Comment #4 on issue 2294 by [email protected]: Apparently incorrect
rounding in Uint8ClampedArray implementation in V8
http://code.google.com/p/v8/issues/detail?id=2294
V8 does rounding as floor(x + 0.5) both in run-time code and generated code:
http://code.google.com/codesearch#OAMlx_jo-ck/src/v8/src/objects.cc&l=11597
http://code.google.com/codesearch#OAMlx_jo-ck/src/v8/src/ia32/macro-assembler-ia32.cc&l=128
We could fix the run-time code the same way as JSC did:
inline long int lrint(double flt)
{
int intgr;
#if CPU(X86)
__asm {
fld flt
fistp intgr
};
#else
#pragma message("Falling back to casting for lrint(), causes rounding
inaccuracy in halfway case.")
intgr = static_cast<int>flt;
#endif
return intgr;
}
But the generated code uses fast SSE instructions and replacing them by FPU
instructions there would degrade performance. I'll look for an SSE
instruction that does the correct rounding.
If I don't find it, I am not sure if it is optimal to fix this issue: we
would sacrifice performance in all cases to fix one corner case.
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
