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).

Bootstrapped and tested on powerpc64-linux {-m64,-m32}.  I'll commit it
if Iain's testing (on darwin) also succeeds.


Segher


2016-10-17  Segher Boessenkool  <seg...@kernel.crashing.org>

        * config/rs6000/rs6000.c: Add include of shrink-wrap.h .
        (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.

---
 gcc/config/rs6000/rs6000.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index df48980..83a7139 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -64,6 +64,7 @@
 #include "target-globals.h"
 #include "builtins.h"
 #include "context.h"
+#include "shrink-wrap.h"
 #include "tree-pass.h"
 #if TARGET_XCOFF
 #include "xcoffout.h"  /* get declarations of xcoff_*_section_name */
@@ -25506,7 +25507,10 @@ rs6000_savres_strategy (rs6000_stack_t *info,
   if (TARGET_MULTIPLE
       && !TARGET_POWERPC64
       && !(TARGET_SPE_ABI && info->spe_64bit_regs_used)
-      && info->first_gp_reg_save < 31)
+      && info->first_gp_reg_save < 31
+      && !(SHRINK_WRAPPING_ENABLED
+          && flag_shrink_wrap_separate
+          && optimize_function_for_speed_p (cfun)))
     {
       /* Prefer store multiple for saves over out-of-line routines,
         since the store-multiple instruction will always be smaller.  */
@@ -27440,6 +27444,9 @@ rs6000_get_separate_components (void)
   sbitmap components = sbitmap_alloc (32);
   bitmap_clear (components);
 
+  gcc_assert (!(info->savres_strategy & SAVE_MULTIPLE)
+             && !(info->savres_strategy & REST_MULTIPLE));
+
   /* The GPRs we need saved to the frame.  */
   if ((info->savres_strategy & SAVE_INLINE_GPRS)
       && (info->savres_strategy & REST_INLINE_GPRS))
-- 
1.9.3

Reply via email to