Re: [PATCH, x86_64]: Provide longlong.h definitions for 128bit operations

[Chris Lattner <clattner at apple dot com>]

> Unfortunatelly, this approach confuses gcc a lot. Attached testcase
> will illustrate the problem on i686 (substitute SI with DI and DI with
> TI on 64bit and change shift values for the same result):
> Compared two approaches head-to-head on i686 target; gcc -O2
> -fomit-frame-pointer generates:
>
> test_c:
>       subl    $28, %esp       #,
>       xorl    %edx, %edx      #
>       movl    40(%esp), %eax  # c, tmp66
>       movl    44(%esp), %ecx  # d, d
>       movl    %esi, 20(%esp)  #,
>       movl    %ebx, 16(%esp)  #,
>       movl    %eax, %edx      # tmp66,
>       movl    $0, %eax        #, tmp66
>       movl    %eax, %esi      #, tmp74
>       movl    32(%esp), %eax  # a, tmp70
>       movl    %edx, %ebx      # tmp75, __c
>       orl     %ecx, %esi      # d, tmp74
>       xorl    %edx, %edx      #
>       movl    %esi, %ecx      # tmp74, __c
>       movl    36(%esp), %esi  # b, b
>       movl    %edi, 24(%esp)  #,
>       movl    24(%esp), %edi  #,
>       movl    %eax, %edx      # tmp70,
>       movl    $0, %eax        #, tmp70
>       orl     %esi, %eax      # b, tmp72
>       movl    20(%esp), %esi  #,
>       addl    %eax, %ecx      # tmp72, __c
>       adcl    %edx, %ebx      #, __c
>       movl    %ecx, 8(%esp)   # __c, y
>       movl    %ebx, %ecx      # __c, __c
>       xorl    %ebx, %ebx      # __c
>       movl    16(%esp), %ebx  #,
>       movl    %ecx, 12(%esp)  # __c, x
>       addl    $28, %esp       #,
>       ret
>
> Looking into this mess, I think that gcc has still a long way to go
> w.r.t to DImode values on 32bit targets (and TImode values on 64bit
> targets) and the only solution for the near future is using asm
> here...

At some point, hacking around these limitations becomes less useful  
than just fixing them once and for all.  LLVM compiles this:

volatile SI x, y;
void test_asm (SI a, SI b, SI c, SI d){
add_ssaaaa_asm (x, y, a, b, c, d);
}


into this:

_test_c:
	movl 12(%esp), %eax
	movl 16(%esp), %ecx
	addl 8(%esp), %ecx
	adcl 4(%esp), %eax
	movl %ecx, _y
	movl %eax, _x
	ret

with llvm-gcc t.c -S -O -static -fomit-frame-pointer

Inline asm really is bad.  It's much better to fix the compiler so  
that users don't have to rely on it to get performance.

-Chris