GAS: Consistently fix labels at the `.end' pseudo-op

Fix a functional regression with the `.end' pseudo-op, introduced with
commit ecb4347adecd ("Last take: approval for MIPS_STABS_ELF killing"),
<https://sourceware.org/ml/binutils/2002-06/msg00443.html>, and commit
dcd410fe1544 ("GNU as 2.14 on IRIX 6: crashes with shared libs"),
<https://sourceware.org/ml/binutils/2003-07/msg00415.html>, which caused
symbol values for labels placed between the end of a function's contents
and its terminating `.end' followed by one of the alignment pseudo-ops
to be different depending on whether either `-mdebug', or `-mno-pdr', or
neither of the command-line options is in effect, be it implied or
specified.

Given debug-label-end.s as follows and the `mips-linux' target we have:

$ cat debug-label-end.s
	.text

	.globl	foo
	.globl	bar
	.align	4, 0
	.ent	foo
foo:
	nop
	.aent	bar
bar:
	.insn
	.end	foo
	.align	4, 0
	.space	16

	.globl	baz
	.ent	baz
baz:
	nop
	.end	baz
	.align	4, 0
	.space	16
$ as -o debug-label-end.o debug-label-end.s
$ readelf -s debug-label-end.o | grep bar
     9: 00000004     0 FUNC    GLOBAL DEFAULT    1 bar
$ as -mdebug -o debug-label-end.o debug-label-end.s
$ readelf -s debug-label-end.o | grep bar
     9: 00000010     0 FUNC    GLOBAL DEFAULT    1 bar
$ as -mno-pdr -o debug-label-end.o debug-label-end.s
$ readelf -s debug-label-end.o | grep bar
     8: 00000010     0 FUNC    GLOBAL DEFAULT    1 bar
$

The reason is the call to `md_flush_pending_output', which in the case
of `mips*-*-*' targets expands to `mips_emit_delays', which in turn
calls `mips_no_prev_insn', which calls `mips_clear_insn_labels', which
clears the list of outstanding labels.  That list is in turn consulted
in `mips_align', called in the interpretation of alignment directives,
and the labels adjusted to the current location.

A call to `md_flush_pending_output' is only made from `s_mips_end' and
then only if `-mpdr' is in effect, which is the default for `*-*-linux*'
and some other `mips*-*-*' targets.  A call to `md_flush_pending_output'
is never made from `ecoff_directive_end', which is used in place of
`s_mips_end' when `-mdebug' is in effect.  Consequently if `-mno-pdr' or
`-mdebug' is in effect the list of outstanding labels makes it through
to any alignment directive that follows and the labels are differently
interpreted depending on the command-lines options used.  And we want
code produced to be always the same.

Call `md_flush_pending_output' unconditionally then in `s_mips_end' and
add such a call from `ecoff_directive_end' as well, as long as the macro
is defined.  While `ecoff_directive_end' is shared among targets, the
only one other than `mips*-*-*' actually using it is `alpha*-*-*' and it
does not define `md_flush_pending_output'.  So the semantics isn't going
to change for it and neither it has to have its `s_alpha_end' updated
or have code in `ecoff_directive_end' conditionalized.

	gas/
	* ecoff.c (ecoff_directive_end) [md_flush_pending_output]: Call
	`md_flush_pending_output'.
	* config/tc-mips.c (s_mips_end) [md_flush_pending_output]: Call
	`md_flush_pending_output' unconditionally.
	* testsuite/gas/mips/debug-label-end-1.d: New test.
	* testsuite/gas/mips/debug-label-end-2.d: New test.
	* testsuite/gas/mips/debug-label-end-3.d: New test.
	* testsuite/gas/mips/debug-label-end.s: New test source.
	* testsuite/gas/mips/mips.exp: Run the new tests.

1 files changed, 4 insertions, 0 deletions

diff --git a/gas/ecoff.c b/gas/ecoff.c
index 5c37312..608d72b 100644
--- a/gas/ecoff.c
+++ b/gas/ecoff.c
@@ -3025,6 +3025,10 @@ ecoff_directive_end (int ignore ATTRIBUTE_UNUSED)
 					 frag_now),
 			     (bfd_vma) 0, (symint_t) 0, (symint_t) 0);
 
+#ifdef md_flush_pending_output
+  md_flush_pending_output ();
+#endif
+
   cur_proc_ptr = (proc_t *) NULL;
 
   (void) restore_line_pointer (name_end);