[PATCH] PR52528, combine fix

[Chung-Lin Tang <cltang at codesourcery dot com>]

Hi,

As described in the PR, a testcase compiled for PowerPC:
struct S {
  unsigned a : 30;
  unsigned b :  2;
};

int foo (int b)
{
  struct S s = {0};
  s.b = b;
  return bar (0x000b0010, 0x00040100ULL, *(unsigned long *)&s);
}

currently this is compiled to:
foo:
        lis 6,0x4
        li 5,0
        ori 6,6,256
        li 7,0
        crxor 6,6,6
        b bar

Notice the incorrect code generated: no construction of the 1st arg (reg
3), and wrong code for the 3rd arg (reg 7)

The problem seems to be in combine, during calls from try_combine() to
can_combine_p():

  can_combine_p() has a call to expand_field_assignment(), which may
call get_last_value() during its simplification operations (through the
reg_nonzero_bits_for_combine() hook); not setting subst_low_luid
properly affects its correctness.

So the fix is a one-liner that sets subst_low_luid before the
expand_field_assignment() call. Bootstrapped and tested under i686,
x86-64, powerpc64. Cross-tested on ARM. I was a bit weary that some
optimization regression might appear, which will complicate things, but
everything looks fine.

I have a larger (customer provided) testcase that exposed this bug after
rev.161655 (the mem-ref2 merge, may be related to effects on bitfields).
So if suitable, please also approve this patch for 4.6/4.7 branches.

Thanks,
Chung-Lin

2012-03-10  Chung-Lin Tang  <cltang@codesourcery.com>

	PR rtl-optimization/52528
	* combine.c (can_combine_p): Add setting of subst_low_luid
	before call to expand_field_assignment().

Attachment: pr52528.diff
Description: Text document