ARM: Work around ICE with gcc-4.3 when building ICU

ARM: Work around ICE with gcc-4.3 when building ICU

Using a crosstool-ng build toolchain, the following error is encountered at the
default optimisation level (-O2) when targetting ARMv5:

third_party/icu/source/i18n/rematch.cpp: In member function 'void icu_4_2::RegexMatcher::MatchAt(int32_t, UBool, UErrorCode&)':
third_party/icu/source/i18n/rematch.cpp:2785: error: unrecognizable insn:
(insn 4627 1680 1682 231 third_party/icu/source/i18n/rematch.cpp:1949 (set (reg:SI 0 r0)
        (plus:SI (reg:SI 3 r3 [1112])
                    (mult:SI (reg/v:SI 14 lr [orig:252 opValue.2394 ] [252])
                                    (const_int 32 [0x20])))) -1 (nil))
third_party/icu/source/i18n/rematch.cpp:2785: internal compiler error: in extract_insn, at recog.c:1990

This can be worked around by building the file at -Os.

Patch by Joel Stanley <joel@jms.id.au>;

BUG=19954

[brettw, 2009-10-10 15:20:10]

Can you add a comment in the gyp file about why this is here like of like the
commit comment?

LGTM otherwise.

[Joel Stanley (old), 2009-10-10 16:19:26]

On 2009/10/10 15:20:10, brettw wrote:
> Can you add a comment in the gyp file about why this is here like of like the
> commit comment?

Done.

> 
> LGTM otherwise.

I don't have commit privs, can you please commit this for me?

Thanks

[Evan Martin, 2009-10-10 17:12:24]

http://codereview.chromium.org/270058/diff/1003/1004
File icu.gyp (right):

http://codereview.chromium.org/270058/diff/1003/1004#newcode216
Line 216: 'cflags': [ '-Os' ],
I'm concerned this change will just be forgotten once it's in.

Have you:
- tried gcc 4.4?
- filed a bug against gcc?
- tried other opt flags/

[Joel Stanley (old), 2009-10-11 03:29:05]

On 2009/10/10 17:12:24, Evan Martin wrote:
> http://codereview.chromium.org/270058/diff/1003/1004
> File icu.gyp (right):
> 
> http://codereview.chromium.org/270058/diff/1003/1004#newcode216
> Line 216: 'cflags': [ '-Os' ],
> I'm concerned this change will just be forgotten once it's in.
> 
> Have you:
> - tried gcc 4.4?

No.  I haven't been successful in building a 4.4 toolchain with crosstool-ng.

> - filed a bug against gcc?

I didn't file a bug myself, but Dean had me send the output of the preprocessor
to Doug (a gcc guy at Google - I don't know any more than that).  The outcome
was a bit vague, but he said that the bug was fixed in later versions.

> - tried other opt flags/

1, 2, and 3 trigger the ICE.  s, 0 and not specifying a optimisation level build
fine.  Building for thumb instead of arm works at all levels.

A problem with my approach is that the added flag gets put in both release and
debug.  I changed it to filter out the known bad flags without adding a new one.

It's still a hack, so I understand if we don't want it in the tree.

[Antoine Labour, 2009-11-13 19:20:17]

http://codereview.chromium.org/270058/diff/4001/5001
File icu.gyp (right):

http://codereview.chromium.org/270058/diff/4001/5001#newcode211
Line 211: ['target_arch=="arm"', {
could you add armv7!=1 ? I never had an issue on armv7.

[Joel Stanley (old), 2009-11-17 01:26:15]

On 2009/11/13 19:20:17, Antoine Labour wrote:
> http://codereview.chromium.org/270058/diff/4001/5001
> File icu.gyp (right):
> 
> http://codereview.chromium.org/270058/diff/4001/5001#newcode211
> Line 211: ['target_arch=="arm"', {
> could you add armv7!=1 ? I never had an issue on armv7.

I've only seen the error with my crosstool built compiler.  I'm now using a
different compiler, so I think we can ignore this unless (until?) someone else
reports the issue.