llvm dev - Mar 2010 - [LLVMdev] Let's -disable-post-ra for x86

We are currently running the post-ra scheduler for x86 with -O3. It is not
really clear that it is worth the compile time.

X86 scheduling is a mystery that post-ra has not solved.

I ran the nightly test suite on x86_64 with LTO enabled and -O3 with and without
-disable-post-ra. Code generation is 10% faster overall with -disable-post-ra,
and the runtime impact varies from a 17% speedup to a 17 percent slowdown.

Below are the tests that had a more than 3% change in runtime.

I think we should disable post-ra scheduling for x86.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: disable-post-ra.pdf
Type: application/pdf
Size: 28700 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20100311/3be6334f/attachment.pdf>

Hi Jakob,

It's not completely clear from your post, I think your chart is just
for x86_64? What does it look like for x86_32?

 - Daniel

On Thu, Mar 11, 2010 at 4:58 PM, Jakob Stoklund Olesen <stoklund at
2pi.dk> wrote:
> We are currently running the post-ra scheduler for x86 with -O3. It is not
really clear that it is worth the compile time.
>
> X86 scheduling is a mystery that post-ra has not solved.
>
> I ran the nightly test suite on x86_64 with LTO enabled and -O3 with and
without -disable-post-ra. Code generation is 10% faster overall with
-disable-post-ra, and the runtime impact varies from a 17% speedup to a 17
percent slowdown.
>
> Below are the tests that had a more than 3% change in runtime.
>
> I think we should disable post-ra scheduling for x86.
>
>
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
>
>

On Mar 12, 2010, at 12:23 AM, Daniel Dunbar wrote:
> Hi Jakob,
> 
> It's not completely clear from your post, I think your chart is just
> for x86_64? What does it look like for x86_32?
You're right, those were 64-bit numbers.

32-bit coming up.

On Mar 12, 2010, at 12:23 AM, Daniel Dunbar wrote:
> Hi Jakob,
> 
> It's not completely clear from your post, I think your chart is just
> for x86_64? What does it look like for x86_32?
Here are the 32-bit numbers. I have omitted the compile times, since that is
pretty clearly a universal win.

The table has runtime of llc-built tests with -O3 and LTO enabled,
TARGET_FLAGS='-arch i386'.

-------------- next part --------------
A non-text attachment was scrubbed...
Name: m32-disable-post-ra.pdf
Type: application/pdf
Size: 30368 bytes
Desc: not available
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20100312/ff9d3b19/attachment.pdf>