[Bug tree-optimization/120702] Extraneous string constant at -Os

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120702

--- Comment #9 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think we can retain all important optimization by restricting the aggressive
replacement by constant to where it is likely creating followup optimization
opportunities - which is for register type replacements.

I'll note that size wins are not always obvious:

000000000000001f <benchmark>:
  1f:   48 8b 05 00 00 00 00    mov    0x0(%rip),%rax        # 26
<benchmark+0x7>
  26:   48 89 05 00 00 00 00    mov    %rax,0x0(%rip)        # 2d
<benchmark+0xe>
  2d:   8b 05 00 00 00 00       mov    0x0(%rip),%eax        # 33
<benchmark+0x14>
  33:   66 89 05 00 00 00 00    mov    %ax,0x0(%rip)        # 3a
<benchmark+0x1b>
  3a:   31 c0                   xor    %eax,%eax
  3c:   c3                      ret

vs.

000000000000001f <benchmark>:
  1f:   48 b8 4a 32 4f 5a 46    movabs $0x463035465a4f324a,%rax
  26:   35 30 46 
  29:   66 c7 05 00 00 00 00    movw   $0x4c59,0x0(%rip)        # 32
<benchmark+0x13>
  30:   59 4c 
  32:   48 89 05 00 00 00 00    mov    %rax,0x0(%rip)        # 39
<benchmark+0x1a>
  39:   31 c0                   xor    %eax,%eax
  3b:   c3                      ret

so for the attached testcase the inline copy is 1 byte shorter.  I'll
note this can be still inline expanded at RTL expansion (block-copy)
or possibly even later.

I'll note the other internal improvement to disallow &STRING_CST in GIMPLE
by introducing CONST_DECLs cannot fix this issue given the STRING_CSTs
can be part of a constant initializer of a larger aggregate that we may
not be able to elide.  That might also be a correctness issue with regard
to the address of such parts.  Not constant folding a string part of such
initializer might also keep that whole variable live only for a possibly
small part.

And I bet not constant folding the STRING_CSTs will break diagnostic
testcases relying on such constant folding (moderately easy to fix
by slapping fold_const_aggregate_ref in those places).

Testing a trivial patch.