Sorry for late update, a little busy in these 2 month : )

NO problem, Thanks!

Hello @morehouse, thanks very much for your careful review, I'll fix all your comments, just for "ClUntagPointer", I feel it very useful in my develop.

Hello, @morehouse, I first spilt the change of compiler to https://reviews.llvm.org/D102901
I'll update this link after that small patch in.

Hello @morehouse, thanks for your work too:
1 OK, I'll break it up into 2 patches, one focus on changes in compiler, the other focus on changes in compiler-rt.

Hello @eugenis @vitalybuka @morehouse I see you change the related code at compiler-rt, Could you help review this patch? This patch just limited the tag for x86_64.

In D102472#2770228, @pcc wrote:

It doesn't look like this was reviewed by anyone. Could you revert this please until it gets reviewed?

Update clang-format

refine, compatible with old AMX API

Fix bkramer's multithread problem.

+1

In D100446#2712464, @djtodoro wrote:

So this patch will remove reporting warnings of missing debug loc to PHIs completely, right?

Yes

Can we at least introduce an option to keep checking the PHIs, because there are cases it can be a real issue detected by debugify.

Of course we can. But this warning just for pass generate PHI, we usually ignore warnings if they are make sense. I think the reason @Nuullll write this patch is just that "PHI existed in too many passes, but very few pass generate PHI"

Use clang-format refine.
And add Fixme: at LowAMXType

rebase

+1 first

Thanks for your detailed explanation, In my eyes, D75242 eliminated uncommon "warnings" (only pass generate PHI node), but generate common "warnings" (all pass with PHI node ).
Let me first +1 for this patch if no other reviewers object.
LGTM

@vks, do you know why we set PHI with debug_loc but not check it before?

Thanks for your reviewing!

I didn't find any problem in the main context of the patch, +1 first.

Refined: No new warning in Release build.

I'll try build it in release, make sure it no warnings, thanks!

Refine Shape info:
In AMX intrinsics we let Shape = {Row, Col (Bytes) }, but the RealCol = Col / ElementSize. We may use the RealCol as a new Row for other new created AMX intrinsics.

The patch is big, thanks for Pengfei's review!

Address pengfei's comments:

We didn't cover this case currently, also O0 + greedy is not support.

Address Pengfei's comments:

LGTM, (spill will only happen in the same shape tile register.)

In D98757#2631042, @lebedev.ri wrote:

The ongoing special-casing of X86_AMXTy through the llvm due to the inability of the existing backend passes to handle certain llvm ir constructs.

In D98757#2631019, @lebedev.ri wrote:

Once again, i suggest to bring this up on llvm-dev.

In D98757#2630968, @lebedev.ri wrote:

I think that is a traditional backend problem that the pass will just have to be updated to deal with.

In D98757#2630942, @lebedev.ri wrote:

I strongly suggest you bring up this ongoing creep of if (DestTy->isX86_AMXTy()) return false; on llvm-dev.
I strongly supsect you are covering up bugs in you backend/pass with them.

In D98757#2630844, @LuoYuanke wrote:

Probably we need a .ll test case to for constant folding.

Fold constant is done in CSE and SCCP which are both passes run in Clang (O2)

In D98757#2630764, @LuoYuanke wrote:

Would you add a test case for it?

+1 first, didn't see key problems.

Just for interest to improve ReloadTileConfig, now Abandoned this patch