[Bug target/95798] [10/11 Regression] Initialization code --- suboptimal

["cvs-commit at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95798

--- Comment #7 from CVS Commits <cvs-commit at gcc dot gnu.org> ---
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:880682e7b2348d66f4089fa4af102b69eaaefbc2

commit r11-7384-g880682e7b2348d66f4089fa4af102b69eaaefbc2
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Thu Feb 25 10:22:53 2021 +0100

    match.pd: Use :s for (T)(A) + CST -> (T)(A + CST) [PR95798]

    The r10-2806 change regressed following testcases, instead of doing
    int -> unsigned long sign-extension once and then add 8, 16, ... 56 to it
    for each of the memory access, it adds 8, 16, ... 56 in int mode and then
    sign extends each.  So that means:
    +       movq    $0, (%rsp,%rax,8)
    +       leal    1(%rdx), %eax
    +       cltq
    +       movq    $1, (%rsp,%rax,8)
    +       leal    2(%rdx), %eax
    +       cltq
    +       movq    $2, (%rsp,%rax,8)
    +       leal    3(%rdx), %eax
    +       cltq
    +       movq    $3, (%rsp,%rax,8)
    +       leal    4(%rdx), %eax
    +       cltq
    +       movq    $4, (%rsp,%rax,8)
    +       leal    5(%rdx), %eax
    +       cltq
    +       movq    $5, (%rsp,%rax,8)
    +       leal    6(%rdx), %eax
    +       addl    $7, %edx
    +       cltq
    +       movslq  %edx, %rdx
    +       movq    $6, (%rsp,%rax,8)
    +       movq    $7, (%rsp,%rdx,8)
    -       movq    $0, (%rsp,%rdx,8)
    -       movq    $1, 8(%rsp,%rdx,8)
    -       movq    $2, 16(%rsp,%rdx,8)
    -       movq    $3, 24(%rsp,%rdx,8)
    -       movq    $4, 32(%rsp,%rdx,8)
    -       movq    $5, 40(%rsp,%rdx,8)
    -       movq    $6, 48(%rsp,%rdx,8)
    -       movq    $7, 56(%rsp,%rdx,8)
    GCC 9 -> 10 change or:
    -       movq    $0, (%rsp,%rdx,8)
    -       movq    $1, 8(%rsp,%rdx,8)
    -       movq    $2, 16(%rsp,%rdx,8)
    -       movq    $3, 24(%rsp,%rdx,8)
    -       movq    $4, 32(%rsp,%rdx,8)
    -       movq    $5, 40(%rsp,%rdx,8)
    -       movq    $6, 48(%rsp,%rdx,8)
    -       movq    $7, 56(%rsp,%rdx,8)
    +       movq    $0, (%rsp,%rax,8)
    +       leal    1(%rdx), %eax
    +       movq    $1, (%rsp,%rax,8)
    +       leal    2(%rdx), %eax
    +       movq    $2, (%rsp,%rax,8)
    +       leal    3(%rdx), %eax
    +       movq    $3, (%rsp,%rax,8)
    +       leal    4(%rdx), %eax
    +       movq    $4, (%rsp,%rax,8)
    +       leal    5(%rdx), %eax
    +       movq    $5, (%rsp,%rax,8)
    +       leal    6(%rdx), %eax
    +       movq    $6, (%rsp,%rax,8)
    +       leal    7(%rdx), %eax
    +       movq    $7, (%rsp,%rax,8)
    change on the other test.  While for the former case of
    int there is due to signed integer overflow (unless -fwrapv)
    the possibility to undo it e.g. during expansion, for the unsigned
    case information is unfortunately lost.

    The following patch adds :s to the convert which restores these
    testcases but keeps the testcases the patch meant to improve as is.

    2021-02-25  Jakub Jelinek  <jakub@redhat.com>

            PR target/95798
            * match.pd ((T)(A) + CST -> (T)(A + CST)): Add :s to convert.

            * gcc.target/i386/pr95798-1.c: New test.
            * gcc.target/i386/pr95798-2.c: New test.