All patches are now committed.

Hi Sam,

LGTM

ping

Approved now that I've properly tested it.

One small comment otherwise LGTM

ping

ping

Hi Andrew, the new version LGTM

Hi Sam,

Ping

I think that the previous comments were addressed. LGTM

Just a small comment, otherwise LGTM, thanks

Hi David,

LGTM

LGTM, given it is needed to bring ARM bots back online

The change makes sense to me and will fix ARM broken bots, I'd say that it can be accepted if there is no objection within a day

LGTM

I'd check if there are some calls of modifiesRegister that pass nullptr, the patch LGTM otherwise

(Sorry for the delay, I took some vacations)

Thanks for raising this issue, I've two inlined comments

Hi Louis,

LGTM

Thsi was committed as D80125 and D80127

Issue fixed by D40061

Hi Sam,

My bad, I applied the previous version where a branch was inserted in those tests but your last version is fine.

Hi Sam,

I think a few more testcases need to be updated:

I've tested the patch and it fixes the liveness issue seen with Machine Outliner enabled.

Removed uneeded LR addition to live-in set.

I've extracted parts which handle calls inside outlined frames (I'll submit that in the next patch) and enhanced the testcases.

In D83313#2138334, @samparker wrote:

I guess I just don't understand why the BX_RET wouldn't be marked with using the link register in the first place?

In D83313#2137190, @efriedma wrote:

I'm not really happy with this approach; if LiveRegUnits isn't producing correct results, we should fix it, not try to hack around it.

Maybe we should revive D40061...

Hi Sam,

No there is 3 more patches in the set, one for the default mode (save/restore of lr on the stack), one which will add the support of calls in the outlined regions and the most tricky one which will fixup the stack to handle load/store. In the end we should reach ~3% of size reduction in T32 and ~5% in ARM (this is what I observed on average for various benchs).

Updated testcase

Change register class to rGPR for both modes and updated testcase

This was splitted from D76068

I found a last issue when doing a full bootstrap with -moutline used to build clang, there was a case of thunk outlining where the original was call a BLX LR which was broken by the outlining call BL OUTLINE_FUNCTION_X I fixed this issue by checking the use of link register before handling calls and I've added a test case.

In D76066#2022025, @samparker wrote:

Ah, yes, good point. Then I have nothing else, but definitely worth waiting to see if Eli has further comments.

In D76066#2014697, @samparker wrote:

Thanks for adding the MVE changes, but I also still don't see any DSP tests, i.e QADD, SADD16.

This update avoids outlining MVE instructions.

LGTM

Here is a new update of the patch.

Ok thanks, LGTM

Hi Sam,