* [PATCH] Fix regmove REG_EQUAL handling (PR rtl-optimization/46804)
@ 2010-12-08 16:47 Jakub Jelinek
2010-12-10 16:57 ` Jeff Law
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2010-12-08 16:47 UTC (permalink / raw)
To: gcc-patches
Hi!
Bernd in 2001 fixed handling of REG_EQUIV and attempted to fix handling of
REG_EQUAL in optimize_reg_copy_3, unfortunately for the latter it wasn't
right - p at the end of the function is no longer the setter of src_reg,
but, as we walked again forwards, insn. We don't want to drop REG_EQUAL
from insn, but from the setter.
This patch fixes it, and also, instead of dropping REG_EQUAL, tries to
replace it with sign or zero extension of it if the REG_EQUAL note was
the same as the original SET_SRC.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2010-12-08 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/46804
* regmove.c (optimize_reg_copy_3): Look for REG_EQUAL note
on the setter of src_reg rather than on insn. If it is
equal to the setter's original SET_SRC, replace it with its
zero or sign extension instead of dropping it.
* gfortran.dg/pr46804.f90: New test.
--- gcc/regmove.c.jj 2010-12-02 11:51:31.000000000 +0100
+++ gcc/regmove.c 2010-12-08 12:36:46.000000000 +0100
@@ -514,7 +514,7 @@ optimize_reg_copy_3 (rtx insn, rtx dest,
rtx src_reg = XEXP (src, 0);
int src_no = REGNO (src_reg);
int dst_no = REGNO (dest);
- rtx p, set;
+ rtx p, set, set_insn;
enum machine_mode old_mode;
basic_block bb = BLOCK_FOR_INSN (insn);
@@ -552,6 +552,7 @@ optimize_reg_copy_3 (rtx insn, rtx dest,
GET_MODE_BITSIZE (GET_MODE (src_reg))))
return;
+ set_insn = p;
old_mode = GET_MODE (src_reg);
PUT_MODE (src_reg, GET_MODE (src));
XEXP (src, 0) = SET_SRC (set);
@@ -584,9 +585,19 @@ optimize_reg_copy_3 (rtx insn, rtx dest,
}
else
{
- rtx note = find_reg_note (p, REG_EQUAL, NULL_RTX);
+ rtx note = find_reg_note (set_insn, REG_EQUAL, NULL_RTX);
if (note)
- remove_note (p, note);
+ {
+ if (rtx_equal_p (XEXP (note, 0), XEXP (src, 0)))
+ {
+ XEXP (note, 0)
+ = gen_rtx_fmt_e (GET_CODE (src), GET_MODE (src),
+ XEXP (note, 0));
+ df_notes_rescan (set_insn);
+ }
+ else
+ remove_note (set_insn, note);
+ }
}
}
--- gcc/testsuite/gfortran.dg/pr46804.f90.jj 2010-12-08 12:41:21.000000000 +0100
+++ gcc/testsuite/gfortran.dg/pr46804.f90 2010-12-08 12:41:02.000000000 +0100
@@ -0,0 +1,36 @@
+! PR rtl-optimization/46804
+! { dg-do run }
+! { dg-options "-O -fPIC -fexpensive-optimizations -fgcse -foptimize-register-move -fpeel-loops -fno-tree-loop-optimize" }
+
+program main
+ integer, parameter :: n1 = 2, n2 = 3, n3 = 4, slen = 3
+ character (len = slen), dimension (n1, n2, n3) :: a
+ integer (kind = 1), dimension (2, 4) :: shift1
+ integer (kind = 2), dimension (2, 4) :: shift2
+ integer (kind = 4), dimension (2, 4) :: shift3
+ do i3 = 1, n3
+ do i2 = 1, n2
+ do i1 = 1, n1
+ a (i1, i2, i3) = 'ab'(i1:i1) // 'cde'(i2:i2) // 'fghi'(i3:i3)
+ end do
+ end do
+ end do
+ shift1 (1, :) = (/ 4, 11, 19, 20 /)
+ shift1 (2, :) = (/ 55, 5, 1, 2 /)
+ shift2 = shift1
+ shift3 = shift1
+ call test (cshift (a, shift2, 2))
+ call test (cshift (a, shift3, 2))
+contains
+ subroutine test (b)
+ character (len = slen), dimension (n1, n2, n3) :: b
+ do i3 = 1, n3
+ do i2 = 1, n2
+ do i1 = 1, n1
+ i2p = mod (shift1 (i1, i3) + i2 - 1, n2) + 1
+ if (b (i1, i2, i3) .ne. a (i1, i2p, i3)) call abort
+ end do
+ end do
+ end do
+ end subroutine test
+end program main
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] Fix regmove REG_EQUAL handling (PR rtl-optimization/46804)
2010-12-08 16:47 [PATCH] Fix regmove REG_EQUAL handling (PR rtl-optimization/46804) Jakub Jelinek
@ 2010-12-10 16:57 ` Jeff Law
0 siblings, 0 replies; 2+ messages in thread
From: Jeff Law @ 2010-12-10 16:57 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: gcc-patches
On 12/08/10 09:42, Jakub Jelinek wrote:
> Hi!
>
> Bernd in 2001 fixed handling of REG_EQUIV and attempted to fix handling of
> REG_EQUAL in optimize_reg_copy_3, unfortunately for the latter it wasn't
> right - p at the end of the function is no longer the setter of src_reg,
> but, as we walked again forwards, insn. We don't want to drop REG_EQUAL
> from insn, but from the setter.
>
> This patch fixes it, and also, instead of dropping REG_EQUAL, tries to
> replace it with sign or zero extension of it if the REG_EQUAL note was
> the same as the original SET_SRC.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
>
> 2010-12-08 Jakub Jelinek<jakub@redhat.com>
>
> PR rtl-optimization/46804
> * regmove.c (optimize_reg_copy_3): Look for REG_EQUAL note
> on the setter of src_reg rather than on insn. If it is
> equal to the setter's original SET_SRC, replace it with its
> zero or sign extension instead of dropping it.
>
> * gfortran.dg/pr46804.f90: New test.
OK. Thanks.
jeff
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2010-12-10 16:33 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2010-12-08 16:47 [PATCH] Fix regmove REG_EQUAL handling (PR rtl-optimization/46804) Jakub Jelinek
2010-12-10 16:57 ` Jeff Law
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).