gcc.c-torture/execute/990208-1.c test failure

[Michael Hope]

Seen in gcc-linaro-4.4+bzr93543 on armv7l only. The gcc testsuite reports:
 FAIL: gcc.c-torture/execute/990208-1.c execution, -Os

See
 http://ex.seabright.co.nz/build/gcc-linaro-4.4-93543/logs/armv7l-maverick-cbuild1-pavo1/gcc-test.txt

[Yao Qi (yao-codesourcery)]

Looks like doit is *not* inlined properly. Root cause is here,

congfig/arm/arm.c:arm_optimization_options(int level, int size ATTRIBUTE_UNUSED)
...
  if (size)
    {
               /* In Thumb mode the function call code size overhead is typically very
  small, and narrow branch instructions have very limited range.
  Inlining even medium sized functions tends to bloat the caller and
  require the use of long branch instructions. On average the long
  branches cost more than eliminating the function call overhead saves,
  so we use extremely restrictive automatic inlining heuristics. In ARM
  mode the results are fairly neutral, probably due to better constant
  pool placement. */
      set_param_value ("max-inline-insns-single", 1);
      set_param_value ("max-inline-insns-auto", 1);
......
    }

Remove two 'set_param_value', problem goes away.

[Yao Qi (yao-codesourcery)]

Hmm... execute/990208-1.x shows that this case should not be run with -Os on ARM.

# On ARM, with -Os, some of the functions that this test
# expects to be inlined are not inlined for code size
# reasons. This is not a bug, it's intentional,
# so stop this test from running.
set torture_eval_before_compile {
  if { [istarget "arm-*-*"] && [string match {*-Os*} "$option"] } {
    continue
  }
}

Target triplet should be "arm*-*-*", which is more general. Configuration 'target=armv7l-unknown-linux-gnueabi' exposes this bug.

[Yao Qi (yao-codesourcery)]

execute/990208-1.x is introduced by this patch,

2009-05-28 Julian Brown <email address hidden>

        Merged from Sourcery G++ 4.3:

        2007-09-19 Vladimir Prus <email address hidden>

        gcc/
        * config/arm/arm.c (arm_optimization_options): Disable
        -fmove-loop-invariants. Use very restrictive inlining heuristics.

        gcc/testsuite/
        * gcc.c-torture/execute/bcp-1.x: New. Don't run bcp-1.c test on arm,
        with -Os.
        * gcc.c-torture/execute/990208-1.x: New. Likewise.

[Michael Hope (michaelh1)]

Related: lp:gcc-linaro/4.4,revno=93544

[Andrew Stubbs (ams-codesourcery)]

Related: lp:gcc-linaro/4.4,revno=93545

[Andrew Stubbs (ams-codesourcery)]

Related: lp:gcc-linaro/4.4,revno=93546
Related: lp:gcc-linaro/4.4,revno=93547

[Yao Qi (yao-codesourcery)]

My patch for "arm target triplet" fix was committed to gcc,
http://gcc.gnu.org/ml/gcc-patches/2010-08/msg00665.html

However, this patch for this bug can't be upstreamable, because gcc.c-torture/execute/990208-1.x is in linaro tree only. What can we do here?

[Yao Qi (yao-codesourcery)]

After talked with Andrew, we agree to mark it as "Fix Committed", because it is not upstreamable.

[Andrew Stubbs (ams-codesourcery)]

Just to be clear, it is 'Fix Committed', not 'Won't Fix', because the rest of the patch has been committed. The remaining part is not so much non-upstreamable, as not-required in later versions of gcc.