tree 8df4f3d92d44dfad2307cc1203982ed1a61e8537
parent b80b72c06c29d257b60b4002bd9d6c40dec94ec9
author Jan Beulich <jbeulich@suse.com> 1659079607 +0200
committer Jan Beulich <jbeulich@suse.com> 1659079607 +0200

Arm64: re-work PR gas/27217 fix

The original approach has resulted in anomalies when . is involved in an
operand of one of the affected insns. We cannot leave . unresolved, or
else it'll be resolved at the end of assembly, then pointing to the
address of a section rather than at the insn of interest. Undo part of
the original change and instead check whether a relocation cannot be
omitted in md_apply_fix().

By resolving the expressions again, equates (see the adjustment of the
respective testcase) will now be evaluated, and hence relocations
against absolute addresses be emitted. This ought to be okay as long as
the equates aren't global (and hence can't be overridden). If a need
for such arises, quite likely the only way to address this would be to
invent yet another expression evaluation mode, leaving everything
_except_ . un-evaluated.

There's a further anomaly in how transitive equates are handled. In

	.set x, 0x12345678
	.eqv bar, x
foo:
	adrp	x0, x
	add	x0, x0, :lo12:x

	adrp	x0, bar
	add	x0, x0, :lo12:bar

the first two relocations are now against *ABS*:0x12345678 (as said
above), whereas the latter two relocations would be against x. (Before
the change here, the first two relocations are against x and the latter
two against bar.) But this is an issue seen elsewhere as well, and would
likely require adjustments in the target-independent parts of the
assembler instead of trying to hack around this for every target.