Missed optimization: unnecessary branching

[jrmuizel]

Bugzilla Link	9661
Resolution	LATER
Resolved on	Apr 21, 2011 20:37
Version	trunk
OS	All
CC	@tkremenek

Extended Description

This code:
float clamp_float(float a) {
if (a > 1.0f)
return 1.0f;
else if (a < 0.0f)
return 0.0f;
else
return a;
}

compiles to:

clamp_float: # @​clamp_float
movss .LCPI0_0(%rip), %xmm1
ucomiss %xmm1, %xmm0
ja .LBB0_2
pxor %xmm1, %xmm1
maxss %xmm0, %xmm1
.LBB0_2: # %return
movaps %xmm1, %xmm0
ret

It should be able to be something like:
clamp_float: # @​clamp_float
movss .LCPI0_0(%rip), %xmm1
minss %xmm1, %xmm0
pxor %xmm1, %xmm1
maxss %xmm1, %xmm0
ret

[tkremenek]

cloned to rdar://problem/9258068

[llvmbot]

What you described is only safe with -funsafe-math-optimizations or -ffast-math.

[jrmuizel]

What you described is only safe with -funsafe-math-optimizations or
-ffast-math.

Indeed.

This version, however, seems to work.

#include <stdio.h>

float clamp(float a) {
if (a > 1.f) {
return 1.f;
} else if (a < 0.f) {
return 0.f;
} else {
return a;
}
}

float clamp2(float a) {
float a2 = 0.f, a3 = 1.f;
asm(
"maxss %2, %0\n"
"minss %0, %1\n" : "+x" (a2), "+x" (a3) : "x" (a));
return a3;
}

typedef union {
unsigned int i;
float f;
} fp;

int main()
{
unsigned int x = 0;
do {
fp in, res1, res2;
in.i = x;
res1.f = clamp(in.f);
res2.f = clamp2(in.f);
if (res1.i != res2.i) {
printf("difference %f:%x -> %f:%x %f:%x\n",
in.f, in.i,
res1.f, res1.i,
res2.f, res2.i);
}
x++;
} while (x != 0);
}

[lattner]

I moved this x86-specific micro-optimization to the X86 readme file in r129527.

[jrmuizel]

This affects more than just x86. On ARM for example:

float clamp(float a)
{
if (a > 1) {
return 1;
} else if (a < 0) {
return 0;
} else {
return a;
}
}

float clamp2(float a)
{
if (a > 1) {
a = 1;
}
if (a < 0) {
return 0;
} else {
return a;
}
}

generates:

clamp: @ @​clamp
vmov.f32 s0, #​1.000000e+00
vmov s1, r0
vcmpe.f32 s1, s0
vmrs apsr_nzcv, fpscr
bgt .LBB0_2
vcmpe.f32 s1, #​0
vldr.32 s0, .LCPI0_0
vmrs apsr_nzcv, fpscr
vmovmi.f32 s1, s0
vmov.f32 s0, s1
.LBB0_2: @ %return
vmov r0, s0
bx lr

clamp2: @ @​clamp2
vmov.f32 s0, #​1.000000e+00
vmov s2, r0
vcmpe.f32 s2, s0
vldr.32 s1, .LCPI1_0
vmrs apsr_nzcv, fpscr
vmovgt.f32 s2, s0
vcmpe.f32 s2, #​0
vmrs apsr_nzcv, fpscr
vmovmi.f32 s2, s1
vmov r0, s2
bx lr