This is an archive of the discontinued LLVM Phabricator instance.

[OpenCL] Use long instead of long long in x86 builtins
ClosedPublic

Authored by alexbatashev on May 29 2019, 5:35 AM.

Details

Summary

According to C99 standard long long is at least 64 bits in size. However, OpenCL C defines long long as 128 bit signed integer. This prevents one to use x86 builtins when compiling OpenCL C code for x86 targets. The patch changes long long to long for OpenCL only.

Diff Detail

Event Timeline

alexbatashev created this revision.May 29 2019, 5:35 AM

'O' is an interesting choice. Any real justification for it, or just "what was available"? It definitely needs to be documented in the top of Builtins.def however.

clang/lib/AST/ASTContext.cpp
9362

Likely need to update the asserts in the other places this is used, such as 'Z'.

Also, add the 'O' to the documentation in Builtins.def

A different (perhaps silly) question is why 'W' isn't sufficient? It represents int64_t, which I wonder if is sufficient.

A different (perhaps silly) question is why 'W' isn't sufficient? It represents int64_t, which I wonder if is sufficient.

I had asked in a separate conversation not to change the signatures outside of opencl. I believe W would use 'long' on 64-bit targets that aren't Windows or X32. The builtins are defined by gcc as well so I wanted to avoid making them different as much as possible.

craig.topper added inline comments.May 29 2019, 10:34 AM
clang/lib/AST/ASTContext.cpp
9308

O should be mentioned here

9314

O should be mentioned here

9324

O should be mentioned here

'O' is an interesting choice. Any real justification for it, or just "what was available"? It definitely needs to be documented in the top of Builtins.def however.

O is for OpenCL. I planned to use 'C' but it is already used.

I'll add docs. Thank you.

alexbatashev marked 4 inline comments as done.
erichkeane accepted this revision.May 29 2019, 12:03 PM

Conditional approval, commit with the comment fixed. Let me know if you need me to commit it for you.

clang/include/clang/Basic/Builtins.def
56

End the comment with a period.

This revision is now accepted and ready to land.May 29 2019, 12:03 PM

What about a testcase? It shouldn't be hard to add a small testcase that demonstrate that the changed builtins now work when compiling OpenCL C code for x86. I don't think you have to add all changed builtins to the testcase (but a few to demonstrate the change).

Do you think it's possible to add a test?

alexbatashev set the repository for this revision to rG LLVM Github Monorepo.
alexbatashev marked an inline comment as done.May 31 2019, 2:50 AM

@Ka-Ka good point. Thank you.
@Anastasia Would such tests be ok with you?
@erichkeane Thank you very much. I think I don't have permissions to commit changes and will need someone's help.

Anastasia accepted this revision.Jun 3 2019, 2:56 AM

LGTM! Thanks!

Ka-Ka added inline comments.Jun 3 2019, 3:02 AM
clang/test/CodeGen/builtins-x86.c
127–129

I don't like the introduced casts. Can we change the testcase to operate on the appropriate types instead?
The above two lines could probably be replaced by:

tmp_V2d = __builtin_ia32_undef128();
tmp_V4d = __builtin_ia32_undef256();

What do you think?

228–231

same here?

250

and here?

428

Can we change this line to?

tmp_V4i = __builtin_ia32_pmaddwd128(tmp_V8s, tmp_V8s);
432–433

Can we change this line to?

tmp_V16c = __builtin_ia32_palignr128(tmp_V16c, tmp_V16c, imm_i);
433

Can we change this line to?

tmp_V8c = __builtin_ia32_palignr(tmp_V8c, tmp_V8c, imm_i);
alexbatashev marked 6 inline comments as done.

@Ka-Ka done

Ka-Ka accepted this revision.Jun 3 2019, 3:51 AM

Thanks!

This revision was automatically updated to reflect the committed changes.
Herald added a project: Restricted Project. · View Herald TranscriptJun 3 2019, 5:34 AM