[llvm] 8d20f2c - Revert "[CodeGenPrepare] Fix isIVIncrement (PR49466)"
Philip Reames via llvm-commits
llvm-commits at lists.llvm.org
Fri Mar 12 15:22:31 PST 2021
On 3/12/21 2:59 PM, Jordan Rupprecht wrote:
>
>
> On Fri, Mar 12, 2021 at 4:29 PM Philip Reames
> <listmail at philipreames.com <mailto:listmail at philipreames.com>> wrote:
>
> Jordan,
>
> Please revert this change.
>
> There's a couple of problems here:
>
> * The change reverted looks obviously innocent. (e.g. it's
> bailing out of a transform slightly more often)
>
> I don't disagree that it "looks obviously innocent", but the proof is
> in the pudding: the reproducer I posted compiles ~instantly at one
> commit prior, and times out at the culprit commit. A change "looking"
> good should never be a basis for saying it must be correct and should
> not be reverted, especially when there is evidence it is a problem. At
> least, that's my personal opinion, but I should think that's a fairly
> basic and widely held belief.
You're correct, but missing the point I was getting at. Admittedly,
poorly worded.
While looking obviously innocent isn't reason not to revert itself, it's
definitely reason to take a slightly closer look and make sure there's
nothing else going on. This is particularly relevant given the other
points in this discussion.
> * The change reverted was itself fixing a functional bug. At a
> minimum, we'd need a larger revert to get ToT back to a sane
> state.
>
> The fix is introducing a different bug, and one which seems more
> widespread. At the very least, we haven't observed any of the crashes
> mentioned in PR49466, but we did notice a compile timeout in several
> different compilation units.
Bug? Discussion? Response to original commit?
Your observation may be true *for you*. It is not necessarily true for
anyone else, and you bear the burden of making the case. Particularly
when reverting a functional fix.
To be clear, I'm not saying you're wrong. I'm saying you haven't given
anyone else enough information to tell if you are or not.
> * It is generally considered reasonable to provide a test case
> and wait a bit before reverting, at least once the patch is
> more than a few hours old.
>
> Sorry about that. I'll concede that I got a little trigger happy here,
> but I was hoping that would be waived by the fact that I gave a
> simple, concrete reproducer.
Thanks for the acknowledgement. To be fair, it would have been less of
an issue if the reproducer worked straight forwardly.
>
> * Failure to provide an IR test case.
>
> (ditto, but see one below)
>
> * Your test case does not reproduce. Or at least, it doesn't
> reproduce when compiled with clang10 to IR and then run
> through (very recent, but without your change) ToT opt -O2.
> If there's something specific about the interaction of clang
> and opt ToT, reducing this down to a IR test case becomes
> particularly important.
>
> This comment is going *way* off track -- the reproducer I posted
> *does* reproduce, at least for me, in the configuration I posted (a
> C++ source file, and just "clang -O2"). By saying it doesn't reproduce
> in a mixed configuration of an old version of clang to do the C++ ->
> IR combined with a ToT version of opt -O2 to do the IR -> object file
> is misleading -- it's true, but that's not at all what I was claiming,
> and I don't know where it's coming from.
I actually don't think this is going off track at all. Our default is
for IR test cases for IR problems. Any reverting commit should
generally include a test case suitable for checkin (w/a bit of cleanup)
once the patch is fixed. I'll admit, we're not strict about this, but
the expectation is definitely there.
The discussion of the hybrid configuration is relevant *because* you
didn't provide a test case in a form I could easily use. I don't work
on clang, don't build it regularly, and shouldn't have to investigate a
LLVM codegen regression.
(See also below.)
> Philip
>
> p.s. To be clear, I'm happy to look at your original issue this
> afternoon if you've got something I can reproduce.
>
> Here's an IR reproducer with IR generated from ToT before my revert
> (at dfd27ebbd0eb137c9a439b7c537bb87ba903efd3):
Ok, we have a problem here. None of the following reproduce for me:
$ ./llc -O2 jordan.ll
$ ./llc -O2 jordan.ll
$ ./llc -O2 jordan.ll --filetype=obj
$ ./llc -O3 jordan.ll --filetype=obj
LLC is ToT, just built with your change reverted locally.
> $ bin/clang -c /tmp/repro.cc -O1 -S -emit-llvm -o /tmp/repro.ll
> $ bin/clang -c /tmp/repro.ll -O2 -o /tmp/repro.o # hangs
Given the preceding, I am now asserting this is likely some clang
specific problem. I've got a build of clang running now, will report
back in a bit.
The other option is that this is someway specific to your configuration.
>
> ; ModuleID = '/tmp/repro.cc'
> source_filename = "/tmp/repro.cc"
> target datalayout =
> "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
> target triple = "x86_64-unknown-linux-gnu"
>
> %class.D = type { i64 }
> %class.a = type { %class.g }
> %class.g = type { i32*, i32* }
>
> $_ZNK1aIliEixEl = comdat any
>
> $_ZNK1aIliE1jEv = comdat any
>
> @o = dso_local local_unnamed_addr global i32 0, align 4
> @p = dso_local local_unnamed_addr global i32 0, align 4
> @.str = private unnamed_addr constant [1 x i8] zeroinitializer, align 1
>
> ; Function Attrs: uwtable mustprogress
> define dso_local void @_ZN1D1qERK1aIliE(%class.D* nocapture nonnull
> dereferenceable(8) %0, %class.a* nonnull align 8 dereferenceable(16)
> %1) local_unnamed_addr #0 align 2 {
> %3 = call i64 @_ZNK1aIliEixEl(%class.a* nonnull dereferenceable(16)
> %1, i64 0) #3
> %4 = icmp eq i64 %3, 0
> br i1 %4, label %26, label %5
>
> 5: ; preds = %2
> %6 = call i64 @_ZNK1aIliE1jEv(%class.a* nonnull dereferenceable(16) %1)
> %7 = icmp eq i64 %6, 0
> br i1 %7, label %26, label %8
>
> 8: ; preds = %5
> %9 = getelementptr inbounds %class.D, %class.D* %0, i64 0, i32 0
> br label %10
>
> 10: ; preds = %8, %18
> %11 = phi i64 [ 0, %8 ], [ %23, %18 ]
> %12 = call i64 @_ZNK1aIliEixEl(%class.a* nonnull dereferenceable(16)
> %1, i64 %11) #3
> %13 = icmp eq i64 %12, 0
> br i1 %13, label %18, label %14
>
> 14: ; preds = %10, %14
> %15 = load i32, i32* @o, align 4, !tbaa !2
> %16 = call i32 @_Z3fn1IiiEiT_T0_PKc(i32 %15, i32 0, i8*
> getelementptr inbounds ([1 x i8], [1 x i8]* @.str, i64 0, i64 0))
> %17 = icmp eq i32 %16, 0
> br i1 %17, label %18, label %14, !llvm.loop !6
>
> 18: ; preds = %14, %10
> %19 = call i64 @_ZNK1aIliEixEl(%class.a* nonnull dereferenceable(16)
> %1, i64 %11) #3
> %20 = load i32, i32* @p, align 4, !tbaa !2
> %21 = sext i32 %20 to i64
> %22 = sdiv i64 %19, %21
> store i64 %22, i64* %9, align 8, !tbaa !9
> %23 = add i64 %11, 1
> %24 = call i64 @_ZNK1aIliE1jEv(%class.a* nonnull dereferenceable(16) %1)
> %25 = icmp eq i64 %24, 0
> br i1 %25, label %26, label %10, !llvm.loop !12
>
> 26: ; preds = %18, %5, %2
> ret void
> }
>
> ; Function Attrs: nounwind uwtable willreturn mustprogress
> define linkonce_odr dso_local i64 @_ZNK1aIliEixEl(%class.a* nonnull
> dereferenceable(16) %0, i64 %1) local_unnamed_addr #1 comdat align 2 {
> %3 = getelementptr inbounds %class.a, %class.a* %0, i64 0, i32 0, i32 0
> %4 = load i32*, i32** %3, align 8, !tbaa !13
> %5 = getelementptr inbounds i32, i32* %4, i64 %1
> %6 = load i32, i32* %5, align 4, !tbaa !2
> %7 = sext i32 %6 to i64
> ret i64 %7
> }
>
> ; Function Attrs: nounwind uwtable willreturn mustprogress
> define linkonce_odr dso_local i64 @_ZNK1aIliE1jEv(%class.a* nonnull
> dereferenceable(16) %0) local_unnamed_addr #1 comdat align 2 {
> %2 = getelementptr inbounds %class.a, %class.a* %0, i64 0, i32 0, i32 1
> %3 = load i32*, i32** %2, align 8, !tbaa !16
> %4 = getelementptr inbounds %class.a, %class.a* %0, i64 0, i32 0, i32 0
> %5 = load i32*, i32** %4, align 8, !tbaa !13
> %6 = ptrtoint i32* %3 to i64
> %7 = ptrtoint i32* %5 to i64
> %8 = sub i64 %6, %7
> %9 = ashr exact i64 %8, 2
> ret i64 %9
> }
>
> declare dso_local i32 @_Z3fn1IiiEiT_T0_PKc(i32, i32, i8*)
> local_unnamed_addr #2
>
> attributes #0 = { uwtable mustprogress "frame-pointer"="none"
> "no-trapping-math"="true" "stack-protector-buffer-size"="8"
> "target-cpu"="x86-64"
> "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
> attributes #1 = { nounwind uwtable willreturn mustprogress
> "frame-pointer"="none" "no-trapping-math"="true"
> "stack-protector-buffer-size"="8" "target-cpu"="x86-64"
> "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
> attributes #2 = { "frame-pointer"="none" "no-trapping-math"="true"
> "stack-protector-buffer-size"="8" "target-cpu"="x86-64"
> "target-features"="+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" }
> attributes #3 = { nounwind }
>
> !llvm.module.flags = !{!0}
> !llvm.ident = !{!1}
>
> !0 = !{i32 1, !"wchar_size", i32 4}
> !1 = !{!"clang version 13.0.0
> (https://github.com/llvm/llvm-project.git
> <https://github.com/llvm/llvm-project.git>
> dfd27ebbd0eb137c9a439b7c537bb87ba903efd3)"}
> !2 = !{!3, !3, i64 0}
> !3 = !{!"int", !4, i64 0}
> !4 = !{!"omnipotent char", !5, i64 0}
> !5 = !{!"Simple C++ TBAA"}
> !6 = distinct !{!6, !7, !8}
> !7 = !{!"llvm.loop.mustprogress"}
> !8 = !{!"llvm.loop.unroll.disable"}
> !9 = !{!10, !11, i64 0}
> !10 = !{!"_ZTS1D", !11, i64 0}
> !11 = !{!"long", !4, i64 0}
> !12 = distinct !{!12, !7, !8}
> !13 = !{!14, !15, i64 0}
> !14 = !{!"_ZTS1g", !15, i64 0, !15, i64 8}
> !15 = !{!"any pointer", !4, i64 0}
> !16 = !{!14, !15, i64 8}
>
> On 3/12/21 1:59 PM, Jordan Rupprecht via llvm-commits wrote:
>> Author: Jordan Rupprecht
>> Date: 2021-03-12T13:59:14-08:00
>> New Revision: 8d20f2c2c66eb486ff23cc3d55a53bd840b36971
>>
>> URL:https://github.com/llvm/llvm-project/commit/8d20f2c2c66eb486ff23cc3d55a53bd840b36971 <https://github.com/llvm/llvm-project/commit/8d20f2c2c66eb486ff23cc3d55a53bd840b36971>
>> DIFF:https://github.com/llvm/llvm-project/commit/8d20f2c2c66eb486ff23cc3d55a53bd840b36971.diff <https://github.com/llvm/llvm-project/commit/8d20f2c2c66eb486ff23cc3d55a53bd840b36971.diff>
>>
>> LOG: Revert "[CodeGenPrepare] Fix isIVIncrement (PR49466)"
>>
>> This reverts commit cf82700af8c658ae09b14c3d01bb1e73e48d3bd3 due to a compile timeout when building the following with `clang -O2`:
>>
>> ```
>> template <class, class = int> class a;
>> struct b {
>> using d = int *;
>> };
>> struct e {
>> using f = b::d;
>> };
>> class g {
>> public:
>> e::f h;
>> e::f i;
>> };
>> template <class, class> class a : g {
>> public:
>> long j() const { return i - h; }
>> long operator[](long) const noexcept;
>> };
>> template <class c, class k> long a<c, k>::operator[](long l) const noexcept {
>> return h[l];
>> }
>> template <typename m, typename n> int fn1(m, n, const char *);
>> int o, p;
>> class D {
>> void q(const a<long> &);
>> long r;
>> };
>> void D::q(const a<long> &l) {
>> int s;
>> if (l[0])
>> for (; l.j(); ++s) {
>> if (l[s])
>> while (fn1(o, 0, ""))
>> ;
>> r = l[s] / p;
>> }
>> }
>> ```
>>
>> Added:
>>
>>
>> Modified:
>> llvm/lib/CodeGen/CodeGenPrepare.cpp
>>
>> Removed:
>> llvm/test/CodeGen/X86/pr49466.ll
>>
>>
>> ################################################################################
>> diff --git a/llvm/lib/CodeGen/CodeGenPrepare.cpp b/llvm/lib/CodeGen/CodeGenPrepare.cpp
>> index 0f698dd3b190..0b1156e2ace7 100644
>> --- a/llvm/lib/CodeGen/CodeGenPrepare.cpp
>> +++ b/llvm/lib/CodeGen/CodeGenPrepare.cpp
>> @@ -1332,7 +1332,7 @@ getIVIncrement(const PHINode *PN, const LoopInfo *LI) {
>>
>> static bool isIVIncrement(const BinaryOperator *BO, const LoopInfo *LI) {
>> auto *PN = dyn_cast<PHINode>(BO->getOperand(0));
>> - if (!PN || LI->getLoopFor(BO->getParent()) != LI->getLoopFor(PN->getParent()))
>> + if (!PN)
>> return false;
>> if (auto IVInc = getIVIncrement(PN, LI))
>> return IVInc->first == BO;
>> @@ -1347,7 +1347,6 @@ bool CodeGenPrepare::replaceMathCmpWithIntrinsic(BinaryOperator *BO,
>> if (!isIVIncrement(BO, LI))
>> return false;
>> const Loop *L = LI->getLoopFor(BO->getParent());
>> - assert(L && "L should not be null after isIVIncrement()");
>> // Do not risk on moving increment into a child loop.
>> if (LI->getLoopFor(Cmp->getParent()) != L)
>> return false;
>>
>> diff --git a/llvm/test/CodeGen/X86/pr49466.ll b/llvm/test/CodeGen/X86/pr49466.ll
>> deleted file mode 100644
>> index 4f6574d9bbf2..000000000000
>> --- a/llvm/test/CodeGen/X86/pr49466.ll
>> +++ /dev/null
>> @@ -1,122 +0,0 @@
>> -; RUN: opt < %s -O2 -codegenprepare -S | FileCheck %s
>> -
>> -target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
>> -target triple = "x86_64-unknown-linux-gnu"
>> -
>> - at b = dso_local local_unnamed_addr global i64 0, align 8
>> - at c = dso_local local_unnamed_addr global i64 0, align 8
>> - at d = dso_local local_unnamed_addr global i64 0, align 8
>> - at e = dso_local local_unnamed_addr global i64 0, align 8
>> - at f = dso_local local_unnamed_addr global i64 0, align 8
>> - at g = dso_local local_unnamed_addr global i64 0, align 8
>> -
>> -; CHECK-LABEL: @m(
>> -
>> -define dso_local i32 @m() local_unnamed_addr {
>> -entry:
>> - %0 = load i64, i64* @f, align 8
>> - %1 = inttoptr i64 %0 to i32*
>> - %2 = load i64, i64* @c, align 8
>> - %conv18 = trunc i64 %2 to i32
>> - %cmp = icmp slt i32 %conv18, 3
>> - %3 = load i64, i64* @d, align 8
>> - %conv43 = trunc i64 %3 to i8
>> - %tobool40.not = icmp eq i8 %conv43, 0
>> - br label %for.cond
>> -
>> -for.cond: ; preds = %for.cond39.preheader, %entry
>> - %j.0 = phi i32 [ undef, %entry ], [ %j.1.lcssa, %for.cond39.preheader ]
>> - %p.0 = phi i64 [ undef, %entry ], [ %p.1.lcssa, %for.cond39.preheader ]
>> - %i.0 = phi i32 [ undef, %entry ], [ %i.1.lcssa, %for.cond39.preheader ]
>> - %cmp73 = icmp slt i32 %i.0, 3
>> - br i1 %cmp73, label %for.body.preheader, label %for.cond39.preheader
>> -
>> -for.body.preheader: ; preds = %for.cond
>> - br label %for.body
>> -
>> -for.cond1.loopexit: ; preds = %for.inc34.preheader, %for.end12
>> - br i1 %cmp, label %for.body, label %for.cond39.preheader.loopexit
>> -
>> -for.cond39.preheader.loopexit: ; preds = %for.cond1.loopexit
>> - br label %for.cond39.preheader
>> -
>> -for.cond39.preheader: ; preds = %for.cond39.preheader.loopexit, %for.cond
>> - %j.1.lcssa = phi i32 [ %j.0, %for.cond ], [ %conv18, %for.cond39.preheader.loopexit ]
>> - %p.1.lcssa = phi i64 [ %p.0, %for.cond ], [ 0, %for.cond39.preheader.loopexit ]
>> - %i.1.lcssa = phi i32 [ %i.0, %for.cond ], [ %conv18, %for.cond39.preheader.loopexit ]
>> - br i1 %tobool40.not, label %for.cond, label %for.inc42.preheader
>> -
>> -for.inc42.preheader: ; preds = %for.cond39.preheader
>> - br label %for.inc42
>> -
>> -for.body: ; preds = %for.body.preheader, %for.cond1.loopexit
>> - %l.176 = phi i8 [ %sub, %for.cond1.loopexit ], [ 0, %for.body.preheader ]
>> - %p.175 = phi i64 [ 0, %for.cond1.loopexit ], [ %p.0, %for.body.preheader ]
>> - %j.174 = phi i32 [ %conv18, %for.cond1.loopexit ], [ %j.0, %for.body.preheader ]
>> - %tobool.not = icmp eq i32 %j.174, 0
>> - br i1 %tobool.not, label %cleanup45, label %for.cond2.preheader
>> -
>> -for.cond2.preheader: ; preds = %for.body
>> - %tobool3.not69 = icmp eq i64 %p.175, 0
>> - %.pr.pre = load i64, i64* @e, align 8
>> - br i1 %tobool3.not69, label %for.end12, label %for.body4.preheader
>> -
>> -for.body4.preheader: ; preds = %for.cond2.preheader
>> - %4 = sub i64 0, %p.175
>> - %xtraiter = and i64 %4, 7
>> - %lcmp.mod.not = icmp eq i64 %xtraiter, 0
>> - br i1 %lcmp.mod.not, label %for.body4.prol.loopexit, label %for.body4.prol.preheader
>> -
>> -for.body4.prol.preheader: ; preds = %for.body4.preheader
>> - %5 = mul nsw i64 %xtraiter, -1
>> - br label %for.body4.prol
>> -
>> -for.body4.prol: ; preds = %for.body4.prol.preheader, %for.body4.prol
>> - %lsr.iv = phi i64 [ 0, %for.body4.prol.preheader ], [ %lsr.iv.next, %for.body4.prol ]
>> - %lsr.iv.next = add nsw i64 %lsr.iv, -1
>> - %prol.iter.cmp.not = icmp eq i64 %5, %lsr.iv.next
>> - br i1 %prol.iter.cmp.not, label %for.body4.prol.loopexit.loopexit, label %for.body4.prol
>> -
>> -for.body4.prol.loopexit.loopexit: ; preds = %for.body4.prol
>> - %6 = sub i64 %p.175, %lsr.iv.next
>> - br label %for.body4.prol.loopexit
>> -
>> -for.body4.prol.loopexit: ; preds = %for.body4.prol.loopexit.loopexit, %for.body4.preheader
>> - %p.270.unr = phi i64 [ %p.175, %for.body4.preheader ], [ %6, %for.body4.prol.loopexit.loopexit ]
>> - %7 = icmp ugt i64 %p.175, -8
>> - br i1 %7, label %for.end12, label %for.body4.preheader89
>> -
>> -for.body4.preheader89: ; preds = %for.body4.prol.loopexit
>> - br label %for.body4
>> -
>> -for.body4: ; preds = %for.body4.preheader89, %for.body4
>> - %p.270 = phi i64 [ %inc11.7, %for.body4 ], [ %p.270.unr, %for.body4.preheader89 ]
>> - %inc11.7 = add i64 %p.270, 8
>> - %tobool3.not.7 = icmp eq i64 %inc11.7, 0
>> - br i1 %tobool3.not.7, label %for.end12.loopexit, label %for.body4
>> -
>> -for.end12.loopexit: ; preds = %for.body4
>> - br label %for.end12
>> -
>> -for.end12: ; preds = %for.end12.loopexit, %for.body4.prol.loopexit, %for.cond2.preheader
>> - %8 = load i32, i32* %1, align 4
>> - %conv23 = zext i32 %8 to i64
>> - %9 = load i64, i64* @b, align 8
>> - %div24 = udiv i64 %9, %conv23
>> - store i64 %div24, i64* @b, align 8
>> - %sub = add i8 %l.176, -1
>> - %tobool32.not72 = icmp eq i64 %.pr.pre, 0
>> - br i1 %tobool32.not72, label %for.cond1.loopexit, label %for.inc34.preheader
>> -
>> -for.inc34.preheader: ; preds = %for.end12
>> - store i64 0, i64* @e, align 8
>> - br label %for.cond1.loopexit
>> -
>> -for.inc42: ; preds = %for.inc42.preheader, %for.inc42
>> - br label %for.inc42
>> -
>> -cleanup45: ; preds = %for.body
>> - %cmp13 = icmp ne i8 %l.176, 0
>> - %conv16 = zext i1 %cmp13 to i32
>> - ret i32 %conv16
>> -}
>>
>>
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
>> https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits <https://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/20210312/b9fcb8b9/attachment.html>
More information about the llvm-commits
mailing list