On Thu, Feb 7, 2013 at 12:01 AM, John Fine <johnsfine@xxxxxxxxxxx> wrote: > I haven't tried constructing a small test case. I fixed the problem by > changing to > if ( expression of *p ) *((char volatile*)p) = '-'; > > I am a skilled enough asm programmer to have zero doubt that I correctly > interpreted the compiled code. It did have that "optimization". > > I don't know what else in that C++ function might have given the compiler > permission to make that optimization. I don't think anything did. I know > there was no unconditional write to *p in C++ code and there was an > unconditional write to *p in the asm code. Despite Jay's skepticism, GCC used to produce speculative stores, so certainly at least one compiler acted that way. It's a natural optimization in the single-threaded world. Stores are much cheaper than conditional branches. Ian > Jay wrote: >> >> This sounds like a bug. But I am skeptical that any compiler has this bug. >> Do you have a small test case that demonstrates it? Compilers aren't >> bug-free (nothing is) but there are FAR more bugs in compiler input than >> compilers themselves. >> - Jay >> >> On Feb 3, 2013, at 9:34 PM, "Mailaripillai, Kannan Jeganathan" >> <kannanmj@xxxxxx> wrote: >> >> >>>> >>>> if ( expression of *p ) *p = '-'; >>>> >>>> The optimizer changes that to the equivalent of >>>> >>>> *p = (expression of *p ) ? '-' : *p; >>>> >>>> Is that a valid optimization? >>>> >> >> >>
