- User Since
- Feb 20 2017, 1:20 PM (22 w, 1 h)
as @grosser pointed out, this is fixed by r308187. Thanks!
Abandoned in favour of D35795 which actually solves the problem.
- [NFC wrt patch] removed return false because we call a noreturn function anyway
- [NFC wrt patch] change to use report_fatal_error
- [NFC wrt patch] copy Array->bound, don't acquire and release
@grosser I do not believe so, release sets the ptr within the isl::multi_pw_aff to nullptr. So, the destructor will not free anything.
The commit r308869 was incorrectly labeled - it was supposed to close D35770. Writing this here as documentation.
This was closed by r308869. Unfortunately, the commit was mis-labeled, and wound up closing D35771
@grosser - Then is every use of llvm_unreachable a bug?
@Meinersbur: I would like an LGTM on your part to be comfortable with upstreaming this.
- [NFC wrt patch] copy the isl_multi_pw_aff object rather than acquiring and releasing it.
@grosser I'm not sure how to build up the correct access union map.
@grosser I'm not entirely sure if this makes it cleaner due to the interfacing with PPCG. I think it's easier to see the transforms that are going on.
[NFC wrt patch] return false indicating failure of regex match if regex did not compile
- [NFC] diff against master.
Fri, Jul 21
Abandoning in favour of allowing regexes to polly-only-func in D35728
Is there some way to test this without having libdevice? The tests break on my mac.
Thu, Jul 20
closed by r308623
closed by r308624
closed by r308623
closed by r308625
- Fix nits.
- Diff against latest version of [2/3] of these changes.
- Remove commented code.
- Change spaces to tabs, looks like PPCG uses tabs.
- Update again because SHA hashes changed, leading to an incorrect diff previously.
- diff with stuff moved to [1/3].
Backport some changes from [2/3] to [1/3].
- update testcase of non-read-only-scalars
- move preloadInvariantLoads to before the RTC is generated.
- [NFC] style fixes
Just to be very sure, I took the .ll file from the testcase and compiled it with the exact commands that we run in the test case.
I compiled the program from non-read-only-scalars.ll and ran in on daint (Piz-daint).
- Use init_device and clear_device nodes in the schedule tree.
marked as WIP because the test cases need another look, and the init/clear_device issue is still being discussed.
@singam-sanjay: We should move the discussion about init / clear device here.
Hm, perhaps updating this diff is too confusing. I am abandoning this huge change and splitting this into 3 chunks.
[Polly] [PPCG] [1/3] Bump up PPCG to vanilla 0.07
@grosser: I wanted to rebase and get a clean history, but PPCGCodeGeneration can be reviewed. The test cases I am dubious about are:
Wed, Jul 19
Tue, Jul 18
I can no longer reproduce this on HEAD. I'll bisect and see what caused the change.
- [WIP] previous commmit makes no sense, that checks if we care about the value at the *end* of a scop, not in between loop iterations
@grosser - with this change, we break the test test/GPGPU/invariant-load-hoisting-with-failing-scop.ll which gets triggered when we have a BuildSuccessful = 0 with invariant load hoisting. However, the only way to trigger this is either
- Cleanup code.
From what I can tell, PPCG does seem to pass "private" scalars:
Mon, Jul 17
A question from someone following this as an "outsider" with no context -
Does this allow others to obtain, say, Scop information (outside of polly) the way one can access DomTree / SCEV information?
@grosser - does the change make sense to you? I'll clean it once I know that this is a sane change.
- [NFC wrt patch] remove debug metadata from testcase.
here's the generated code for this example.
- Call preloadInvariantLoads() before invoking GPUNodeBuilder::create.
Fri, Jul 14
Nope, this is definitely incorrect, we generate incorrect IR in the kernel module.
@grosser: It would appear that since you blocked on the "changes required", I can no longer close the revision (which is upstreamed with r308010). Hence, I am abandoning this revision.