[prev in list] [next in list] [prev in thread] [next in thread] 

List:       gcc-bugs
Subject:    [Bug middle-end/30996]  New: Performance problem with structure assignment on MIPS
From:       "ska at bth dot se" <gcc-bugzilla () gcc ! gnu ! org>
Date:       2007-02-28 16:09:41
Message-ID: bug-30996-12925 () http ! gcc ! gnu ! org/bugzilla/
[Download RAW message or body]

GCC 4.1 seems to sometimes generate inefficient code when doing
structure assignments directly when compiling for MIPS1. When assigning
to structure members manually, it generates regular lw/sw sequences.
When assigning to the structure, you instead get lwl/lwr and swl/swr
pairs (for no reason, since the data is aligned).

I'm translating the MIPS code to Java bytecode, so the problem is even
worse for me as I then have to read four separate bytes and or
these together to implement a lwl/lwr pair.


I've attached a preprocessed file which exhibits this behavior. The two
functions below show this, the first doing a structure assignment and
the second member assignment.

  void dummy1(NavigateSegment *segments,
             int count,
             RoadMapPosition *src_pos,
             RoadMapPosition *dst_pos) {

      int i;
      int group_id = 0;
      NavigateSegment *segment;

      for (i=0; i < count; i++) {
         set_from_pos (&segments[i].from_pos);
         segments[i].shape_initial_pos = segments[i].from_pos;

      }
   }

   void dummy2(NavigateSegment *segments,
               int count,
               RoadMapPosition *src_pos,
               RoadMapPosition *dst_pos) {

      int i;
      int group_id = 0;
      NavigateSegment *segment;

      for (i=0; i < count; i++) {
         set_from_pos (&segments[i].from_pos);
         segments[i].shape_initial_pos.longitude =
segments[i].from_pos.longitude;
         segments[i].shape_initial_pos.latitude =
segments[i].from_pos.latitude;
      }
   }

And if you disassemble it you see that they look like

00000014 <dummy1>:
  14:   00003021        move    a2,zero
  18:   240804d2        li      t0,1234
  1c:   08000014        j       50 <dummy1+0x3c>
  20:   2407162e        li      a3,5678
  24:   ac88001c        sw      t0,28(a0)
  28:   ac870020        sw      a3,32(a0)
  2c:   8882001c        lwl     v0,28(a0)
  30:   88830020        lwl     v1,32(a0)
  ...

00000064 <dummy2>:
  64:   00001821        move    v1,zero
  68:   240604d2        li      a2,1234
  6c:   08000022        j       88 <dummy2+0x24>
  70:   2407162e        li      a3,5678
  74:   ac870020        sw      a3,32(a0)
  78:   ac86001c        sw      a2,28(a0)
  7c:   ac86002c        sw      a2,44(a0)
  80:   ac870030        sw      a3,48(a0)
  ....

If the loop is removed, both functions generate the same code (with
regular loads/stores). The GCC version is

   mips-linux-gnu-gcc (GCC) 4.1.2 20061115 (prerelease) (Debian 4.1.1-21)

(Ehud Shabtai discovered this problem)


-- 
           Summary: Performance problem with structure assignment on MIPS
           Product: gcc
           Version: 4.1.2
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: ska at bth dot se
 GCC build triplet: mips-linux-gnu
GCC target triplet: mips-linux-gnu


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=30996

[prev in list] [next in list] [prev in thread] [next in thread]