A few inline assembly questions

Hey,

I'm contemplating adding llvm-gcc as a core platform to Qt, meaning we'd
release Qt unless we would make sure it compiles and works with llvm-gcc but
I have some problems with making it work cleanly so I just wanted to ask a
few simple questions.
The problems refer to llvm and llvm-gcc from respectively cvs/svn which, I'm
assuming, means upcoming 2.0 release.

1) we added a little inline assembly to do runtime detection of a few cpu
features. llvm-g++ crashes on the attached testcase (test.cpp).

2) i have code utilizing whatever vector insructions i can get my hands on to
speed up some very common rendering operations. llvm-gcc doesn't seem to
support mmx intrinsics which makes Qt compiled with llvm-gcc quite a bit
slower when it comes to graphics. test2.cpp shows in essence what we're doing
in Qt. we basically have a very simple class that defines simple static
methods like negate/add/byte_mul/interpolate_pixel. now the question is if
there is any way of getting code like this to utilize vector instructions
with llvm-gcc. i wouldn't mind having llvm specific path there as long as it
works.

Zack

test2.cpp (5.43 KB)

test.cpp (1.29 KB)

I'm contemplating adding llvm-gcc as a core platform to Qt, meaning we'd
release Qt unless we would make sure it compiles and works with llvm-gcc but
I have some problems with making it work cleanly so I just wanted to ask a
few simple questions.

Very nice!

The problems refer to llvm and llvm-gcc from respectively cvs/svn which, I'm
assuming, means upcoming 2.0 release.

Great, that is very useful.

1) we added a little inline assembly to do runtime detection of a few cpu
features. llvm-g++ crashes on the attached testcase (test.cpp).

Anton filed a bug for this, I'll try to make sure it gets fixed this week.

2) i have code utilizing whatever vector insructions i can get my hands on to
speed up some very common rendering operations. llvm-gcc doesn't seem to
support mmx intrinsics which makes Qt compiled with llvm-gcc quite a bit
slower when it comes to graphics.

As Anton mentioned, work is actively underway to add MMX intrinsic support. Right now we have very basic support for a few simple operations like add and multiply. Bill Wendling is the one working on this, I forwarded your testcase to him so that he can focus on those operations first.

test2.cpp shows in essence what we're doing
in Qt. we basically have a very simple class that defines simple static
methods like negate/add/byte_mul/interpolate_pixel. now the question is if
there is any way of getting code like this to utilize vector instructions
with llvm-gcc. i wouldn't mind having llvm specific path there as long as it
works.

Absolutely. Until now, there hasn't been anything pushing for MMX intrinsic support (most clients we've worked with so far use SSE2 instead of MMX). LLVM definitely needs to support the MMX intrinsics, when Bill has your testcase working, I'd appreciate it if you could try Qt with MMX enabled again. I filed http://llvm.org/PR1260 to track your specific testcase (PR1222 is the bug for MMX intrinsics in general). If you want to see when this is completed, it is easiest to CC yourself on the bug.

If there is anything else we can do, please don't hesitate to email the list further (or file bugs) :slight_smile:

-Chris

2) i have code utilizing whatever vector insructions i can get my hands on to
speed up some very common rendering operations. llvm-gcc doesn't seem to
support mmx intrinsics which makes Qt compiled with llvm-gcc quite a bit
slower when it comes to graphics.

As Anton mentioned, work is actively underway to add MMX intrinsic
support. Right now we have very basic support for a few simple operations
like add and multiply. Bill Wendling is the one working on this, I
forwarded your testcase to him so that he can focus on those operations
first.

Got 'em. :slight_smile: I think my schedule's slightly more open this week, so I should be able to add the missing operators easily.

test2.cpp shows in essence what we're doing
in Qt. we basically have a very simple class that defines simple static
methods like negate/add/byte_mul/interpolate_pixel. now the question is if
there is any way of getting code like this to utilize vector instructions
with llvm-gcc. i wouldn't mind having llvm specific path there as long as it
works.

Absolutely. Until now, there hasn't been anything pushing for MMX
intrinsic support (most clients we've worked with so far use SSE2 instead
of MMX). LLVM definitely needs to support the MMX intrinsics, when Bill
has your testcase working, I'd appreciate it if you could try Qt with MMX
enabled again. I filed http://llvm.org/PR1260 to track your specific
testcase (PR1222 is the bug for MMX intrinsics in general). If you want
to see when this is completed, it is easiest to CC yourself on the bug.

Also, if you could do some performance comparisons between GCC and LLVM, it would be a good measurement of how we're doing. From the few instructions I've added, we tend to generate better code. :slight_smile:

-bw

> As Anton mentioned, work is actively underway to add MMX intrinsic
> support. Right now we have very basic support for a few simple
> operations
> like add and multiply. Bill Wendling is the one working on this, I
> forwarded your testcase to him so that he can focus on those
> operations
> first.

Got 'em. :slight_smile: I think my schedule's slightly more open this week, so I
should be able to add the missing operators easily.

Ah, that would be wonderful :slight_smile:

Also, if you could do some performance comparisons between GCC and
LLVM, it would be a good measurement of how we're doing. From the few
instructions I've added, we tend to generate better code. :slight_smile:

I'll definitely do that :slight_smile:
I'm planning to run some of the other benchmarks we have here but I'm afraid I
won't get to it until Friday. We have a few text decoding/manipulation
benchmarks where GCC is generating suboptimal code (Visual C++ on identical
machine manged to outperform it by almost two-fold) so I'm eager to see how
LLVM stacks up. As soon as I have some results I'll post them here.

Zack