| Summary: | setjmp/longjmp do not save/restore correct FPU registers in N32 ABI mode | ||
|---|---|---|---|
| Product: | glibc | Reporter: | Nicolas S. Dade <n-redhat-glibc> |
| Component: | ports | Assignee: | Roland McGrath <roland> |
| Status: | RESOLVED INVALID | ||
| Severity: | normal | CC: | glibc-bugs |
| Priority: | P2 | Flags: | fweimer:
security-
|
| Version: | 2.8 | ||
| Target Milestone: | --- | ||
| Host: | mips64-unknown-linux-gnu | Target: | |
| Build: | i686-unknown-linux-gnu | Last reconfirmed: | |
My copy of the MIPSpro(TM) N32 ABI Handbook says "There are eight callee-saved floating point registers, $f24..$f31 for the 64-bit interface. There are six for the n32 ABI, the six even registers in $f20..$f30.". This agrees with glibc and GCC. |
In the files glibc/ports/sysdeps/mips/mips64/setjmp_aux.c glibc/ports/sysdeps/mips/mips64/__longjmp.c the innermost #if checks look like this: #if _MIPS_SIM == _ABI64 [ save/restore f24, f25, ... f31 ] #else [ save/restore f20,, f22, ... f30 ] (Even registers only b/c on 32-bit CPUs the registers are paired even/odd to hold double precision FP numbers) #endif Thus in _ABIN32 (64-bit CPU using 32-bit pointers) the 32-bit CPU code is selected, which fails to save or restore any odd numbered registers. The N32 ABI defines f24-f31 as callee-saved, same as _ABI64 does. One fix would be to change the tests from - #if _MIPS_SIM == _ABI64 + #if _MIPS_SIM != _ABIO32 This is present in at least glibc-ports-2.7 (where we discovered it), 2.8 and in CVS HEAD. -Nicolas Dade