"Ben Voigt" <rbv (AT) nospam (DOT) nospam> wrote

True for optimized builds, the call is hoisted, but the loop is not hoisted by the JIT, the
C++ compiler (optimized build) effectively hoists the loop.
What's produced by the JIT depends on the version of the CLR.

this snip of the code:
for (int i = 0; i < 50000; i++)
{
Run(i);
}

is turned into into:

xor r11d,r11d
add r11d,4
cmp r11d,0C350h
jl 00000642`80150341 (jump to add r11d, 4 if less than)


by the JIT64, while the JIT32 (both v2 of the CLR), produces

xor eax,eax
add eax,1
cmp eax,0C350h
jl 001e014b (jump to add eax, 1 if less than)

see the subtle difference:
add r11d, 4
and
add eax, 1

here the JIT64 is cheating , no big deal in this case, but I would prefer some more
consistent behavior across JIT versions, here I mean hoist the loop, or keep the loop as is,
but don't cheat.

Willy.

"Ben Voigt" <rbv (AT) nospam (DOT) nospam> wrote

True, with the following results:

C# (JIT32)
244
0,0000681650880209635582175947
C# (JIT64)
86
0,0000240253998762412541258735

C++ (O2 switch) 32 bit and 64 bit
164
0.000045815878834

see, C++ is faster than JIT 32 but slower than JIT64 code, for this particular case (50000
iterations), however such kind of micro-benchmarks have absolutely no value. For instance
make the loop count an odd number (eg. 49999) and you get the same results for C# JIT64
and C++.
Note that the JIT32 is known as not to being a great loop optimizer ;-).

Willy.