- User Since
- Nov 23 2015, 1:58 PM (95 w, 4 d)
Mon, Sep 18
LGTM. It would be nice if we could defer ISel lowering parts occuring during legalization so we can remove the pre-legalization step,
Fri, Sep 8
Yet another case of unused nodes being left behind and causing degradation. We should change the behavior of the DAG ReplaceAllUses-ish functions to check if the old source node becomes unused and if so delete the node. This would obviate the need for this patch and catch the lion share of these sorts of issues but will cause assertions in cases a bunch of cases (which is a better situation than the current silent degradation). I have a first-cut patch (D36643) which fixes just the DAGCombiner CombineTo function and all of the assertion failures from update ordering / double deletions.
Tue, Sep 5
The attributes in the test case can probably be pruned, but otherwise LGTM. Thanks.
Mon, Aug 28
Aug 23 2017
Missed line added. This should be okay. Can I get an LGTM?
Aug 11 2017
This is not ready yet, as it needs much more testing to shake out any missing cases. dbabokin should be able to find a number with his testing infrastructure.
Add missing swap
Add swap when both operands are exapnded loads w/ a dependence.
D'oh. It possible though hard. I'll add a cheap check and swap.
Aug 10 2017
Excise NFC change and address comments.
Aug 9 2017
Aug 8 2017
Pull out some NFC changes and use peekThroughBitcast function
Hmm, looking around the only cases I see is that after an assembly label # non-directives are accepted by GCC (here's an example in the wild: https://github.com/android/platform_bionic/blob/master/libc/arch-x86/bionic/__bionic_clone.S). Looking at the parser we do have an explicit check for that in the parser so this should be okay. I also see that at least the arm GCC 4.8.4 errors on hash non-directives at the end of line with various assembly directives / instructions. I suspect there are more test cases that will need to be fixed.
Aug 7 2017
LexLineComment will only be called when we # is a asm string line comment or we're at the start of a statement. This change would miss the case of a hash-based (preprocessor) EOL comment trailing a non-empty line which we do need to catch. I believe (but am not 100% certain) that I did find a live assembly case where the # as line comment interpreation for AArch64 was used.
LGTM. I'll run it locally myself and land it. Thanks.
Aug 4 2017
Based on the UB Error, it seems like the CVT_tied check should be removed and the values of DefaultOffsets shifted over one.
Aug 3 2017
Looks Good. I'll just recheck it again and land it.
Aug 2 2017
Cleanup FirstZeroAfterNonZero calc.
I suspect that the issue is variation in our cmake configurations. I just rebuilt from a clean build and got the same issues.
Aug 1 2017
Rebase past landed NFC commits.
Fold away type modification into landed NFC patch
Looks like this is raising an assertion. Reopening until I can look at it again.
This is factored from D34569
I'm seeing it after this:
Jul 31 2017
Undo unrelated updated.
Unfold unrelated commit.
Unfold unrelated commit.
Remove redundant check
I just did a sanity check on this patch upstream before committing and it crashes on a number of AMDGPU tests. Can you take a quick look at what's wrong?
Jul 30 2017
Great. It looks like you don't have commit access. Would you like me to commit this patch for you?
Jul 28 2017
Can you add a test case?
Jul 27 2017
Rebase and general update. No real change.
Rework to regularize and simplify bitcasting logic.
Add type check in truncate case.
Comment cleanup and apply same logic to final leg of store merge reasoning.
Jul 26 2017
In addition to alignment, we less restrictive of const zero stores over other constant stores. Add logic to prevent skipping zero cases.
Jul 25 2017
Update to weaken type restrictions
Jul 22 2017
Jul 20 2017
Jul 19 2017
Jul 18 2017
Rebase past PR34095 change