Make the test case insensitive to compiler optimizations.

In D50304#3611390, @goncharov wrote:

FYI this caused a test breakage at lldb-aarch64-ubuntu, https://lab.llvm.org/buildbot/#/builders/96/builds/25128. I have reverted this in fe6db8d03ff16a65f57af24d2cb04f489e2e9b0c

Thank you.

Sorry for spamming. But I am not sure if you have received my previous messages.

Ping.

In D50304#3586710, @jingham wrote:

For some reason I'm not getting mail notifications for review changes, sorry about that.

This is certainly better than the original implementation. Among other things, if we find an exact match, we really shouldn't be doing any more inexact matches, so after the first exact hit it should have switched exact to true.

But if you had line tables laid out like (this is in increasing order of address:

10
20
30
10
16

and you did "thread until 15", this would find the inexact match at 20, switch to an exact match for line 20 and find no other matches. But the gap between 10 & 16 in the line table is maybe an even more plausible place to put the line 15 until breakpoint, so maybe we did want to throw a breakpoint there as well?

@jingham
Nope, with the current patch, we would find the inexact match at 16.

@jingham looks like you missed to notice this one as well.

Address feedback.

Ping.

Ping.

Removed llvm/IR/Metadata.h header inclusion.

Address feedback.

Ping.

Ping.

Thank you @aprantl.

Address review feedback.

Thank you very much @djtodoro.

Ping.

Does this look good now to land?

Does this look good now to land?

Went ahead and added the error capturing for setting incorrect thread id and also the fix.

Added the test coverage.

I have tried to add a test but it's becoming difficult for me to come-up with a failing test scenario. Is it okay to commit this without a test?

Ping

Ping.

Ping

@djtodoro thanks for the quick response

Address feedback.

Could someone review/comment/approve this revision?

Rebase onto ToT.

In D114652#3367310, @arsenm wrote:

I'm not sure I understand why a separate isPhysRegUsed check is needed since ultimately that's what IPRA is doing, but I don't see how it could hurt

Address review feedback.

Fixed the blender application failure seen in the downstream testing.

In D120922#3363739, @arsenm wrote:

An artificial test with high pressure would be more reliable long term. Branches on constants could be fragile

Rebase and ping.

Fixed an issue exposed by OpenMP LIT test failures.

This test will be used to verify the fix that will be added to https://reviews.llvm.org/D114652 for resolving the OpenMP LIT failures.

In D117064#3236907, @Flakebi wrote:

This looks like it was fixed by D114652, but that got reverted?

I think I have addressed all the outstanding feedback. Is this good to go now?

Rebase onto ToT and use copyImplicitOps.

Rebase onto ToT and also made the following changes to address review feedback.

Rebase and fix tests in llvm/test/CodeGen/AMDGPU/callee-frame-setup.ll.

Re-use the existing constexpr option table.

Remove the unnecessary if {} gaurding around range for loop.

Address review comments.

In D114448#3150900, @DavidSpickett wrote:

Not sure if you're a regular contributor and already know this but in case not, I tend to use https://clang.llvm.org/docs/ClangFormat.html#script-for-patch-reformatting for doing clang-format stuff. It'll format only the lines you've changed.

Thanks for the comments.