syscall: 0.432
instruction: 0.356
runtime: 0.212



ppc64le linux user emulation w/ 64KiB pages seems broken since v5.0.0
Description of problem:
[Our (snmalloc's)](https://github.com/microsoft/snmalloc) CI includes running a PowerPC64 little-endian Linux build inside qemu, running with 64KiB pages as, at least, Debian runs them by default.  As reported [over there](https://github.com/microsoft/snmalloc/issues/576), this broke when GitHub's CI runners moved from Ubuntu Focal (20.04) to Jammy (22.04), bringing qemu from v4.2 to v6.2.

The failing test case appears to die of an erroneous `SIGSEGV` `SEGV_MAPERR`:
```
--- SIGSEGV {si_signo=SIGSEGV, si_code=1, si_addr=0x0000004001be5000} ---
```
despite that address nominally being mapped by the last memory syscall to touch that area
```
openat(AT_FDCWD,"/usr/powerpc64le-linux-gnu/lib/libstdc++.so.6",O_RDONLY|O_CLOEXEC) = 4
[...]
mmap(0x0000004001bd0000,131072,PROT_READ|PROT_WRITE,MAP_PRIVATE|MAP_DENYWRITE|MAP_FIXED,4,0x2f0000) = 0x4001bd0000
```

Bisection reveals that the breakage first occurred with 4dcf078f094d436866ef793aa25c96fba85ac8d0, though I suspect this is merely the commit that exposes some underlying bug rather than being the actual root cause.
Steps to reproduce:
Run a ppc64el Linux executable under `qemu-user` with `-p 65536`.
Additional information:
Please advise what more would be useful.