On 12/21/2016 07:29 AM, David Wohlferd wrote: > Ahh, the wonders of google. I believe I have found the problem kai's > change was originally intended to fix: > http://mingw-w64-public.narkive.com/jr5jMaJ2/quadmath-h-s-expq-crashes-at-runtime > > > And, I believe I have found what the actual problem was. I don't think > what kai did was the right fix. Let me explain why. > > There is a big old structure full of obscure x87 floating point settings > defined as fenv_t in fenv.h (make sure you don't look at the ARM > definition by mistake). The format of this structure is defined by the > fldenv assembler instruction (so blame Intel). > > Now, someone decided to be tricky and used some of the 'unused' data > members of this struct to store, not just the x87 FPU settings, but also > the SSE settings. It does this by overwriting the __unused0 and > __unused1 members of the struct. > > So, when you call fegetenv(fenv_t * envp), what you actually get back is > a mixture of settings for both x87 and SSE. This isn't (normally) a > problem, since when you call fesetenv (const fenv_t * envp), it extracts > the SSE values back out of the struct, puts back the normal 'default' > values for __unused0 and __unused1, then calls fldenv to set the x87 > values, followed by ldmxcsr for the SSE values. > > However, quadmath doesn't use fegetenv to retrieve the current > settings. It uses feholdexcept. Unfortunately, feholdexcept doesn't > set __unused0 and __unused1 using the SSE flags the way fegetenv does. > As a result, when quadmath eventually calls fesetenv (&oldenv) to put > things back, the values in __unused0 and __unused1 are not valid SSE > flag settings. This is why quadmath was crashing. > > Instead of fixing feholdexcept to use the same format that the rest of > the fenv_t functions use, kai changed fesetenv as described below. > > I think the correct fix is something like the attached. This undoes > kai's change, and fixes feholdexcept instead. I've taken a brief look, > and I don't see any other functions that are like this. > > Comments? Zidane, can you try this? >
David, can you also test against the gfortran testsuites? Kai asks to make sure there are no new failures introduced by this change. Thanks for looking into it.
signature.asc
Description: OpenPGP digital signature
------------------------------------------------------------------------------ Developer Access Program for Intel Xeon Phi Processors Access to Intel Xeon Phi processor-based developer platforms. With one year of Intel Parallel Studio XE. Training and support from Colfax. Order your platform today.http://sdm.link/intel
_______________________________________________ Mingw-w64-public mailing list Mingw-w64-public@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/mingw-w64-public
