Instcombine hangs with "opt -functionattrs -disable-basicaa -O1"

[mikaelholmen]

Bugzilla Link	44541
Resolution	FIXED
Resolved on	Feb 10, 2020 02:36
Version	trunk
OS	Linux
Blocks	#43900
Attachments	bbi-37458_x86.ll reproducer
CC	@davidbolvansky,@zmodem,@RKSimon,@nikic,@rotateright
Fixed by commit(s)	a148b9e

Extended Description

Running
opt -functionattrs -disable-basicaa -O1 -S -o - ./bbi-37458_x86.ll
seems to hang forever.

Adding -debug-pass=Executions shows

[2020-01-14 15:08:50.872376271] 0x611e690 Executing Pass 'Combine redundant instructions' on Function 'test_sub3_4'...

and adding -debug gives repeated occurences of

IC: Visiting: %_tmp2687 = select i1 %1808, i16 0, i16 %_tmp2685
IC: Visiting: %1808 = icmp sgt i16 0, %_tmp2685
IC: Mod = %1808 = icmp sgt i16 0, %_tmp2685
New = %1808 = icmp slt i16 %_tmp2682, 0
IC: ADD: %_tmp2687 = select i1 %1808, i16 0, i16 %_tmp2685
IC: ADD: %1808 = icmp slt i16 %_tmp2682, 0
IC: Visiting: %1808 = icmp slt i16 %_tmp2682, 0
IC: Visiting: %_tmp2687 = select i1 %1808, i16 0, i16 %_tmp2685

I've tried to reduce the command line to some minor variation of just
opt -instcombine
but I can't seem to manage that. :/

The problem seems to be fairly old, I've seen it in builds from around late March 2019.

[mikaelholmen]

The problem seems to be fairly old, I've seen it in builds from around late
March 2019.

It starts failing with 3af5b28:

[ValueTracking] Use ConstantRange based overflow check for signed sub

No idea if it's a coincidence or if 3af5b28 is actually doing something weird.

[nikic]

The reason this might not be reproducible standalone is that apparently there is no way right now to disable expensive combines in "opt -instcombine". I always thought those were default disabled in that case, but looking at the code, even doing something like "opt -expensive-combines=0" is still going to run them.

This is relevant here, because it determines whether the @​idusf() calls get constant folded.

[nikic]

Here's a reduced test case based on https://reviews.llvm.org/D72861 and opt -instcombine -expensive-combines=0:

declare i16 @​llvm.ssub.sat.i16(i16, i16)

define i16 @​passthru(i16 returned %x) {
ret i16 %x
}

define i16 @​test(i16 %arg) {
%zero = call i16 @​passthru(i16 0)
%sub = call i16 @​llvm.ssub.sat.i16(i16 %arg, i16 %zero)
%cmp = icmp slt i16 %sub, 0
%ret = select i1 %cmp, i16 0, i16 %sub
ret i16 %ret
}

[mikaelholmen]

The reason this might not be reproducible standalone is that apparently
there is no way right now to disable expensive combines in "opt
-instcombine". I always thought those were default disabled in that case,
but looking at the code, even doing something like "opt
-expensive-combines=0" is still going to run them.

This is relevant here, because it determines whether the @​idusf() calls get
constant folded.

Aha nice, that explains why it was hard. Thanks for looking into this.

[nikic]

Candidate patch: https://reviews.llvm.org/D72944

[nikic]

Alternative fix from https://reviews.llvm.org/D73849 has landed, keeping this open to track LLVM 10 backport.

[zmodem]

Alternative fix from https://reviews.llvm.org/D73849 has landed, keeping
this open to track LLVM 10 backport.

Cherry-picked as fc12083

[zmodem]

mentioned in issue #43900