[Bug target/39942] Nonoptimal code - leaveq; xchg %ax,%ax; retq

["jakub at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #31 from jakub at gcc dot gnu dot org  2009-05-14 15:15 -------
Some -O2 code size data from today's trunk bootstraps.  The first .text line
is always vanilla bootstrap, the second one with
http://gcc.gnu.org/ml/gcc-patches/2009-05/msg00702.html
only, the third one with that patch and
http://gcc.gnu.org/ml/gcc-patches/2009-05/msg00703.html
and the last with additional:
--- i386.c.jj3  2009-05-14 12:41:24.000000000 +0200
+++ i386.c      2009-05-14 14:48:24.000000000 +0200
@@ -27202,7 +27202,7 @@ x86_function_profiler (FILE *file, int l
 static int
 min_insn_size (rtx insn)
 {
-  int l = 0;
+  int l = 0, len;

   if (!INSN_P (insn) || !active_insn_p (insn))
     return 0;
@@ -27222,7 +27222,8 @@ min_insn_size (rtx insn)
       && symbolic_reference_mentioned_p (PATTERN (insn))
       && !SIBLING_CALL_P (insn))
     return 5;
-  if (get_attr_length (insn) <= 1)
+  len = get_attr_length (insn);
+  if (len <= 1)
     return 1;

   /* For normal instructions we may rely on the sizes of addresses
@@ -27230,6 +27231,9 @@ min_insn_size (rtx insn)
      This is not the case for jumps where references are PC relative.  */
   if (!JUMP_P (insn))
     {
+      if (get_attr_type (insn) != TYPE_MULTI)
+       return len;
+
       l = get_attr_length_address (insn);
       if (l < 4 && symbolic_reference_mentioned_p (PATTERN (insn)))
        l = 4;
to see how the code size changes with much more accurate (though sometimes not
minimum but maximum bound) insn sizing for the algorithm.
64-bit cc1plus
  [12] .text             PROGBITS        000000000047f990 07f990 8c3ba8 00  AX 
0   0 16
  [12] .text             PROGBITS        000000000047f990 07f990 89b1e8 00  AX 
0   0 16
  [12] .text             PROGBITS        000000000047f9c0 07f9c0 899f78 00  AX 
0   0 16
  [12] .text             PROGBITS        000000000047f9c0 07f9c0 88eaf8 00  AX 
0   0 16
32-bit cc1plus
  [12] .text             PROGBITS        080b24e0 06a4e0 8f8cac 00  AX  0   0
16
  [12] .text             PROGBITS        080b24e0 06a4e0 8d516c 00  AX  0   0
16
  [12] .text             PROGBITS        080b2510 06a510 8d507c 00  AX  0   0
16
  [12] .text             PROGBITS        080b2510 06a510 8cbd7c 00  AX  0   0
16
For 64-bit cc1plus that's 1.8%, 1.86%, 2.36% smaller binary with the 1, 2 resp.
3 patches, for 32-bit cc1plus 1.55%, 1.56%, 1.96% smaller binary.
So the first patch is the most important and something like the third one,
perhaps with more exceptions, also makes a difference.  I'll now try to update
my awk script to check for the AMD rules, namely that the last byte of the
branch insn counts rather than the first.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=39942