558aa0a740e9 _start /usr/grte/v4/debug-src/src/csu/../sysdeps/x86_64/start.S:121:0
[1] 17713 segmentation fault opt -O2 -o /dev/null bugpoint-reduced-simplified.ll
gdb's backtrace is a bit more useful:
#0 0x000055555958b1d4 in llvm::Use::get (this=<optimized out>) at third_party/llvm/llvm/include/llvm/IR/Use.h:108
#1 llvm::PHINode::getOperand (this=<optimized out>, i_nocapture=4294967295) at third_party/llvm/llvm/include/llvm/IR/Instructions.h:2714
#2 llvm::PHINode::getIncomingValue (this=<optimized out>, i=4294967295) at third_party/llvm/llvm/include/llvm/IR/Instructions.h:2603
#3 0x000055555c41abba in llvm::InstCombiner::visitPHINode (this=0x7ffff4e3b970, PN=...) at third_party/llvm/llvm/lib/Transforms/InstCombine/InstCombinePHI.cpp:1238
#4 0x000055555c378489 in llvm::InstCombiner::run (this=<optimized out>) at third_party/llvm/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp:3192
#5 0x000055555c37b6d2 in combineInstructionsOverFunction (F=..., Worklist=..., AA=0x60600002b7c0, AC=..., TLI=..., DT=..., ORE=..., ExpensiveCombines=<optimized out>, LI=<optimized out>)
at third_party/llvm/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp:3424
#6 0x000055555c37be37 in llvm::InstructionCombiningPass::runOnFunction (this=<optimized out>, F=...) at third_party/llvm/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp:3483
#7 0x000055555d39b609 in llvm::FPPassManager::runOnFunction (this=<optimized out>, F=...) at third_party/llvm/llvm/lib/IR/LegacyPassManager.cpp:1644
#8 0x000055555ce8f11b in (anonymous namespace)::CGPassManager::RunPassOnSCC (this=<optimized out>, P=0x614000009a40, CurSCC=..., CG=..., CallGraphUpToDate=<optimized out>, DevirtualizedCall=<optimized out>)
at third_party/llvm/llvm/lib/Analysis/CallGraphSCCPass.cpp:178
#9 (anonymous namespace)::CGPassManager::RunAllPassesOnSCC (this=<optimized out>, CurSCC=..., CG=..., DevirtualizedCall=<optimized out>) at third_party/llvm/llvm/lib/Analysis/CallGraphSCCPass.cpp:442
#10 (anonymous namespace)::CGPassManager::runOnModule (this=0x614000009840, M=...) at third_party/llvm/llvm/lib/Analysis/CallGraphSCCPass.cpp:498
#11 0x000055555d39ca7a in (anonymous namespace)::MPPassManager::runOnModule (this=<optimized out>, M=...) at third_party/llvm/llvm/lib/IR/LegacyPassManager.cpp:1744
#12 llvm::legacy::PassManagerImpl::run (this=0x619000020f50, M=...) at third_party/llvm/llvm/lib/IR/LegacyPassManager.cpp:1857
#13 0x00005555588bfb58 in main (argc=<optimized out>, argv=<optimized out>) at third_party/llvm/llvm/tools/opt/opt.cpp:828
I've spent the last 48 hours attempting to reduce a test case (both mechanically and manually). Unfortunately, I don't feel I have a useful reproducer to share... between creduce and bugpoint, I've gotten one target down to 68k of .bc (which disassembles to about 180k of .ll). (And it would take a lot of effort to redact before I can share it.)
The stack trace pretty clearly points to an unchecked out-of-range return value here:
https://git.llvm.org/klaus/llvm/blob/master/lib/Transforms/InstCombine/InstCombinePHI.cpp#L-1238
I suspect your change is exposing a pre-existing bug; but unfortunately, it does very cleanly resolve to exactly r347934, and it occurs with multiple targets.
Could you advise on a fix? Unfortunately, I'm not familiar with this code, so the best option I can offer is a revert. (If you won't able to author a timely fix, please let me know so I can submit the revert for you.)
--David Jones
On Thu, Nov 29, 2018 at 4:05 PM Warren Ristow via llvm-commits <llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>> wrote:
Author: wristow
Date: Thu Nov 29 16:02:54 2018
New Revision: 347934
URL: http://llvm.org/viewvc/llvm-project?rev=347934&view=rev
Log:
[SCEV] Guard movement of insertion point for loop-invariants
r320789 suppressed moving the insertion point of SCEV expressions with
dev/rem operations to the loop header in non-loop-invariant situations.
This, and similar, hoisting is also unsafe in the loop-invariant case,
since there may be a guard against a zero denominator. This is an
adjustment to the fix of r320789 to suppress the movement even in the
loop-invariant case.
This fixes PR30806.
Differential Revision: https://reviews.llvm.org/D54713
Added:
llvm/trunk/test/Transforms/LoopVectorize/pr30806.ll
Modified:
llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp
Modified: llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp?rev=347934&r1=347933&r2=347934&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp (original)
+++ llvm/trunk/lib/Analysis/ScalarEvolutionExpander.cpp Thu Nov 29 16:02:54 2018
@@ -1732,49 +1732,50 @@ Value *SCEVExpander::expand(const SCEV *
// Compute an insertion point for this SCEV object. Hoist the instructions
// as far out in the loop nest as possible.
Instruction *InsertPt = &*Builder.GetInsertPoint();
- for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());;
- L = L->getParentLoop())
- if (SE.isLoopInvariant(S, L)) {
- if (!L) break;
- if (BasicBlock *Preheader = L->getLoopPreheader())
- InsertPt = Preheader->getTerminator();
- else {
- // LSR sets the insertion point for AddRec start/step values to the
- // block start to simplify value reuse, even though it's an invalid
- // position. SCEVExpander must correct for this in all cases.
- InsertPt = &*L->getHeader()->getFirstInsertionPt();
- }
- } else {
- // We can move insertion point only if there is no div or rem operations
- // otherwise we are risky to move it over the check for zero denominator.
- auto SafeToHoist = [](const SCEV *S) {
- return !SCEVExprContains(S, [](const SCEV *S) {
- if (const auto *D = dyn_cast<SCEVUDivExpr>(S)) {
- if (const auto *SC = dyn_cast<SCEVConstant>(D->getRHS()))
- // Division by non-zero constants can be hoisted.
- return SC->getValue()->isZero();
- // All other divisions should not be moved as they may be
- // divisions by zero and should be kept within the
- // conditions of the surrounding loops that guard their
- // execution (see PR35406).
- return true;
- }
- return false;
- });
- };
- // If the SCEV is computable at this level, insert it into the header
- // after the PHIs (and after any other instructions that we've inserted
- // there) so that it is guaranteed to dominate any user inside the loop.
- if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L) &&
- SafeToHoist(S))
- InsertPt = &*L->getHeader()->getFirstInsertionPt();
- while (InsertPt->getIterator() != Builder.GetInsertPoint() &&
- (isInsertedInstruction(InsertPt) ||
- isa<DbgInfoIntrinsic>(InsertPt))) {
- InsertPt = &*std::next(InsertPt->getIterator());
+
+ // We can move insertion point only if there is no div or rem operations
+ // otherwise we are risky to move it over the check for zero denominator.
+ auto SafeToHoist = [](const SCEV *S) {
+ return !SCEVExprContains(S, [](const SCEV *S) {
+ if (const auto *D = dyn_cast<SCEVUDivExpr>(S)) {
+ if (const auto *SC = dyn_cast<SCEVConstant>(D->getRHS()))
+ // Division by non-zero constants can be hoisted.
+ return SC->getValue()->isZero();
+ // All other divisions should not be moved as they may be
+ // divisions by zero and should be kept within the
+ // conditions of the surrounding loops that guard their
+ // execution (see PR35406).
+ return true;
+ }
+ return false;
+ });
+ };
+ if (SafeToHoist(S)) {
+ for (Loop *L = SE.LI.getLoopFor(Builder.GetInsertBlock());;
+ L = L->getParentLoop()) {
+ if (SE.isLoopInvariant(S, L)) {
+ if (!L) break;
+ if (BasicBlock *Preheader = L->getLoopPreheader())
+ InsertPt = Preheader->getTerminator();
+ else
+ // LSR sets the insertion point for AddRec start/step values to the
+ // block start to simplify value reuse, even though it's an invalid
+ // position. SCEVExpander must correct for this in all cases.
+ InsertPt = &*L->getHeader()->getFirstInsertionPt();
+ } else {
+ // If the SCEV is computable at this level, insert it into the header
+ // after the PHIs (and after any other instructions that we've inserted
+ // there) so that it is guaranteed to dominate any user inside the loop.
+ if (L && SE.hasComputableLoopEvolution(S, L) && !PostIncLoops.count(L))
+ InsertPt = &*L->getHeader()->getFirstInsertionPt();
+ while (InsertPt->getIterator() != Builder.GetInsertPoint() &&
+ (isInsertedInstruction(InsertPt) ||
+ isa<DbgInfoIntrinsic>(InsertPt)))
+ InsertPt = &*std::next(InsertPt->getIterator());
+ break;
}
- break;
}
+ }
// Check to see if we already expanded this here.
auto I = InsertedExpressions.find(std::make_pair(S, InsertPt));
Added: llvm/trunk/test/Transforms/LoopVectorize/pr30806.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopVectorize/pr30806.ll?rev=347934&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/LoopVectorize/pr30806.ll (added)
+++ llvm/trunk/test/Transforms/LoopVectorize/pr30806.ll Thu Nov 29 16:02:54 2018
@@ -0,0 +1,65 @@
+; RUN: opt -loop-vectorize -S < %s 2>&1 | FileCheck %s
+
+; Produced from test-case:
+;
+; void testGuardedInnerLoop(uint32_t *ptr, uint32_t denom, uint32_t numer, uint32_t outer_lim)
+; {
+; for(uint32_t outer_i = 0; outer_i < outer_lim; ++outer_i) {
+; if (denom > 0) {
+; const uint32_t lim = numer / denom;
+;
+; for (uint32_t i = 0; i < lim; ++i)
+; ptr[i] = 1;
+; }
+; }
+; }
+
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128-ni:1"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @testGuardedInnerLoop(i32* %ptr, i32 %denom, i32 %numer, i32 %outer_lim) {
+entry:
+ %cmp1 = icmp eq i32 %outer_lim, 0
+ br i1 %cmp1, label %exit, label %loop1.preheader
+
+; Verify that a 'udiv' does not appear between the 'loop1.preheader' label, and
+; whatever label comes next.
+loop1.preheader:
+; CHECK-LABEL: loop1.preheader:
+; CHECK-NOT: udiv
+; CHECK-LABEL: :
+ br label %loop1
+
+loop1:
+ %outer_i = phi i32 [ %inc1, %loop2.exit ], [ 0, %loop1.preheader ]
+ %0 = add i32 %denom, -1
+ %1 = icmp ult i32 %0, %numer
+ br i1 %1, label %loop2.preheader, label %loop2.exit
+
+; Verify that a 'udiv' does appear between the 'loop2.preheader' label, and
+; whatever label comes next.
+loop2.preheader:
+; CHECK-LABEL: loop2.preheader:
+; CHECK: udiv
+; CHECK-LABEL: :
+ %lim = udiv i32 %numer, %denom
+ %2 = zext i32 %lim to i64
+ br label %loop2
+
+loop2:
+ %indvar.loop2 = phi i64 [ 0, %loop2.preheader ], [ %indvar.loop2.next, %loop2 ]
+ %arrayidx = getelementptr inbounds i32, i32* %ptr, i64 %indvar.loop2
+ store i32 1, i32* %arrayidx, align 4
+ %indvar.loop2.next = add nuw nsw i64 %indvar.loop2, 1
+ %cmp2 = icmp ult i64 %indvar.loop2.next, %2
+ br i1 %cmp2, label %loop2, label %loop2.exit
+
+loop2.exit:
+ %inc1 = add nuw i32 %outer_i, 1
+ %exitcond = icmp eq i32 %inc1, %outer_lim
+ br i1 %exitcond, label %exit, label %loop1
+
+exit:
+ ret void
+}
_______________________________________________
llvm-commits mailing list
llvm-commits at lists.llvm.org<mailto:llvm-commits at lists.llvm.org>
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20181205/f792daeb/attachment.html>
More information about the llvm-commits
mailing list