[backport patch approved for 4.9 stage1, I believe native compilation
 on ARM is affected by the missed-optimization <mikpe@it.uu.se>]

Date: Thu, 28 Feb 2013 00:30:24 +0800
Subject: [PATCH, combine] Fix host-specific behavior in simplify_compare_const()
From: Chung-Ju Wu <jasonwucj at gmail dot com>
List-Archive: <http://gcc.gnu.org/ml/gcc-patches/>

Hi all,

I notice the function simplify_compare_const() in gcc/combine.c presents
different behavior between 32-bit host machine and 64-bit host machine,
causing it to perform different RTL transformation for a 32-bit target.

In the function: simplify_compare_const (enum rtx_code code, rtx op0,
rtx *pop1),
there is a statement "HOST_WIDE_INT const_op = INTVAL (*pop1);".
Although it uses "trunc_int_for_mode()" to make const_op
present the value for target machine mode,
the subsequent condition checking does not filter out the unnecessary bits:

  if (...
      ...
      && exact_log2 (const_op) >= 0
      && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) const_op)
    {
      ...
    }

For example, given that:

op0: (and:SI (reg/v:SI 54)
             (const_int -2147483648))
*pop1: (const_int -2147483648)


Since mode comes from GET_MODE (op0), we know that mode is SImode.


on 32-bit host machine,

  rtx *pop1: (const_int -2147483648 [0x80000000])
  const_op: 0x80000000

  exact_log2 (const_op): 31
  nonzero_bits (op0, mode)         : 0x80000000
  (unsigned HOST_WIDE_INT) const_op: 0x80000000


on 64-bit host machine,

  rtx *pop1: (const_int -2147483648 [0xffffffff80000000])
  const_op: 0xffffffff80000000

  exact_log2 (const_op): -1
  nonzero_bits (op0, mode)         : 0x0000000080000000
  (unsigned HOST_WIDE_INT) const_op: 0xffffffff80000000


In this case, the condition results on 32-bit and 64-bit host machine
are different.
The result on 64-bit host machine cannot pass the condition test
so that we have no chance to do optimization.

In other words, any 32-bit target with 'need_64bit_hwint=yes' in config.gcc
is not able to have benefit from this optimization because it never
passes the condition test.


My solution is to use GET_MODE_MASK(mode) to filter out all bits not
in target mode.
The following is my patch:


gcc/ChangLog:

        * gcc/combine.c: Use GET_MODE_MASK() to filter out
        unnecessary bits in simplify_compare_const().

--- gcc-4.7.2/gcc/combine.c.~1~	2012-08-09 17:33:28.000000000 +0200
+++ gcc-4.7.2/gcc/combine.c	2013-03-09 13:32:13.830901513 +0100
@@ -10917,8 +10917,8 @@ simplify_compare_const (enum rtx_code co
       && (code == EQ || code == NE || code == GE || code == GEU
 	  || code == LT || code == LTU)
       && mode_width <= HOST_BITS_PER_WIDE_INT
-      && exact_log2 (const_op) >= 0
-      && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) const_op)
+      && exact_log2 (const_op & GET_MODE_MASK (mode)) >= 0
+      && nonzero_bits (op0, mode) == (unsigned HOST_WIDE_INT) (const_op & GET_MODE_MASK (mode)))
     {
       code = (code == EQ || code == GE || code == GEU ? NE : EQ);
       const_op = 0;