[PATCH] Avoid shifts using hard regs from likely spilled classes for BLKmode returning calls (PR rtl-optimization/41646)
Richard Guenther
richard.guenther@gmail.com
Fri Oct 9 19:02:00 GMT 2009
On Fri, Oct 9, 2009 at 8:40 PM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> This testcase ICEs on x86_64-linux, because in expand copy_blkmode_from_reg
> is called with %rax and thus a hard register is used in several shift insns.
> Later on combiner extends lifetime of that register across multiplication
> insns that require this register during reload and reload ICEs.
>
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux.
> Ok for trunk?
Ok.
Thanks,
Richard.
> 2009-10-09 Jakub Jelinek <jakub@redhat.com>
>
> PR rtl-optimization/41646
> * calls.c (expand_call): For BLKmode types returned in registers
> avoid likely spilled hard regs in copy_blkmode_from_reg generated
> insns.
>
> * gcc.c-torture/compile/pr41646.c: New test.
>
> --- gcc/calls.c.jj 2009-09-23 00:57:18.000000000 +0200
> +++ gcc/calls.c 2009-10-09 17:25:32.000000000 +0200
> @@ -3014,7 +3014,10 @@ expand_call (tree exp, rtx target, int i
> }
> else if (TYPE_MODE (rettype) == BLKmode)
> {
> - target = copy_blkmode_from_reg (target, valreg, rettype);
> + rtx val = valreg;
> + if (GET_MODE (val) != BLKmode)
> + val = avoid_likely_spilled_reg (val);
> + target = copy_blkmode_from_reg (target, val, rettype);
>
> /* We can not support sibling calls for this case. */
> sibcall_failure = 1;
> --- gcc/testsuite/gcc.c-torture/compile/pr41646.c.jj 2009-10-09 17:29:51.000000000 +0200
> +++ gcc/testsuite/gcc.c-torture/compile/pr41646.c 2009-10-09 17:30:17.000000000 +0200
> @@ -0,0 +1,28 @@
> +/* PR rtl-optimization/41646 */
> +
> +struct A { unsigned long a; };
> +struct B { unsigned short b, c, d; };
> +struct B bar (unsigned long);
> +
> +char *
> +foo (char *a, struct A *x)
> +{
> + struct B b = bar (x->a);
> + unsigned char c;
> + unsigned short d;
> + a[3] = ((unsigned char) (b.b % 10) + 48);
> + d = b.b / 10;
> + a[2] = ((unsigned char) (d % 10) + 48);
> + d = d / 10;
> + a[1] = ((unsigned char) (d % 10) + 48);
> + a[0] = ((unsigned char) ((d / 10) % 10) + 48);
> + a[4] = 46;
> + c = (unsigned char) b.c;
> + a[6] = (c % 10 + 48);
> + a[5] = ((c / 10) % 10 + 48);
> + a[7] = 46;
> + c = b.d;
> + a[9] = (c % 10 + 48);
> + a[8] = ((c / 10) % 10 + 48);
> + return a + 10;
> +}
>
> Jakub
>
More information about the Gcc-patches
mailing list