[backport gcc-4.7/trunk r182500 <mikpe@it.uu.se>]

From: Richard Sandiford <rdsandiford at googlemail dot com>
Subject: Invalid hard-reg decomposition in lower-subreg
Date: Sat, 17 Dec 2011 11:26:11 +0000
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>

lower-subreg.c:can_decompose_p uses the following condition to test
whether a multiword hard register can be decomposed into words:

	return (validate_subreg (word_mode, GET_MODE (x), x, UNITS_PER_WORD)
		&& HARD_REGNO_MODE_OK (regno, word_mode));

This doesn't work reliably on MIPS, where a doubleword HI-LO value
cannot be split into independent HI and LO words; the LO word is OK,
but the HI word isn't.  This means that the test works correctly
on big-endian targets, where the invalid HI case comes first,
but not on little-endian ones, where the valid LO one does.

This endian difference is the cause of the mips-sde-elf build failure
that Maciej reported earlier in the week.  Tested on that target and
on x86_64-linux-gnu.  OK to install?

Richard

gcc/

2011-12-19  Richard Sandiford  <rdsandiford@googlemail.com>

	* lower-subreg.c (can_decompose_p): Check every word of a hard
	register.

--- gcc-4.6.2/gcc/lower-subreg.c.~1~	2011-01-03 21:52:22.000000000 +0100
+++ gcc-4.6.2/gcc/lower-subreg.c	2011-12-27 12:57:09.000000000 +0100
@@ -634,8 +634,15 @@ can_decompose_p (rtx x)
       unsigned int regno = REGNO (x);
 
       if (HARD_REGISTER_NUM_P (regno))
-	return (validate_subreg (word_mode, GET_MODE (x), x, UNITS_PER_WORD)
-		&& HARD_REGNO_MODE_OK (regno, word_mode));
+	{
+	  unsigned int byte, num_bytes;
+
+	  num_bytes = GET_MODE_SIZE (GET_MODE (x));
+	  for (byte = 0; byte < num_bytes; byte += UNITS_PER_WORD)
+	    if (simplify_subreg_regno (regno, GET_MODE (x), byte, word_mode) < 0)
+	      return false;
+	  return true;
+	}
       else
 	return !bitmap_bit_p (subreg_context, regno);
     }