Assertion `!isa<CallBrInst>(Preds[i]->getTerminator()) && "Cannot split an edge from a CallBrInst"' failed. #896
Comments
|
This is still reproducible on ToT LLVM and Linux. I'll reduce this down and report it upstream. |
|
This is a regression in 10.0.0, as 9.0.1 is fine. I'm running a bisect now. creduce spits out: |
|
Run the below with |
|
I think this is the fix: |
|
Bisect points to llvm/llvm-project@cc95a45, don’t know if that is accurate or I took a wrong turn somewhere (I tried to do it automatically and might have messed something up). I can try that patch later. |
|
I think llvm/llvm-project@cc95a45 just exposed the issue. The change of the |
|
This is weird, I cannot reproduce on master/ToT llvm. Is this only reproducible in clang-10? |
|
I could reproduce it with ToT LLVM this morning. |
$ make CC=clang i386_defconfig
$ make CC=clang drivers/char/random.o
CALL scripts/checksyscalls.sh
CALL scripts/atomic/check-atomics.sh
CC drivers/char/random.o
$ git show -1 | head -n1
commit ca7e1fd1026c5af6a533b4b5447e1d2f153e28f2
$ clang -O2 -no-integrated-as -march=i686 -m32 -c -o /dev/null tmp.c
$ echo $?
0Compiler is haunted, 👻 , confirmed. |
$ opt -licm foo.ll -S
; ModuleID = 'foo.ll'
source_filename = "foo.ll"
target datalayout = "e-m:e-p:32:32-p270:32:32-p271:32:32-p272:64:64-f64:32:64-f80:32-n8:16:32-S128"
target triple = "i386-unknown-linux-gnu"
define dso_local i32 @j() local_unnamed_addr #0 {
entry:
br label %for.cond
for.cond: ; preds = %cond.true.i, %entry
callbr void asm sideeffect "", "X,~{dirflag},~{fpsr},~{flags}"(i8* blockaddress(@j, %for.end)) #1
to label %cond.true.i [label %for.end]
cond.true.i: ; preds = %for.cond
br i1 true, label %for.end, label %for.cond
for.end: ; preds = %cond.true.i, %for.cond
%asmresult1.i.i2 = phi i32 [ 0, %cond.true.i ], [ undef, %for.cond ]
ret i32 undef
}
attributes #0 = { "use-soft-float"="false" }
attributes #1 = { nounwind }
!llvm.ident = !{!0}
!0 = !{!"clang version 11.0.0 (https://github.com/llvm/llvm-project.git faf74f0b2b2e51989780a95d4f1466fd6e1beb80)"}in Bill's reproducer, there's a weird branch on |
|
Sorry, needs OpenMandriva’s https://github.com/OpenMandrivaAssociation/kernel-release-clang/blob/master/common.config |
|
I have a Clang built from yesterday: With @nathanchance's reproducer, I see the expected crash. |
|
thanks for the hot tipz, but I still cannot reproduce. I wonder if something covered this back up since yesterday, let me do some bisection. |
|
crashing at f5efa08247c7 (yesterday morning). Bisecting forward. |
|
Just tried TOT right now. Can confirm that it is passing again, so compiler isn't necessarily as haunted as previously indicated. 👻 |
|
Fixed by @gwelymernans two hours ago llvm/llvm-project@2fe4576 (without code review...) |
|
Notes to self:
before: callbr void asm sideeffect "", "X,~{dirflag},~{fpsr},~{flags}"(i8* blockaddress(@j, %for.end)) #1
to label %cond.true.i [label %for.end]which produces: callbr void asm sideeffect "", "X,~{dirflag},~{fpsr},~{flags}"(i8* blockaddress(@j, %for.end)) #1
to label %cond.true.i [label %for.end.split.loop.exit1]which fails module verification because the indirect label list was updated but not the argument to the asm. We've had this issue in the past, but |
|
I've posted https://reviews.llvm.org/D74947 to demonstrate some breakage in If I remove the |
|
This has been fixed in LLVM 11.0.0 and it was picked into LLVM 10.0.0: |
nathanchance
added
[FIXED][LLVM] 10
I'm compiling kernel-5.5.4 with LLVM/clang-10.0.x on i686
Build details and logs can be found here
https://abf.openmandriva.org/build_lists/709792
The text was updated successfully, but these errors were encountered: