On Mon, 10 Nov 2014, NightStrike wrote:
On Mon, Oct 20, 2014 at 7:18 PM, NightStrike <nightstrike@xxxxxxxxx> wrote:I have been studying the asm generated by a typical clamping function, and I am confused about the results. This is done on an Opteron 6k series compiled with -fverbose-asm, -O3 and -march=native. float clamp(float const x, float const min, float const max) { #if defined (BRANCH) if ( x > max ) return max; else if ( x < min ) return min; else return x; #elif defined (BRANCH2) return x > max ? max : ( x < min ? min : x ); #elif defined (CALL) return __builtin_fminf(__builtin_fmaxf(x, min), max); #else float const t = x < min ? min : x; return t> max ? max : t; #endif } -DBRANCH / -DBRANCH2: The first two approaches are obviously identical, and produce: clamp: .LFB0: .cfi_startproc vucomiss %xmm2, %xmm0 # max, x ja .L3 #, vmaxss %xmm0, %xmm1, %xmm0 # x, min, D.2214 ret .p2align 4,,7 .p2align 3 .L3: vmovaps %xmm2, %xmm0 # max, D.2214 ret .cfi_endproc -DCALL: This one I figured would be great, given the use of builtins: clamp: .LFB0: .cfi_startproc subq $24, %rsp #, .cfi_def_cfa_offset 32 vmovss %xmm2, 12(%rsp) # max, %sfp call fmaxf # vmovss 12(%rsp), %xmm2 # %sfp, max addq $24, %rsp #, .cfi_def_cfa_offset 8 vmovaps %xmm2, %xmm1 # max, jmp fminf # .cfi_endproc
I guess -ffast-math (or some weaker option) would let it generate the same as the next version.
But then we have what appears to be the best of them all.... just a couple instructions, no branches, no calls, nothing: .LFB0: .cfi_startproc vmaxss %xmm0, %xmm1, %xmm0 # x, min, D.2219 vminss %xmm0, %xmm2, %xmm0 # D.2219, max, D.2219 ret .cfi_endproc So I'm curious.... why is the last approach optimized better than the naive approach of some nested if statements?
Optimization is done locally, it detects a "max" pattern and a "min" pattern. The BRANCH versions are a more complicated pattern. There is code in phiopt that is supposed to handle it, but apparently it requires that we can prove at compile-time that min<=max. If you believe it can be generalized, you could file an enhancement PR with details about what to transform to what, and why this is valid whatever the ordering of x, min and max.
-- Marc Glisse
