Differing results from cross and native compilers

["Eric DeVolder" <edevolder@xxxxxxxxxxxxxxxxxxxxxxxx>]

 

Hi, I have a MIPS native and cross compiler (gcc 3.4.4 and binutils 2.15)
both built using crosstool-0.42. I have an Alchemy system (MIPS32)
for which I am building code. When I compile a simple example, I get
different instructions from the cross and native compilers.

 

Specifically, the source file is:

 

 #include <stdio.h>
 int
 main (int argc, char **argv)
 {
     return -printf("Hello, world!\n");
 }

 

To compile natively, I use:

 

 % gcc -c hello.c -o obj1.o

 

To cross compile, I use:

 

 % mipsel-unknown-linux-gnu-gcc -c hello.c -o obj2.o

 

If I use the -O0 or -O1 flags with each, I get identical output
(for the .text), but when I compile in both environments with -O2;
the compilers produce different results.

 

The native compiler on the Alchemy board results in the following:

 

obj1.o:     file format elf32-tradlittlemips

Disassembly of section .text:

00000000 <main>:
   0:   3c1c0000        lui     gp,0x0
   4:   279c0000        addiu   gp,gp,0
   8:   0399e021        addu    gp,gp,t9
   c:   27bdffe0        addiu   sp,sp,-32
  10:   afbf0018        sw      ra,24(sp)
  14:   afbc0010        sw      gp,16(sp)
  18:   8f840000        lw      a0,0(gp)
  1c:   00000000        nop
  20:   24840000        addiu   a0,a0,0
  24:   8f990000        lw      t9,0(gp)
  28:   00000000        nop
  2c:   0320f809        jalr    t9
  30:   00000000        nop
  34:   8fbc0010        lw      gp,16(sp)
  38:   8fbf0018        lw      ra,24(sp)
  3c:   00021023        negu    v0,v0
  40:   03e00008        jr      ra
  44:   27bd0020        addiu   sp,sp,32

 

The cross compiler results in:

 

obj2.o:     file format elf32-tradlittlemips

Disassembly of section .text:

00000000 <main>:
   0:   3c1c0000        lui     gp,0x0
   4:   279c0000        addiu   gp,gp,0
   8:   0399e021        addu    gp,gp,t9
   c:   27bdffe0        addiu   sp,sp,-32
  10:   afbf0018        sw      ra,24(sp)
  14:   afbc0010        sw      gp,16(sp)
  18:   8f840000        lw      a0,0(gp)
  1c:   8f990000        lw      t9,0(gp)
  20:   00000000        nop
  24:   0320f809        jalr    t9
  28:   24840000        addiu   a0,a0,0
  2c:   8fbc0010        lw      gp,16(sp)
  30:   8fbf0018        lw      ra,24(sp)
  34:   00021023        negu    v0,v0
  38:   03e00008        jr      ra
  3c:   27bd0020        addiu   sp,sp,32

 

I would have thought that for this simple example I'd get identical
results. For reasons I have yet to figure out, the cross compiler is
finding better optimizations than the native, though both are the
same gcc version. I checked the specs files, and the two are identical
other than the "cross_compile" setting.

 

I'd like to solve this little mystery to get to the bigger prize, and
that is the ability to actually compile code (e.g. the entire system)
for MIPS32. As is demonstrated above, the compilers are generating
MIPS1 code (based on the NOPs between the load-to-use), but if I
could get the compiler to honor -mips32 (this discovery I originally

made while trying to get native -mips32 to work), I can get even

tighter code...

 

I've done quite a bit of googling to no avail. I've also done some
digging into gcc/binutils sources, and it appears that binutils has a
"from-abi" option for choosing its architecture, which is o32 which
probably means just MIPS1. I thought about trying to rebuild all the
tools for mipsisa32-unknown-linux-gnu, but am concerned I'd still
have the same problem of a particular compiler optimziation option
not being honored by the compiler.

 

Any tips/advice much appreciated, thanks!
Eric