Hi Vlad,
Vladimir Makarov <vmaka...@redhat.com> writes:
> The following patch fixes
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87305
>
> The patch was bootstrapped and tested on x86-64 and ppc64 (be).
>
> Committed as rev. 267823.
>
> Index: ChangeLog
> ===================================================================
> --- ChangeLog (revision 267822)
> +++ ChangeLog (working copy)
> @@ -1,3 +1,12 @@
> +2019-01-10 Vladimir Makarov <vmaka...@redhat.com>
> +
> + PR rtl-optimization/87305
> + * lra-assigns.c
> + (setup_live_pseudos_and_spill_after_risky_transforms): Check
> + allocation for big endian pseudos used as paradoxical subregs and
> + spill them if it is wrong.
> + * lra-constraints.c (lra_constraints): Add a comment.
> +
> 2019-01-10 Richard Biener <rguent...@suse.de>
>
> PR tree-optimization/88792
> Index: lra-assigns.c
> ===================================================================
> --- lra-assigns.c (revision 267822)
> +++ lra-assigns.c (working copy)
> @@ -1146,12 +1146,12 @@ static void
> setup_live_pseudos_and_spill_after_risky_transforms (bitmap
> spilled_pseudo_bitmap)
> {
> - int p, i, j, n, regno, hard_regno;
> + int p, i, j, n, regno, hard_regno, biggest_nregs, nregs_diff;
> unsigned int k, conflict_regno;
> poly_int64 offset;
> int val;
> HARD_REG_SET conflict_set;
> - machine_mode mode;
> + machine_mode mode, biggest_mode;
> lra_live_range_t r;
> bitmap_iterator bi;
> int max_regno = max_reg_num ();
> @@ -1166,8 +1166,26 @@ setup_live_pseudos_and_spill_after_risky
> for (n = 0, i = FIRST_PSEUDO_REGISTER; i < max_regno; i++)
> if ((pic_offset_table_rtx == NULL_RTX
> || i != (int) REGNO (pic_offset_table_rtx))
> - && reg_renumber[i] >= 0 && lra_reg_info[i].nrefs > 0)
> - sorted_pseudos[n++] = i;
> + && (hard_regno = reg_renumber[i]) >= 0 && lra_reg_info[i].nrefs > 0)
> + {
> + biggest_mode = lra_reg_info[i].biggest_mode;
> + biggest_nregs = hard_regno_nregs (hard_regno, biggest_mode);
> + nregs_diff = (biggest_nregs
> + - hard_regno_nregs (hard_regno, PSEUDO_REGNO_MODE (i)));
> + enum reg_class rclass = lra_get_allocno_class (i);
> +
> + if (WORDS_BIG_ENDIAN
> + && (hard_regno - nregs_diff < 0
> + || !TEST_HARD_REG_BIT (reg_class_contents[rclass],
> + hard_regno - nregs_diff)))
> + {
> + /* Hard registers of paradoxical sub-registers are out of
> + range of pseudo register class. Spill the pseudo. */
> + reg_renumber[i] = -1;
> + continue;
> + }
I think for !WORDS_BIG_ENDIAN the equivalent problem to:
|| !TEST_HARD_REG_BIT (reg_class_contents[rclass],
hard_regno - nregs_diff)))
would be:
|| !TEST_HARD_REG_BIT (reg_class_contents[rclass],
hard_regno + biggest_nregs - 1)))
Where does that little-endian check happen? I wasn't sure why
"biggest_mode goes outside the class" problems only occured here
for big-endian.
Thanks,
Richard
