[PATCH 3/3] gas: avoid (scrubber) diagnostics for stuff past .end
Jan Beulich
jbeulich@suse.com
Mon Sep 2 09:35:24 GMT 2024
What's past an active .end directive (when that has its default purpose)
is supposed to be entirely ignored. That should be true not just for
regular processing, but also for "pre-processing" (aka scrubbing). A
complication is that such a directive may of course occur inside a
(false) conditional or a macro definition. To deal with that make sure
we can continue as usual if called another time.
Note however that .end inside a macro will still have the full macro
body expanded; dealing with that would require further (perhaps
intrusive) adjustments in sb_scrub_and_add_sb() and/or callers thereof.
However, at least some of the warnings issued by do_scrub_chars() are
unlikely to occur when expanding a macro. (If we needed to go that far,
presumably .exitm would also want recognizing.)
---
Various targets assign different meaning to .end. For those the extra
logic could in principle be skipped. Thoughts?
The "inside a false conditional or macro definition" aspect applies to
the .mri special handling as well, but isn't handled there. Nor is that
handling covering that s_mri() actually parses an expression, rather
than just looking for 0 and 1 as arguments. Nor is that handling dealing
with the absence of a leading . in the NO_PSEUDO_DOT case.
--- a/gas/app.c
+++ b/gas/app.c
@@ -58,6 +58,15 @@ static const char symver_pseudo[] = ".
static const char * symver_state;
#endif
+/* The pseudo-op (without leading dot) at which we want to (perhaps just
+ temporarily) stop processing. See the comments in do_scrub_chars(). */
+static const char end_pseudo[] = "end ";
+static const char * end_state;
+
+/* Whether, considering the state at start of assembly, NO_PSEUDO_DOT is
+ active. */
+static bool no_pseudo_dot;
+
static char last_char;
#define LEX_IS_SYMBOL_COMPONENT 1
@@ -161,6 +170,12 @@ do_scrub_begin (int m68k_mri ATTRIBUTE_U
{
const char *p;
+ /* Latch this once at start. xtensa uses a hook function, yet context isn't
+ meaningful for scrubbing (or else we'd need to sync scrubber behavior as
+ state changes). */
+ if (lex['/'] == 0)
+ no_pseudo_dot = NO_PSEUDO_DOT;
+
#ifdef TC_M68K
scrub_m68k_mri = m68k_mri;
@@ -287,6 +302,7 @@ struct app_save
int add_newlines;
char * saved_input;
size_t saved_input_len;
+ const char * end_state;
#ifdef TC_M68K
int scrub_m68k_mri;
const char * mri_state;
@@ -317,6 +333,7 @@ app_push (void)
memcpy (saved->saved_input, saved_input, saved_input_len);
saved->saved_input_len = saved_input_len;
}
+ saved->end_state = end_state;
#ifdef TC_M68K
saved->scrub_m68k_mri = scrub_m68k_mri;
saved->mri_state = mri_state;
@@ -357,6 +374,7 @@ app_pop (char *arg)
saved_input_len = saved->saved_input_len;
free (saved->saved_input);
}
+ end_state = saved->end_state;
#ifdef TC_M68K
scrub_m68k_mri = saved->scrub_m68k_mri;
mri_state = saved->mri_state;
@@ -805,6 +823,43 @@ do_scrub_chars (size_t (*get) (char *, s
recycle:
+ /* We need to watch out for .end directives: We should in particular not
+ issue diagnostics for anything after an active one. */
+ if (end_state == NULL)
+ {
+ if ((state == 0 || state == 1)
+ && (ch == '.'
+ || (no_pseudo_dot && ch == end_pseudo[0])))
+ end_state = end_pseudo + (ch != '.');
+ }
+ else if (ch != '\0'
+ && (*end_state == ch
+ /* Avoid triggering on directives like .endif or .endr. */
+ || (*end_state == ' ' && !IS_SYMBOL_COMPONENT (ch))))
+ {
+ if (IS_NEWLINE (ch) || IS_LINE_SEPARATOR (ch))
+ goto end_end;
+ ++end_state;
+ }
+ else if (*end_state != '\0')
+ /* We did not get the expected character, or we didn't
+ get a valid terminating character after seeing the
+ entire pseudo-op, so we must go back to the beginning. */
+ end_state = NULL;
+ else if (IS_NEWLINE (ch) || IS_LINE_SEPARATOR (ch))
+ {
+ end_end:
+ /* We've read the entire pseudo-op. If this is the end of the line,
+ bail out now by (ab)using the output-full path. This allows the
+ caller to process input up to here and terminate processing if this
+ directive is actually active (not on the false branch of a
+ conditional and not in a macro definition). */
+ end_state = NULL;
+ state = 0;
+ PUT (ch);
+ goto tofull;
+ }
+
#if defined TC_ARM && defined OBJ_ELF
/* We need to watch out for .symver directives. See the comment later
in this function. */
@@ -1445,7 +1500,7 @@ do_scrub_chars (size_t (*get) (char *, s
#if defined TC_ARM && defined OBJ_ELF
&& symver_state == NULL
#endif
- )
+ && end_state == NULL)
{
char *s;
ptrdiff_t len;
--- /dev/null
+++ b/gas/testsuite/gas/all/end.l
@@ -0,0 +1,3 @@
+# No diagnostics should appear for anything past .end.
+>1<
+>2<
--- /dev/null
+++ b/gas/testsuite/gas/all/end.s
@@ -0,0 +1,11 @@
+ .if 0
+ .end a b c
+ .endif
+
+ .irpc n,12
+ .print ">\n<"
+ .endr
+
+ .end q r, s
+ "\z"
+ äöü'\
\ No newline at end of file
--- /dev/null
+++ b/gas/testsuite/gas/all/end-no-dot.l
@@ -0,0 +1,3 @@
+# No diagnostics should appear for anything past "end".
+>3<
+>4<
--- /dev/null
+++ b/gas/testsuite/gas/all/end-no-dot.s
@@ -0,0 +1,11 @@
+ if 0
+ end a b c
+ endif
+
+ irpc n,34
+ print ">\n<"
+ endr
+
+ end q r, s
+ "\z"
+ äöü'\
\ No newline at end of file
--- a/gas/testsuite/gas/all/gas.exp
+++ b/gas/testsuite/gas/all/gas.exp
@@ -464,6 +464,30 @@ switch -glob $target_triplet {
run_dump_test weakref1w
}
}
+
+# .end works differently on some targets. Also make sure to test the dot-less
+# form on targets setting NO_PSEUDO_DOT (and not overriding the directive).
+switch -glob $target_triplet {
+ alpha*-*-* { }
+ hppa*-*-* { }
+ iq2000-*-* { }
+ microblaze-*-* { }
+ mips*-*-* { }
+ score*-*-* { }
+ xtensa*-*-* { }
+ m68hc1*-*-* -
+ s12z-*-* -
+ spu-*-* -
+ xgate-*-* -
+ z80-*-* {
+ run_list_test "end"
+ run_list_test "end-no-dot"
+ }
+ default {
+ run_list_test "end"
+ }
+}
+
gas_test_error "weakref2.s" "" "e: would close weakref loop: e => a => b => c => d => e"
gas_test_error "weakref3.s" "" "a: would close weakref loop: a => b => c => d => e => a"
gas_test_error "weakref4.s" "" "is already defined"
More information about the Binutils
mailing list