This is a problem discovered while working on new FreeBSD plugin for LLDB, as contracted to Moritz Systems by FreeBSD Foundation.

TL;DR: We'd need to replace 'plain' *xsave* and *xrstor* calls used to populate GETFPREGS/SETFPREGS on amd64 (or possibly all of them) with *xsave*64 and *xrstor*64.  This changes the FIP/FDP representation inside the struct not to be truncated to 32 bits (without changing anything else).


Long version:

The fxsave, xsave... and fxrstor, xrstor... instructions have two variants on amd64.  The variant without prefix uses a 32-bit compatible x87 register dump structure while the variant prefixed by rex.w=1 uses a 64-bit x87 register dump.  The latter can be done in gas by appending '64' suffix to the command, e.g. fxsave64.

The only difference in these two variants is how FIP/FDP registers are written.  In the 32-bit compatible variant, they are written as a 16-bit segment register (FCS/FDS) and an actual pointer truncated to 32 bits (FIP[31:0], FDP[31:0]), plus 16 bits of padding.  In the 64-bit version, they are written as full 64-bit register instead.

FreeBSD currently calls the 32-bit compatible variant of instructions.  This is problematic because it means that pointers exceeding 2^32 are truncated.  Switching to the 64-bit variant would solve that.

This is a potentially breaking change but we don't think it is likely to break anything.  The FIP/FDP registers are used to locate instruction and its memory operand (if any) when software handling of x87 exceptions is used.  We don't think modern debuggers are used often in that context and even if they are, the current state is broken as they get a truncated pointer.

We'd lose access to respective segment registers but unless I'm mistaken they aren't really used much these days.

The problem also affects current versions of gdb.  Linux gdb and lldb are using the 64-bit variant.  Although it is displayed split as fiseg/fioff, foseg/fooff, the *seg register does not contain the 16-bit segment but instead FIP[63:32], FDP[63:32].