Re: GNAT and register allocation

["Markus Schöpflin" <no.spam@spam.spam>]

Am 20.04.2012 13:17, schrieb Georg Bauhaus:
> On 20.04.12 12:21, Georg Bauhaus wrote:
>> Hi,
>>
>> in short, is there anything that one can do to make
>> GNAT use many registers? MMX registers in particular,
>> but not only.
>
> SSE, that is, darn, the registers being named xmmN.
>
>> It appears that the number of registers allocated increases
>> with inline expansion enabled. But only so much.
>
> In the (meaningless) example below, the translations of both
> Comp.A and Comp.B will use the same set of registers, xmm0 .. xmm1.
> If they were instead using xmm0 .. xmm1 and xmm2 .. xmm3,
> respectively, speed might well double.

[snip code]

As long as the functions are not inlined, I'd expect them to obey the calling 
convention of the ABI of the platform on which they are compiled. Therefore 
you won't see A using xmm0/1 and B using xmm2/3.

If they are inlined however, things look quite different. Here is what I get 
when I compile your code:

---%<---
 > gnatmake -g -O3 comp
gcc-4.4 -c -g -O3 comp.adb

 > objdump -d comp.o

comp.o:     file format elf64-x86-64


Disassembly of section .text:
...
0000000000000050 <comp___elabb>:
...
   5b:   e8 00 00 00 00          callq  60 <comp___elabb+0x10>
   60:   84 c0                   test   %al,%al
   62:   0f 84 8f 00 00 00       je     f7 <comp___elabb+0xa7>
   68:   66 0f 57 c9             xorpd  %xmm1,%xmm1
   6c:   31 c0                   xor    %eax,%eax
   6e:   f2 0f 10 05 00 00 00    movsd  0x0(%rip),%xmm0        # 76 
<comp___elabb+0x26>
   75:   00
   76:   f2 0f 10 15 00 00 00    movsd  0x0(%rip),%xmm2        # 7e 
<comp___elabb+0x2e>
   7d:   00
   7e:   f2 0f 5c c8             subsd  %xmm0,%xmm1
   82:   f2 0f 10 35 00 00 00    movsd  0x0(%rip),%xmm6        # 8a 
<comp___elabb+0x3a>
   89:   00
   8a:   f2 0f 10 2d 00 00 00    movsd  0x0(%rip),%xmm5        # 92 
<comp___elabb+0x42>
   91:   00
   92:   f2 0f 10 25 00 00 00    movsd  0x0(%rip),%xmm4        # 9a 
<comp___elabb+0x4a>
   99:   00
   9a:   66 0f 1f 44 00 00       nopw   0x0(%rax,%rax,1)
   a0:   66 0f 28 d9             movapd %xmm1,%xmm3
   a4:   f2 0f 59 cd             mulsd  %xmm5,%xmm1
   a8:   f2 0f 5c c2             subsd  %xmm2,%xmm0
   ac:   83 c0 01                add    $0x1,%eax
   af:   f2 0f 5e de             divsd  %xmm6,%xmm3
   b3:   3d 40 42 0f 00          cmp    $0xf4240,%eax
   b8:   f2 0f 58 c9             addsd  %xmm1,%xmm1
   bc:   f2 0f 58 c0             addsd  %xmm0,%xmm0
   c0:   f2 0f 58 c3             addsd  %xmm3,%xmm0
   c4:   f2 0f 59 c0             mulsd  %xmm0,%xmm0
   c8:   66 0f 28 d8             movapd %xmm0,%xmm3
   cc:   f2 0f 58 da             addsd  %xmm2,%xmm3
   d0:   f2 0f 58 cb             addsd  %xmm3,%xmm1
   d4:   f2 0f 59 c9             mulsd  %xmm1,%xmm1
   d8:   f2 0f 5e cc             divsd  %xmm4,%xmm1
   dc:   75 c2                   jne    a0 <comp___elabb+0x50>
   de:   f2 0f 58 c1             addsd  %xmm1,%xmm0
   e2:   f2 0f 59 05 00 00 00    mulsd  0x0(%rip),%xmm0        # ea 
<comp___elabb+0x9a>
   e9:   00
   ea:   f2 0f 11 05 00 00 00    movsd  %xmm0,0x0(%rip)        # f2 
<comp___elabb+0xa2>
--->%---

As you can see, the xmmN registers up to xmm6 are used. Do you get different 
results?

Markus