LoopUnroll introduces infinite loop with nested loops

[arsenm]

Bugzilla Link	37859
Version	trunk
OS	All
Attachments	Testcase, More reduced testcase
CC	@hfinkel

Extended Description

In the attached testcase, there are two nested loops with constant trip counts. The exit isn't fully eliminated.

opt -loop-unroll -gvn shows a strange infinite loop is introduced instead.

[llvmbot]

I am unable to reproduce the infinite loop from your reduced example. It is producing a terminating example.

Please give the following clarifications

• Please give us your version of llvm
• your ll file is crashing with the following error

Intrinsic has incorrect argument type!
void (i8 addrspace(5), i8, i32, i1) @​llvm.memset.p5i8.i32

I had to comment out the relevant instruction.

• If possible, please also post the resulting IR with infinite loop
• Can you make the example smaller? It seems to have a lot of unrelated instructions.

[arsenm]

I am unable to reproduce the infinite loop from your reduced example. It is
producing a terminating example.

Please give the following clarifications

• Please give us your version of llvm
• your ll file is crashing with the following error

Intrinsic has incorrect argument type!
void (i8 addrspace(5), i8, i32, i1) @​llvm.memset.p5i8.i32

I had to comment out the relevant instruction.

• If possible, please also post the resulting IR with infinite loop
• Can you make the example smaller? It seems to have a lot of unrelated
  instructions.

This is on trunk. You see the intrinsic type error with any released version. This is about as reduced as I've gotten it since it depends on unroll thresholds etc. The instructions before the loop also matter (although I expect this has more to do with the cleanup GVN manages)

I'm not claiming it doesn't terminate dynamically. Both of these loops should be completely unrolled (and used to be in a previous version in the original testcase), but now this fails to be fully unrolled.

I think now this might be a heuristic failure to unroll the outermost loop and an artifact of test reduction that produces a useless loop after GVN. If I force unrolling the loops or boost the thresholds, something sensible is produced.

[llvmbot]

My guess

[llvmbot]

My guess is that there is a limit on the size on the unrolled code. If your loop has a large body and unrolled loop will be greater than the limit, the LoopUnroll will refuse to unroll. You may be hitting the limit. There is a way to change the default limit. However, I am not clear how to do this from command line.

[arsenm]

My guess is that there is a limit on the size on the unrolled code. If your
loop has a large body and unrolled loop will be greater than the limit, the
LoopUnroll will refuse to unroll. You may be hitting the limit. There is a
way to change the default limit. However, I am not clear how to do this from
command line.

AMDGPU sets the limit to be enormous, so it should easily fit. https://reviews.llvm.org/D48424 fixes unrolling in the original case I set out to fix