* [PATCH] optabs: Don't reuse target for multi-word expansions if it overlaps operand(s) [PR97073]
@ 2020-09-18 8:30 Jakub Jelinek
2020-09-26 12:03 ` Eric Botcazou
0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2020-09-18 8:30 UTC (permalink / raw)
To: Richard Biener, Eric Botcazou, Jeff Law; +Cc: gcc-patches
Hi!
The following testcase is miscompiled on i686-linux, because
we try to expand a double-word bitwise logic operation with op0
being a (mem:DI u) and target (mem:DI u+4), i.e. partial overlap, and
thus end up with:
movl 4(%esp), %eax
andl u, %eax
movl %eax, u+4
! movl u+4, %eax optimized out
andl 8(%esp), %eax
movl %eax, u+8
rather than with the desired:
movl 4(%esp), %edx
movl 8(%esp), %eax
andl u, %edx
andl u+4, %eax
movl %eax, u+8
movl %edx, u+4
because the store of the first word to target overwrites the second word of
the operand.
expand_binop for this (and several similar places) already check for target
== op0 or target == op1, this patch just adds reg_overlap_mentioned_p calls
next to it.
Pedantically, at least for some of these it might be sufficient to force
a different target if there is overlap but target is not rtx_equal_p to
the operand (e.g. in this bitwise logical case, but e.g. not in the shift
cases where there is reordering), though that would go against the
preexisting target == op? checks and the rationale that REG_EQUAL notes in
that case isn't correct.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and
release branches?
2020-09-18 Jakub Jelinek <jakub@redhat.com>
PR middle-end/97073
* optabs.c (expand_binop, expand_absneg_bit, expand_unop,
expand_copysign_bit): Check reg_overlap_mentioned_p between target
and operand(s) and if it returns true, force a pseudo as target.
* gcc.c-torture/execute/pr97073.c: New test.
--- gcc/optabs.c.jj 2020-07-28 15:39:09.000000000 +0200
+++ gcc/optabs.c 2020-09-17 22:07:53.238396458 +0200
@@ -1395,6 +1395,8 @@ expand_binop (machine_mode mode, optab b
if (target == 0
|| target == op0
|| target == op1
+ || reg_overlap_mentioned_p (target, op0)
+ || reg_overlap_mentioned_p (target, op1)
|| !valid_multiword_target_p (target))
target = gen_reg_rtx (int_mode);
@@ -1475,6 +1477,8 @@ expand_binop (machine_mode mode, optab b
if (target == 0
|| target == op0
|| target == op1
+ || reg_overlap_mentioned_p (target, op0)
+ || reg_overlap_mentioned_p (target, op1)
|| !valid_multiword_target_p (target))
target = gen_reg_rtx (int_mode);
@@ -1533,6 +1537,8 @@ expand_binop (machine_mode mode, optab b
|| target == op0
|| target == op1
|| !REG_P (target)
+ || reg_overlap_mentioned_p (target, op0)
+ || reg_overlap_mentioned_p (target, op1)
|| !valid_multiword_target_p (target))
target = gen_reg_rtx (int_mode);
@@ -2670,6 +2676,7 @@ expand_absneg_bit (enum rtx_code code, s
if (target == 0
|| target == op0
+ || reg_overlap_mentioned_p (target, op0)
|| (nwords > 1 && !valid_multiword_target_p (target)))
target = gen_reg_rtx (mode);
@@ -2951,7 +2958,10 @@ expand_unop (machine_mode mode, optab un
int i;
rtx_insn *insns;
- if (target == 0 || target == op0 || !valid_multiword_target_p (target))
+ if (target == 0
+ || target == op0
+ || reg_overlap_mentioned_p (target, op0)
+ || !valid_multiword_target_p (target))
target = gen_reg_rtx (int_mode);
start_sequence ();
@@ -3472,6 +3482,8 @@ expand_copysign_bit (scalar_float_mode m
if (target == 0
|| target == op0
|| target == op1
+ || reg_overlap_mentioned_p (target, op0)
+ || reg_overlap_mentioned_p (target, op1)
|| (nwords > 1 && !valid_multiword_target_p (target)))
target = gen_reg_rtx (mode);
--- gcc/testsuite/gcc.c-torture/execute/pr97073.c.jj 2020-09-17 22:00:39.778614611 +0200
+++ gcc/testsuite/gcc.c-torture/execute/pr97073.c 2020-09-17 22:06:08.870893863 +0200
@@ -0,0 +1,21 @@
+/* PR middle-end/97073 */
+/* { dg-additional-options "-mno-stv" { target i?86-*-* x86_64-*-* } } */
+
+typedef unsigned long long L;
+union U { L i; struct T { unsigned k; L l; } j; } u;
+
+__attribute__((noinline,noclone)) void
+foo (L x)
+{
+ u.j.l = u.i & x;
+}
+
+int
+main ()
+{
+ u.i = 5;
+ foo (-1ULL);
+ if (u.j.l != 5)
+ __builtin_abort ();
+ return 0;
+}
Jakub
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH] optabs: Don't reuse target for multi-word expansions if it overlaps operand(s) [PR97073]
2020-09-18 8:30 [PATCH] optabs: Don't reuse target for multi-word expansions if it overlaps operand(s) [PR97073] Jakub Jelinek
@ 2020-09-26 12:03 ` Eric Botcazou
0 siblings, 0 replies; 2+ messages in thread
From: Eric Botcazou @ 2020-09-26 12:03 UTC (permalink / raw)
To: Jakub Jelinek; +Cc: Richard Biener, Jeff Law, gcc-patches
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and
> release branches?
>
> 2020-09-18 Jakub Jelinek <jakub@redhat.com>
>
> PR middle-end/97073
> * optabs.c (expand_binop, expand_absneg_bit, expand_unop,
> expand_copysign_bit): Check reg_overlap_mentioned_p between target
> and operand(s) and if it returns true, force a pseudo as target.
>
> * gcc.c-torture/execute/pr97073.c: New test.
This looks good to me.
--
Eric Botcazou
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2020-09-26 12:03 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2020-09-18 8:30 [PATCH] optabs: Don't reuse target for multi-word expansions if it overlaps operand(s) [PR97073] Jakub Jelinek
2020-09-26 12:03 ` Eric Botcazou
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).