[PATCH, rtl-optimization] Fix PR63475, Postreload CSE propagates aliased memory operand

[Uros Bizjak <ubizjak at gmail dot com>]

Hello!

Attached patch fixes PR63475, where postreload CSE propagates aliased
memory operand.

The core of the problem was with the call to base_alias_check when
VALUE RTXes are involved. Before the call, find_base_term is used to
extract the base of x_addr and mem_addr. Please note that
find_base_term is able to extract the bases from VALUE RTXes. These
extracted bases were passed to base_alias_check, together with
original VALUE RTXes x_addr and mem_addr.

The problem begins here. base_alias_check doesn't handle VALUE RTXes,
and uses e.g. canon_rtx on VALUEs and various GET_CODE accessors to
determine various properties of passed x_addr and mem_addr. One of
these check checks for the AND alignment addresses to prevent:

  /* Differing symbols not accessed via AND never alias.  */
  if (GET_CODE (x_base) != ADDRESS && GET_CODE (y_base) != ADDRESS)
    return 0;

early exit. However, when x and y are passed as VALUE RTXes (that
corresponds and hides the address with AND), and preceding calls to
find_base_term are nevertheless able to extract the bases of x and y,
this condition fires erroneously and invalid return value is returned
(with 0 meaning that the addresses X and Y are known to point to
different objects).

The solution is to always extract values for x_addr and mem_addr and
use them in the calls to find_base_term and base_alias_check.

[It can happen that get_addr is not able to match VALUE RTX with some
address, so it is not possible to simply add a bunch of "GET_CODE (x)
!= VALUE" asserts in base_alias_check. But in this case find_base_term
returns ADDRESS RTX, so we stay in sync as far as base_alias_check is
concerned (see the quoted code above).]

Added benefit of the patch is, that canon_rtx now works as expected.
canon_rtx does NOT handle VALUE RTXes.

A small optimization is also present. If the address is already
canonicalized, we pass original address to memrefs_conflict_p, but we
have to extract original address for preceding functions nevertheless.
Also, we use extracted original address in recently added check for
AND aligned addresses when checking for MEM_READONLY_P.

The patch also removes a couple of unneeded and unused calls to
canon_rtx, also to show the level of bitrot in this area ...

2014-10-14  Uros Bizjak  <ubizjak@gmail.com>

    PR rtl-optimization/63475
    * alias.c (true_dependence_1): Always use get_addr to extract
    true address operands from x_addr and mem_addr.  Use extracted
    address operands to check for references with alignment ANDs.
    Use extracted address operands with find_base_term and
    base_alis_check. For noncanonicalized operands call canon_rtx with
    extracted address operand.
    (write_dependence_1): Ditto.
    (may_alias_p): Ditto.  Remove unused calls to canon_rtx.

Patch was thoroughly tested on x86_64-linux-gnu {,-m32} and
alpha-linux-gnu for all default languages plus obj-c++ and go. While
there was no differences on x86_64-linux-gnu (as expected),
alpha-linux-gnu improved the result [1] for some hundred of PASSes in
gfortran testsuite [2].

OK for mainline?

[1] https://gcc.gnu.org/ml/gcc-testresults/2014-10/msg01151.html
[2] https://gcc.gnu.org/ml/gcc-testresults/2014-10/msg01478.html

Uros.

Attachment: alias-7.diff.txt
Description: Text document