Re: [IRA, regresssion] code size regression with IRA

Re: [IRA, regresssion] code size regression with IRA

[Vladimir Makarov]

Hariharan Sandanagobalane wrote:
> Hi Vlad,
> I am attaching a test function which has just a huge switch statement 
> with multiple cases doing different things. This function register 
> allocates on GCC 4.3.2 with no spills at all (at -Os). The function 
> has lots of live ranges that span function calls, but the old 
> allocator seemed to handle it quite well. With IRA, i get lots of 
> spills and a huge frame usage (160 bytes). Can you have a look and see 
> if this is some heuristic going wrong in IRA or just a degenerate case?
>
> I tried this testcase with all the IRA options, but nothing helped 
> much in reducing code size/frame size.
>
> Picochip's assembly is RISC and the instructions should be obvious, 
> but if you need anything, please refer 
> http://www.picochip.com/downloads/picoGcc_reference_manual.pdf or ask 
> me! You can ofcourse build picochip target (picochip-unknown-none) on 
> mainline if you need to.
>
> Attachments:
> 4_3.tgz - Contains source, assembly and all the dump files from -da 
> from GCC 4.3.2.
> mainline.tgz - Similar tar from mainline(as of 11/08/2009).
>
> Your help is greatly appreciated. Let me know if you need any more 
> information.
>
>
As I understood correctly, the  gcc4.3 did not generate spills.  I've 
checked gcc4.4, there are a lot of spills and there is no possibility to 
rid off them because IRA reports register pressure equal to 88 and the 
processor has less 16 regs.

So I guess some other optimization changed between 4.3 and 4.4 started 
to generate such big register pressure.  But unfortunately, I have no 
idea what optimization it is.  Probably you should compare each pass 
dump to find this.  I can not do this because 4.3 has no picochip support.

[IRA, regresssion] code size regression with IRA

[Hariharan]

[-- Attachment #1: Type: text/plain, Size: 1085 bytes --]

Hi Vlad,
I am attaching a test function which has just a huge switch statement 
with multiple cases doing different things. This function register 
allocates on GCC 4.3.2 with no spills at all (at -Os). The function has 
lots of live ranges that span function calls, but the old allocator 
seemed to handle it quite well. With IRA, i get lots of spills and a 
huge frame usage (160 bytes). Can you have a look and see if this is 
some heuristic going wrong in IRA or just a degenerate case?

I tried this testcase with all the IRA options, but nothing helped much 
in reducing code size/frame size.

Picochip's assembly is RISC and the instructions should be obvious, but 
if you need anything, please refer 
http://www.picochip.com/downloads/picoGcc_reference_manual.pdf or ask 
me! You can ofcourse build picochip target (picochip-unknown-none) on 
mainline if you need to.

Attachments:
ira_regression.tgz - Contains source and assembly from GCC 4.3.2, 4.4.0 
and mainline.

Your help is greatly appreciated. Let me know if you need any more 
information.

Thanks and regards
Hari


[-- Attachment #2: ira_regression.tgz --]
[-- Type: application/x-compressed-tar, Size: 48031 bytes --]