[PATCH] Fix find_reloads_subreg_address (PR rtl-optimization/42429)

[Jakub Jelinek <jakub at redhat dot com>]

Hi!

As described in the PR, DSE2 removes a needed store into a spill slot
because it thinks it doesn't overlap with the use.
Before reload we have:
(set (reg:QI 5621) ...)
...
(set ... (xor:SI (subreg:SI (reg:QI 5621) 0) ...))
(and &= 255 later on, so the higher bits are ignored).  As the register
pressure is extremely high, pseudo 5621 is spilled:
(set (reg:SI 14 %r14) (const_int 4096 [0x1000]))
(set (mem/c:QI (plus:SI (plus:SI (reg/f:SI 15 %r15) (reg:SI 14 %r14)) (const_int 519 [0x207])) [7 %sfp+-137 S1 A8]) ...)
...
(set (reg:SI 4 %r4) (const_int 4096 [0x1000]))
(set ... (xor:SI ... (mem/c:SI (plus:SI (plus:SI (reg/f:SI 15 %r15) (reg:SI 4 %r4)) (const_int 516 [0x204])) [7 %sfp+-140 S1 A8])))
The spill slot for QImode had MEM_ATTRS [7 %sfp+-137 S1 A8] and
when find_reloads_subreg_address widened that to SImode, it didn't adjust
MEM_SIZE, so nonoverlapping_memrefs_p during DSE2 thinks that
QImode [7 %sfp+-137 S1 A8] and SImode [7 %sfp+-140 S1 A8] can't overlap
(it doesn't consider MEM's mode when MEM_SIZE is set).
Fixed by adjusting also MEM_SIZE.

Attached are two patches, the first attempt where it adjusts MEM_SIZE only
when MEM_OFFSET is set has been bootstrapped/regtested on x86_64-linux/trunk
and {x86_64,i686,powerpc,powerpc64,s390,s390x}-linux on 4.4 branch.  I did
it because I wasn't sure what does MEM_SIZE mean when MEM_OFFSET isn't set,
if it isn't the size of the whole decl.  But after reading adjust_address_1
etc., I believe it should be for non-BLKmode MEMs always the size of the
mode, so probably the second patch (which fixes the testcase too) should be
used instead.  The MEM_ATTRS check is there to avoid creating MEM_ATTRS if
it has none (in that case MEM_SIZE gives GET_MODE_SIZE of the mode, and we
would needlessly create MEM_ATTRS).

Ok for trunk/4.4?

	Jakub

Attachment: Y498
Description: Text document

Attachment: Y498a
Description: Text document