rs6000: Fix separate shrink-wrapping for TARGET_MULTIPLE

We cannot use {SAVE,REST}_MULTIPLE and separate shrink-wrapping together,
not without checking when actually emitting the prologue/epilogue that the
registers to save/restore are actually still one contiguous block up to
(and including) 31.  So either:

1) We delay the decision of whether to use lmw/stmw to later;
2) We disallow shrink-wrapping separate (integer) components when those
strategies are selected; or
3) We don't use those strategies if we use separate shrink-wrapping.

This patch does 3).  In the long term it may be best to do 1) instead,
it can be slightly more efficient.

This caused problems on darwin (it is the only config that uses lmw/stmw
instructions by default).


	* config/rs6000/rs6000.c (rs6000_savres_strategy): Do not select
	{SAVE,REST}_MULTIPLE if shrink-wrapping separate components.
	(rs6000_get_separate_components): Assert we do not have those
	strategies selected.

From-SVN: r241297
This commit is contained in:
Segher Boessenkool 2016-10-18 10:00:00 +02:00 committed by Segher Boessenkool
parent 62869a1ca1
commit 8f784f6bde
2 changed files with 14 additions and 1 deletions

View File

@ -1,3 +1,10 @@
2016-10-18 Segher Boessenkool <segher@kernel.crashing.org>
* config/rs6000/rs6000.c (rs6000_savres_strategy): Do not select
{SAVE,REST}_MULTIPLE if shrink-wrapping separate components.
(rs6000_get_separate_components): Assert we do not have those
strategies selected.
2016-10-18 Richard Biener <rguenther@suse.de> 2016-10-18 Richard Biener <rguenther@suse.de>
* tree-ssa-propagate.h (substitute_and_fold): Adjust prototype. * tree-ssa-propagate.h (substitute_and_fold): Adjust prototype.

View File

@ -25518,7 +25518,10 @@ rs6000_savres_strategy (rs6000_stack_t *info,
if (TARGET_MULTIPLE if (TARGET_MULTIPLE
&& !TARGET_POWERPC64 && !TARGET_POWERPC64
&& !(TARGET_SPE_ABI && info->spe_64bit_regs_used) && !(TARGET_SPE_ABI && info->spe_64bit_regs_used)
&& info->first_gp_reg_save < 31) && info->first_gp_reg_save < 31
&& !(flag_shrink_wrap
&& flag_shrink_wrap_separate
&& optimize_function_for_speed_p (cfun)))
{ {
/* Prefer store multiple for saves over out-of-line routines, /* Prefer store multiple for saves over out-of-line routines,
since the store-multiple instruction will always be smaller. */ since the store-multiple instruction will always be smaller. */
@ -27452,6 +27455,9 @@ rs6000_get_separate_components (void)
sbitmap components = sbitmap_alloc (32); sbitmap components = sbitmap_alloc (32);
bitmap_clear (components); bitmap_clear (components);
gcc_assert (!(info->savres_strategy & SAVE_MULTIPLE)
&& !(info->savres_strategy & REST_MULTIPLE));
/* The GPRs we need saved to the frame. */ /* The GPRs we need saved to the frame. */
if ((info->savres_strategy & SAVE_INLINE_GPRS) if ((info->savres_strategy & SAVE_INLINE_GPRS)
&& (info->savres_strategy & REST_INLINE_GPRS)) && (info->savres_strategy & REST_INLINE_GPRS))