[Bug rtl-optimization/48596] [4.7 Regression] [SH] unable to find a register to spill in class 'FPUL_REGS'

[kkojima at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48596

--- Comment #5 from Kazumoto Kojima <kkojima at gcc dot gnu.org> 2011-08-02 23:53:30 UTC ---
I was trying to find a way that solves it without penalizing -O2
or the higher cases, though it's not easy to me.  It seems that
the target's register_move_cost is the way to discourage trying
to use FP registers for a pointer.  Unfortunately, Pmode is simply
SImode for our case and it also discourages using a FP reg as
a cheap storage for SImode.  I've tried

--- ORIG/trunk/gcc/config/sh/sh.c    2011-08-01 09:22:27.000000000 +0900
+++ trunk/gcc/config/sh/sh.c    2011-08-01 09:41:25.000000000 +0900
@@ -11472,8 +11472,18 @@ sh_register_move_cost (enum machine_mode
        && REGCLASS_HAS_GENERAL_REG (srcclass))
       || (REGCLASS_HAS_GENERAL_REG (dstclass)
       && REGCLASS_HAS_FP_REG (srcclass)))
-    return ((TARGET_SHMEDIA ? 4 : TARGET_FMOVD ? 8 : 12)
-        * ((GET_MODE_SIZE (mode) + 7) / 8U));
+    {
+      if (TARGET_SHMEDIA)
+    return 4 * ((GET_MODE_SIZE (mode) + 7) / 8U);
+      else
+    {
+      /* Discourage trying to use fp regs for a pointer.  */
+      int addend = (mode == Pmode) ? 40 : 0;
+
+      return (((TARGET_FMOVD ? 8 : 12) + addend)
+          * ((GET_MODE_SIZE (mode) + 7) / 8U));
+    }
+    }

   if ((dstclass == FPUL_REGS
        && REGCLASS_HAS_GENERAL_REG (srcclass))

on the current trunk and observed some CSiBE testresults.  A bit
surprisingly, there are no code size regressions and one 2%
improvement for teem-1.6.0-src src/bane/gkmsTxf which reduces
to 3192 bytes from 3256 bytes.  Now I'm inclined to apply it
on trunk if it passes the bootstrap/regression/other tests.