[SYCL] Turn off part of SimplifyCFG optimizations in SYCL mode.

[maksimsab]

Usually clang with SimplifyCFG pass optimizes constructs like:

if (i % 2 == 0)
  func();
else
 func();

into one simple func() invocation.
This behaviour might be wrong in cases when func's behaviour depends on
a place where it is written. For this cases we turn off part of optimizations of SimplifyCFG pass.
There is a relevant discussion about introducing a reliable tool for such cases: https://reviews.llvm.org/D85603

[maksimsab]

/summary:run

[ghost]

@maksimsab Why is this change marked as [NFC]?

[AlexeySachkov]

I wonder, would it be better to simply add a "complete switch off" option to the pass itself? Something similar to what is done in #5149?

[maksimsab]

I thought about that. It seemed for me that compiler should behave as usual in case when sycl isn't involved (when -fsycl option is not present).

[AlexeySachkov]

It seemed for me that compiler should behave as usual in case when sycl isn't involved (when -fsycl option is not present).

So the idea is to have that new switch off by default, i.e. by default the pass is working as-is. Under SYCL mode we will turn it on. Anyway, it seems to me that we can minimize amount of changes by simply disabling hoisting common instructions which is only done once (see my other comment)

[AlexeySachkov]

@maksimsab Why is this change marked as [NFC]?

I guess technically, it is supposed to be a non-functional change, because optimizations must not alter the functionality. But the tag on this PR still makes me a bit nervous :)

[keryell]

I guess technically, it is supposed to be a non-functional change, because optimizations must not alter the functionality. But the tag on this PR still makes me a bit nervous :)

I prefer a lacking [NFC] to a misleading [NFC]. :-)

[xtian-github]

Alexey, thanks for creating this new patch. This patch has a better trade-off than previous one. Is it feasible to disable SimplifyCFG for “common call instructions” for SYCL? If it will take a long time to do so. I am ok for you to get your current patch merged first.

[bader]

Is it feasible to disable SimplifyCFG for “common call instructions” for SYCL? If it will take a long time to do so. I am ok for you to get your current patch merged first.

It should be feasible to disable host/sink of "convergent" instructions. I think it's not as simple as config pass update, but we can investigate this option.

@maksimsab, please fix linter complains to unblock further pre-commit testing.

[bader]

@maksimsab, #5305 has been merged. Please, update your branch.

[bader]

/summary:run