LLVM 3.0rc2: popcnt instruction fails on OSX (invalid suffix)

[llvmbot]

Bugzilla Link	11291
Resolution	WONTFIX
Resolved on	Nov 03, 2011 01:56
Version	2.9
OS	MacOS X
Attachments	testcase
Reporter	LLVM Bugzilla Contributor
CC	@lattner

Extended Description

I've attached a small test case that produces assembly code that the assemblers on OSX can't handle. OSX doesn't handle the length suffixes on the popcnt instruction (e.g. {w,l,q}).

Here is a related GCC bug for this issue:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=34497#c14

I believe they fixed it by not adding the suffix on OSX, which works fine. Indeed changing instruction definition in lib/Target/X86/X86InstrSSE.td to drop the suffix, works fine.

The integrated assembler of llc works fine but GHC which uses LLVM can't use the integrated assembler at this stage as we do some assembly post processing.

== To Reproduce ==

1. llc testcase.ll -o testcase.s
2. gcc -m32 -c testcase.s -o b.o
   (or as -arch i386 testcase.s)

== Expected Output ==

Shiny object file

== Actual Output ==

testcase.s:7:suffix or operands invalid for `popcnt'

[llvmbot]

This bug is pretty important to GHC as it is a serious issue with using LLVM 3.0 on OSX.

[llvmbot]

Oh and this is a regression from 2.9, since 2.9 didn't use the popcnt machine instruction but reduced popcnt to a series of other machine instructions.

[lattner]

David, I'm not sure what you're requesting here. The Mac OS system assembler isn't moving forward, and we have no plans to keep llvm compatible with it. Why can't you use the integrated assembler?

-Chris

[llvmbot]

Chris, I was proposing making LLVM compatible with the Mac OS system assembler. However, Evan also pointed out that the system assembler isn't moving forward and that LLVM doesn't want to concern itself with maintaining comparability. I wasn't aware of this but will try to change over to using clang to compile the assembly files, I expect that will work fine but can't test right now.

We can't use the integrated assembler at this stage as we do some post processing on the assembly produced to implement an optimization that we can't achieve with LLVM. (Basically we need control over how functions and globals are arranged in memory, so our post processing stage picks up all the functions and global data and arranges them into an optimized order/layout). I hope someday to submit patches to LLVM to get support for this feature but haven't had time for this and don't expect any in the near future.

[lattner]

Hi David,

Does it work to run clang on the .s file? That should produce a .o file from your (munged) .s file, using the integrated assembler.

-Chris

[llvmbot]

Hi Chris,

Yes using clang works fine. Thanks! Sorry for the noise. I'll switch GHC over to using clang. You can close this bug.

Cheeers,
David