Pass -mtls-use-call when building libraries used in IRT.

Pass -mtls-use-call when building libraries used in IRT.

This adds -mtls-use-call to the compilation options on x86-32 for the
platform, srpc, and ppapi_proxy libraries.  These use TLS (either
themselves or by dint of using newlib's stdout et al) and are linked
into the IRT blob, which must segregate IRT-private TLS uses from user TLS.

We also add a test run_irt_tls_test that does a sanity check with objdump
that no direct-register TLS use leaked into irt.nexe in fact.  Note this
will later need to be extended to check that the ARM irt.nexe does not use
r9 in the same way.

Finally, we change the conditionalization of the building of irt_tls.c now that x86-32 can also use it.  Only bitcode builds are still broken.

BUG=http://code.google.com/p/nativeclient/issues/detail?id=1595
TEST=trybots

Committed: http://src.chromium.org/viewvc/native_client?view=rev&revision=4866

[Roland McGrath, 2011-04-13 19:47:09]

Passing trybots modulo one flaky and the mac10.5 bot having unexplained slowness
(reproduced again) on the objdump test.

[Mark Seaborn, 2011-04-13 20:32:30]

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/check_tls.py
File src/untrusted/irt/check_tls.py (right):

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/check_tls....
src/untrusted/irt/check_tls.py:14: objdump = args[0]
Maybe also do "assert len(args) == 2" (or a more informative check).

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/check_tls....
src/untrusted/irt/check_tls.py:19: print 'TLS access found: %s' % line
Can add some more description to the error?

e.g. "x86-32 direct TLS access found.  This means that a library was linked in
that was not compiled with '-mtls-use-call'."

This error is most likely to be triggered on the trybots when someone changes a
library and is not expecting to see this error.

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/check_tls....
src/untrusted/irt/check_tls.py:20: exit(1)
I think sys.exit is the more correct function to use.

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/nacl.scons
File src/untrusted/irt/nacl.scons (right):

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/nacl.scons...
src/untrusted/irt/nacl.scons:45: files = ['irt_entry.c', 'irt_sbrk.c',
'irt_elf_utils.c', 'irt_tls.c',
Nit: we tend to split lists like this into one-entry-per-line when they need
multiple lines.

http://codereview.chromium.org/6839002/diff/1001/src/untrusted/irt/nacl.scons...
src/untrusted/irt/nacl.scons:53: # Make sure that the linked IRT nexe has no TLS
uses.
"TLS uses" -> "direct TLS uses via %gs"?

[Roland McGrath, 2011-04-13 21:12:12]

Did requested cosmetic improvements.

[robertm, 2011-04-13 21:21:04]

one more cosmetic thing ;-)
(applies to three files)

http://codereview.chromium.org/6839002/diff/3002/src/shared/platform/nacl.scons
File src/shared/platform/nacl.scons (right):

http://codereview.chromium.org/6839002/diff/3002/src/shared/platform/nacl.sco...
src/shared/platform/nacl.scons:9: env.Append(CCFLAGS=['-mtls-use-call'])
# can you motivate this special casing in a brief comment maybe copied from the
CL descriptions.

[Mark Seaborn, 2011-04-13 21:24:41]

http://codereview.chromium.org/6839002/diff/3002/src/untrusted/irt/check_tls.py
File src/untrusted/irt/check_tls.py (right):

http://codereview.chromium.org/6839002/diff/3002/src/untrusted/irt/check_tls....
src/untrusted/irt/check_tls.py:21: print 'This looks like ann x86-32 direct TLS
use.'
typo: "ann"

http://codereview.chromium.org/6839002/diff/3002/src/untrusted/irt/nacl.scons
File src/untrusted/irt/nacl.scons (right):

http://codereview.chromium.org/6839002/diff/3002/src/untrusted/irt/nacl.scons...
src/untrusted/irt/nacl.scons:50: 'irt_tls.c',
It just occurred to me that putting irt_tls.c in unconditionally ought to fail
on ARM (and PNaCl x86-32).  You probably need to conditionalise on
Bit('bitcode').

[Roland McGrath, 2011-04-13 21:38:54]

Done.

[Mark Seaborn, 2011-04-13 21:45:04]

LGTM

[Mark Seaborn, 2011-04-13 23:27:33]

http://codereview.chromium.org/6839002/diff/15/src/untrusted/irt/nacl.scons
File src/untrusted/irt/nacl.scons (right):

http://codereview.chromium.org/6839002/diff/15/src/untrusted/irt/nacl.scons#n...
src/untrusted/irt/nacl.scons:70: if env.Bit('build_x86_32'):
Actually this is not right:  it needs to be
  "env.Bit('build_x86_32') and not env.Bit('bitcode')"

This would have been caught by the trybots.  Sorry I didn't spot it.

This test is failing on the PNaCl bots.  See:
http://build.chromium.org/p/client.nacl.toolchain/console
Specifically:
http://build.chromium.org/p/client.nacl.toolchain/builders/lucid32-toolchain_...

[Roland McGrath, 2011-04-14 00:35:24]

http://codereview.chromium.org/6849008/