Comment #1 on issue 1975 by [email protected]: Precision of trignometric
functions is low
http://code.google.com/p/v8/issues/detail?id=1975
Observations for test262/S15.8.2.18_A7:
1) V8 built for x64 using GCC does not exhibit the problem with Math.tan.
2) V8 built on Windows does not does not exhibit the problem either.
3) V8 built for ia32 using GCC exhibits the problem.
4) 32-bit Firefox on Linux exhibits the problem as well.
5) V8 built for ia32 using GCC with -ffloat-store does not have the problem.
Conclusion: when compiled with GCC, calculating tan on ia32 utilizes the
full 80 bit of FPU registers during calculations, making the result more
accurate, which however deviates from the expected value in Test262. This
happens only on GCC/ia32. With other compilers/architectures the results
are correct.
--
v8-dev mailing list
[email protected]
http://groups.google.com/group/v8-dev
