Richard Guenther <richard.guent...@gmail.com> writes: > On Thu, Feb 9, 2012 at 6:32 PM, Ian Lance Taylor <i...@google.com> wrote: >> Richard Guenther <richard.guent...@gmail.com> writes: >> >>> On Wed, Feb 8, 2012 at 8:38 PM, Ian Lance Taylor <i...@google.com> wrote: >>>> The master Go math library uses assembler code on 386 processors to take >>>> advantage of 387 instructions. This patch lets gccgo do the same thing, >>>> by compiling the math library with -funsafe-math-optimizations. I also >>>> pass -mfancy-math-387, although that is the default. It would not be >>>> appropriate to compile all Go code with -funsafe-math-optimizations, of >>>> course, but the math library is designed to handle it. >>> >>> Huh ... I'd rather not do that if I were you. Instead I'd say we lack a >>> machine specific flag to enable the fancy-x87-math patterns which >>> then -funsafe-math-optimizations should enable. The x87 math >>> routines are the only thing you are after, right? No math-library >>> can be _safe_ against -funsafe-math-optimizations I believe. >> >> Yes, that approach would make sense, but this doesn't seem like the >> right time to do it. >> >> The -funsafe-math-optimizations option does not permit arbitrary >> behaviour. It merely permits a set of optimizations which violate >> strict IEEE conformance. I believe the Go math library can be safe in >> the presence of those optimizations, because the library does explicit >> checks for NaN and infinity, where necessary, before it does the actual >> operation. The math library has a fairly extensive set of tests, >> including tests of exceptional conditions, and it passes the tests when >> using -funsafe-math-optimizations. Note that I'm only using >> -funsafe-math-optimizations on x86. > > I see. -funsafe-math-optimizations affects precision though (it doesn't > assume NaNs or Infinities do not happen), see the docs - it enables > -fno-signed-zeros, -fno-trapping-math, -fassociative-math and > -freciprocal-math.
Affecting precision is OK for the Go math library, since the master version of the library is written to use the 80-bit 80387 registers on 386, in order to use the magic 80387 instructions. Ian