[M3devel] cg.compare_exchange should take more parameters in function call, fewer on stack?

[Jay K]

It is easier to generate more efficient code if the variable is passed as a parameter to cg.compare_exchange instead of on the stack.

We can defer register allocation until we know we must use eax, instead of allocating it "wherever" and then moving it to eax once we know that is where it is needed, as it always will be.

We can also generate more efficient code for fetch_op if passed function parameters since we don't need to get the address in a register.

 

Again, if m3back were multi pass or somehow smarter, it'd be moot.

I'll be looking into make it one or both of these maybe sometime soon.

 

 - Jay
 


From: hosking at cs.purdue.edu
Date: Fri, 12 Feb 2010 09:49:14 -0500
To: jay.krell at cornell.edu
CC: m3devel at elegosoft.com
Subject: Re: [M3devel] cg.compare_exchange should take more parameters in function call, fewer on stack?

Here's what Boehm produces in his libatomic package.  Don't know if it helps or answers your question.


The C source:




int Atomic__CompareSwap_ordered (volatile AO_t *addr, AO_t *old, AO_t new)
{ return AO_compare_and_swap_full(addr, *old, new); }


generates:



.globl _Atomic__CompareSwap_ordered
_Atomic__CompareSwap_ordered:
LFB526:
pushq %rbp
LCFI42:
movq %rsp, %rbp
LCFI43:
movq (%rsi), %rax
lock; cmpxchgq %rdx, (%rdi); setz %al
movsbl %al,%eax
leave
ret
LFE526:
.align 4,0x90






On 12 Feb 2010, at 04:17, Jay K wrote:

Tony, currently I have atomic.compare_and_exchange generating e.g.:
 
 
 load_address      gv.2[_MM_Main] 56
 0000018E: 8D3538000000        LEA ESI gv.2[_MM_Main]+56
 load_address      gv.2[_MM_Main] 152
 00000194: 8D3D98000000        LEA EDI gv.2[_MM_Main]+152
 load              gv.2[_MM_Main] 156 Int.32 Int.32
 compare_exchange   Int.32 Int.32 Int.32
 0000019A: 8BC7                MOV EAX EDI
 0000019C: 8B1D9C000000        MOV EBX gv.2[_MM_Main]+156
 declare_temp      4 4 Addr F tv.35[T$35] -4
 000001A2: 8975FC              MOV tv.35[T$35] ESI
 000001A5: F00FB15DFC          LOCK CMPXCHG  tv.35[T$35]:Addr EBX
 

This might not be correct, but it is close.
 

Now, the thing is, the eax register is special to lock cmpxchg.
  It always holds the comparison value.

The point is, if cg.compare_exchange took its parameters as
parameters instead of on the stack, it could chose the
right register right away, instead of as above where it
choses the wrong one and then must mov it.
 
Easy to change it like that?
 

I realize this is an m3back deficiency.
It does the register allocation too early.

 - Jay



 		 	   		  
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://m3lists.elegosoft.com/pipermail/m3devel/attachments/20100212/1ebee2b2/attachment-0002.html>