On Dec 4, 2010, at 5:22 AM, Florian Weimer wrote:
> * Joe Buck:
>
>> It's wasted code if the multiply instruction detects the overflow.
>> It's true that the cost is small (maybe just one extra instruction
>> and the same number of tests, maybe one more on architectures where you
>> have to load a large constant), but it is slightly worse code than what
>> Chris Lattner showed.
>
> It's possible to improve slightly on the LLVM code by using the
> overflow flag (at least on i386/amd64), as explained in this blog
> post:
>
> <http://blogs.msdn.com/b/michael_howard/archive/2005/12/06/500629.aspx>
Ah, great point. I improved the clang codegen to this:
$ cat t.cc
void *test(long count) {
return new int[count];
}
$ clang t.cc -S -o - -O3 -mkernel -fomit-frame-pointer -mllvm -show-mc-encoding
.section __TEXT,__text,regular,pure_instructions
.globl __Z4testl
.align 4, 0x90
__Z4testl: ## @_Z4testl
## BB#0: ## %entry
movl $4, %ecx ## encoding: [0xb9,0x04,0x00,0x00,0x00]
movq %rdi, %rax ## encoding: [0x48,0x89,0xf8]
mulq %rcx ## encoding: [0x48,0xf7,0xe1]
movq $-1, %rdi ## encoding:
[0x48,0xc7,0xc7,0xff,0xff,0xff,0xff]
cmovnoq %rax, %rdi ## encoding: [0x48,0x0f,0x41,0xf8]
jmp __Znam ## TAILCALL
## encoding: [0xeb,A]
## fixup A - offset: 1, value:
__Znam-1, kind: FK_PCRel_1
.subsections_via_symbols
This could be further improved by inverting the cmov condition to avoid the
first movq, which we'll tackle as a general regalloc improvement.
Thanks for the pointer!
-Chris
