Remove some load/store instructions from the dynamic tlsdesc resolver
fast path. This gives around 20% faster tls access in dlopened shared
libraries (assuming glibc ran out of static tls space).
* sysdeps/aarch64/dl-tlsdesc.S (_dl_tlsdesc_dynamic): Optimize.
Backtrace through _dl_tlsdesc_resolve_rela was broken because the offset
of x30 from cfa was not in the debug info.
Add enough annotation so backtracing from the dynamic linker through
tlsdesc entry points works and the debugger shows registers correctly.
Some of the cfi annotations used incorrect sign.
* sysdeps/aarch64/dl-tlsdesc.S (_dl_tlsdesc_return_lazy): Fix
cfi_adjust_cfa_offset argument.
(_dl_tlsdesc_undefweak, _dl_tlsdesc_dynamic): Likewise.
(_dl_tlsdesc_resolve_rela, _dl_tlsdesc_resolve_hold): Likewise.
Lazy TLSDESC initialization needs to be synchronized with concurrent TLS
accesses. The TLS descriptor contains a function pointer (entry) and an
argument that is accessed from the entry function. With lazy initialization
the first call to the entry function updates the entry and the argument to
their final value. A final entry function must make sure that it accesses an
initialized argument, this needs synchronization on systems with weak memory
ordering otherwise the writes of the first call can be observed out of order.
There are at least two issues with the current code:
tlsdesc.c (i386, x86_64, arm, aarch64) uses volatile memory accesses on the
write side (in the initial entry function) instead of C11 atomics.
And on systems with weak memory ordering (arm, aarch64) the read side
synchronization is missing from the final entry functions (dl-tlsdesc.S).
This patch only deals with aarch64.
* Write side:
Volatile accesses were replaced with C11 relaxed atomics, and a release
store was used for the initialization of entry so the read side can
synchronize with it.
* Read side:
TLS access generated by the compiler and an entry function code is roughly
ldr x1, [x0] // load the entry
blr x1 // call it
entryfunc:
ldr x0, [x0,#8] // load the arg
ret
Various alternatives were considered to force the ordering in the entry
function between the two loads:
(1) barrier
entryfunc:
dmb ishld
ldr x0, [x0,#8]
(2) address dependency (if the address of the second load depends on the
result of the first one the ordering is guaranteed):
entryfunc:
ldr x1,[x0]
and x1,x1,#8
orr x1,x1,#8
ldr x0,[x0,x1]
(3) load-acquire (ARMv8 instruction that is ordered before subsequent
loads and stores)
entryfunc:
ldar xzr,[x0]
ldr x0,[x0,#8]
Option (1) is the simplest but slowest (note: this runs at every TLS
access), options (2) and (3) do one extra load from [x0] (same address
loads are ordered so it happens-after the load on the call site),
option (2) clobbers x1 which is problematic because existing gcc does
not expect that, so approach (3) was chosen.
A new _dl_tlsdesc_return_lazy entry function was introduced for lazily
relocated static TLS, so non-lazy static TLS can avoid the synchronization
cost.
[BZ #18034]
* sysdeps/aarch64/dl-tlsdesc.h (_dl_tlsdesc_return_lazy): Declare.
* sysdeps/aarch64/dl-tlsdesc.S (_dl_tlsdesc_return_lazy): Define.
(_dl_tlsdesc_undefweak): Guarantee TLSDESC entry and argument load-load
ordering using ldar.
(_dl_tlsdesc_dynamic): Likewise.
(_dl_tlsdesc_return_lazy): Likewise.
* sysdeps/aarch64/tlsdesc.c (_dl_tlsdesc_resolve_rela_fixup): Use
relaxed atomics instead of volatile and synchronize with release store.
(_dl_tlsdesc_resolve_hold_fixup): Use relaxed atomics instead of
volatile.
* elf/tlsdeschtab.h (_dl_tlsdesc_resolve_early_return_p): Likewise.
This patch moves the AArch64 port to the main sysdeps hierarchy. The
move is essentially:
git mv ports/sysdeps/aarch64 sysdeps/aarch64
git mv ports/sysdeps/unix/sysv/linux/aarch64 sysdeps/unix/sysv/linux/aarch64
The README is updated and I've updated ChangeLog.aarch64 along the
lines of the ARM move. The AArch64 build has been tested to confirm
that there were no changes in objdump -dr output or the shared
objects.
