On Tue, Jun 20, 2023 at 3:37 PM Ian Lance Taylor <i...@golang.org> wrote:
> On Tue, Jun 20, 2023 at 11:35 AM Andreas Schwab <sch...@linux-m68k.org> > wrote: > > > > On Jun 20 2023, Ian Lance Taylor via Gcc-patches wrote: > > > > > This libgo patches changes the runtime pacakge to use a C function to > call mmap. > > > > > > The final argument to mmap, of type off_t, varies. In > > > https://go.dev/cl/445375 > > > (https://gcc.gnu.org/pipermail/gcc-patches/2022-October/604158.html) > > > we changed it to always use the C off_t type, but that broke 32-bit > > > big-endian Linux systems. > > > > This has nothing to do with big-endian, armv7 isn't big-endian. > > OK, but I think that it does have something to do with big-endian. > The bug was that on some 32-bit systems it was passing a 64-bit value > to a function that expected a 32-bit value. The problem didn't show > up on 32-bit x86 because it is little-endian, and did show up on > 32-bit PPC because it is big-endian. I guess the armv7 case was > failing for a different reason. I think there is a calling convention issue. On 32-bit ARM, for the case of mmap, if the last argument is 32-bit, it is passed 4 bytes at sp+4. If it is 64-bit, the offset is aligned and it is stored as 8 bytes at sp+8. So if the callee tries to read at sp+4, it gets the wrong value, even for little endian. On 32-bit x86 it doesn't seem to have that alignment padding.
