My apologies for not responding to this bug earlier. I'm currently the upstream maintainer for the glibc port to hppa. I also try to do all the debian libc6 work. I'm suffering from the "must finish thesis" syndrome and I had to drop everything. A consequence is that hppa has some mis-merged patched and missing fixes.
We fixed this bug already in cvs.parisc-linux.org, you may want to look at our solution, since that is what I'll be submitting upstream in mid-November (that's my estimate for when I get free time back). We had to tackle all of these problems when we wrote the NPTL port of hppa. The TLS changes were for the 4.x gcc branch so we had to fixup glibc so that it worked with that compiler. A warning, the following tree is setup to also build the TLS/NPTL threading code, but the generic sysdeps and math code should be fine to copy. http://cvs.parisc-linux.org/glibc/sysdeps/hppa/fpu/ You also must be *very* careful with the fstd, fldd operations since they are used to "store" and "restore" the state of FPU. This is a tricky operation since hppa can and does make use of delayed floating point exceptions (e.g. It might look like it works but if you run the test again it might fail). The code I wrote for fesetround was: --- int fesetround (int round) { struct { unsigned int sw[2]; } s; /* Get the current status word. */ __asm__ ("fstd %%fr0,0(%1)" : "=m" (s) : "r" (&s) : "%r0"); s.sw[0] &= ~FE_DOWNWARD; s.sw[0] |= round & FE_DOWNWARD; __asm__ ("fldd 0(%0),%%fr0" : : "r" (&s), "m" (s) : "%r0"); return 0; } --- Comments: - Please do not return 1 if the rounding mode is invalid. Try as hard as possible to set something reasonable. - Use a struct where possible for two integers since GCC will generally have an easier time following the memory. - These are double word stores, so you need sw[2]. Again, sorry for not submitting upstream or helping out with the state of libc6 for hppa. I'm damned busy. I hope for this state to change with the completion of my thesis. Cheers, Carlos -- To UNSUBSCRIBE, email to [EMAIL PROTECTED] with a subject of "unsubscribe". Trouble? Contact [EMAIL PROTECTED]
