libc: Use the initial-exec TLS model for libc on amd64 and arm64

[markj]

Summary

This permits more efficient accesses of thread-local variables, which
are heavily used at least by jemalloc and locale-aware code. Note that
on amd64, jemalloc's thread-local variables already have their TLS model
overridden by defining JEMALLOC_TLS_MODEL.

PR: 255840
Suggested by: jrtc27

Test Plan

A test program which malloc()s a 64 byte buffer and frees it, in a loop,
completes about 25% faster than before on a EC2 Graviton instance. On
the same system I see no significant difference in buildkernel times.

On an amd64 NUC, a sort -n of some test data was roughly 10% faster with
this change. On the Graviton it was roughly 5% faster.

[jrtc27]

lib/libc/Makefile

57

Why only these architectures? All of them can benefit from this change.

[markj]

lib/libc/Makefile

57

Mostly for lack of testing. I can try powerpc, i386 and riscv but can't easily test 32-bit arm or mips. In the meantime, it seemed reasonable to enable it first on "tier 1" platforms.

In particular, I had to make a change to rtld to avoid breaking systems with this change on arm64: D31069.

[kib]

So this assumes that libc is never dlopened, which is practically true but not neccessary hold.

[markj]

Enable -ftls-model=initial-exec on i386 and riscv.
Add a comment.

[kib]

lib/libc/Makefile

57

s/safely/reasonably/

59

should aarch64 come first, alphabetically?

[markj]

Address feedback.