[PATCH] LD: Always make a SEGMENT_START expression section-relative
Maciej W. Rozycki
macro@mips.com
Thu Jul 19 17:37:00 GMT 2018
Fix an issue with the SEGMENT_START builtin function where its result is
absolute when taken from the default supplied, and section-relative when
taken from a `-T' command-line override. This is against documentation,
inconsistent and unexpected, and with PIE executables gives an incorrect
result with the `__executable_start' symbol.
Make the result of SEGMENT_START always section-relative then.
ld/
* ldexp.c (fold_binary): Always make the result of SEGMENT_START
section-relative.
* testsuite/ld-scripts/segment-start.d: New test.
* testsuite/ld-scripts/segment-start.ld: New test linker script.
* testsuite/ld-scripts/segment-start.s: New test source.
* testsuite/ld-scripts/script.exp: Run the new test.
---
Hi Alan,
I will appreciate if you review this quickly as my time is running out
and this blocks PR ld/21375, because it makes MIPS programs that use the
`__executable_start' symbol stop working with unfixed `ld.so'.
Continuing the saga of absolute symbols...
I got tempted to run final native MIPS testing (which is not something I
do regularly) before committing the fix for PR ld/21375 and lo and behold,
it did trigger a regression. I got a bit concerned I did something wrong
with my fix, but not. Here's what the regression looked like:
Running: tmpdir/pr14525 > tmpdir/pr14525.out
tmpdir/pr14525: symbol lookup error: tmpdir/pr14525: undefined symbol: __executable_start
FAIL: PIE PR ld/14525
This looked remarkably familiar to me, and yes, this is another case of
glibc BZ #23307. Indeed looking at `tmpdir/pr14525' I discovered it has
an absolute `__executable_start' symbol, any my PR ld/21375 fix has duly
exported it so as to make a GOT entry that does not get relocated at load
time:
$ readelf --dyn-syms -A tmpdir/pr14525 | grep __executable_start
14: 00000000 0 NOTYPE GLOBAL DEFAULT ABS __executable_start
00010a10 -32672(gp) 00000000 00000000 NOTYPE ABS __executable_start
$ objdump -d tmpdir/pr14525 | grep -- -32672\(gp\)
814: 8f828060 lw v0,-32672(gp)
$
And my native test environment has an `ld.so' binary that does not carry
my recent BZ #23307 fix.
Obviously `__executable_start' is a program address and is not supposed
to be absolute, not the least because this is going to break with PIE
executables. The culprit turned out to be the SEGMENT_START function
whose result is only section-relative if an override has been made at the
command line (`-Ttext-segment=' in this case). This is covered by the
test case provided.
No regressions across my usual targets, though the new test case does
introduce a couple of new failures for a bunch of obscure targets:
arm-wince-pe +FAIL: SEGMENT_START expression not absolute (overridden)
i686-pe +FAIL: SEGMENT_START expression not absolute (overridden)
mcore-pe +FAIL: SEGMENT_START expression not absolute (overridden)
mmix +FAIL: SEGMENT_START expression not absolute (default)
mmix +FAIL: SEGMENT_START expression not absolute (overridden)
ns32k-netbsd +FAIL: SEGMENT_START expression not absolute (overridden)
pdp11-dec-aout +FAIL: SEGMENT_START expression not absolute (default)
pdp11-dec-aout +FAIL: SEGMENT_START expression not absolute (overridden)
powerpc-beos +FAIL: SEGMENT_START expression not absolute (default)
powerpc-beos +FAIL: SEGMENT_START expression not absolute (overridden)
powerpc-ibm-aix5.2.0 +FAIL: SEGMENT_START expression not absolute (default)
powerpc-ibm-aix5.2.0 +FAIL: SEGMENT_START expression not absolute (overridden)
rs6000-aix4.3.3 +FAIL: SEGMENT_START expression not absolute (default)
rs6000-aix4.3.3 +FAIL: SEGMENT_START expression not absolute (overridden)
rs6000-aix5.1 +FAIL: SEGMENT_START expression not absolute (default)
rs6000-aix5.1 +FAIL: SEGMENT_START expression not absolute (overridden)
sh-pe +FAIL: SEGMENT_START expression not absolute (default)
sh-pe +FAIL: SEGMENT_START expression not absolute (overridden)
tic30-unknown-aout +FAIL: SEGMENT_START expression not absolute (default)
tic30-unknown-aout +FAIL: SEGMENT_START expression not absolute (overridden)
tic30-unknown-coff +FAIL: SEGMENT_START expression not absolute (overridden)
tic4x-coff +FAIL: SEGMENT_START expression not absolute (overridden)
tic54x-coff +FAIL: SEGMENT_START expression not absolute (default)
tic54x-coff +FAIL: SEGMENT_START expression not absolute (overridden)
x86_64-mingw32 +FAIL: SEGMENT_START expression not absolute (overridden)
z8k-coff +FAIL: SEGMENT_START expression not absolute (overridden)
Some of them do not enter `__executable_start' into the symbol table
produced for some reason, and I don't know offhand if that is expected or
not with these targets and/or binary formats. Most crash in LD with a
floating-point exception where `-Ttext-segment=0x10000000' has been used,
with no limitation to a specific binary format, which suggests a generic
problem somewhere, possibly the generic BFD linker, and clearly a bug.
I'd rather pushed this change for these problems to stick out and bring
maintainers' attention and not papered them over with XFAILs right away.
OK to apply then?
Maciej
---
ld/ldexp.c | 4 +++-
ld/testsuite/ld-scripts/script.exp | 4 ++++
ld/testsuite/ld-scripts/segment-start.d | 16 ++++++++++++++++
ld/testsuite/ld-scripts/segment-start.ld | 12 ++++++++++++
ld/testsuite/ld-scripts/segment-start.s | 2 ++
5 files changed, 37 insertions(+), 1 deletion(-)
binutils-ld-segment-start-default-rel.diff
Index: binutils/ld/ldexp.c
===================================================================
--- binutils.orig/ld/ldexp.c 2018-07-19 16:05:39.935270697 +0100
+++ binutils/ld/ldexp.c 2018-07-19 16:05:49.440283166 +0100
@@ -534,6 +534,7 @@ fold_binary (etree_type *tree)
operand, binary.rhs is first operand. */
if (expld.result.valid_p && tree->type.node_code == SEGMENT_START)
{
+ bfd_vma value = expld.result.value;
const char *segment_name;
segment_type *seg;
@@ -550,9 +551,10 @@ fold_binary (etree_type *tree)
"isn't multiple of maximum page size\n"),
segment_name);
seg->used = TRUE;
- new_rel_from_abs (seg->value);
+ value = seg->value;
break;
}
+ new_rel_from_abs (value);
return;
}
Index: binutils/ld/testsuite/ld-scripts/script.exp
===================================================================
--- binutils.orig/ld/testsuite/ld-scripts/script.exp 2018-07-19 16:05:40.881609839 +0100
+++ binutils/ld/testsuite/ld-scripts/script.exp 2018-07-19 16:05:52.128902848 +0100
@@ -231,3 +231,7 @@ foreach test_script $test_script_list {
run_dump_test "align-with-input"
run_dump_test "pr20302"
+
+run_dump_test "segment-start" {{name (default)}}
+run_dump_test "segment-start" {{name (overridden)} \
+ {ld -Ttext-segment=0x10000000}}
Index: binutils/ld/testsuite/ld-scripts/segment-start.d
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ binutils/ld/testsuite/ld-scripts/segment-start.d 2018-07-19 16:05:52.156325561 +0100
@@ -0,0 +1,16 @@
+#PROG: nm
+#name: SEGMENT_START expression not absolute
+#source: segment-start.s
+#ld: -e 0 -u __executable_start -T segment-start.ld
+
+# Make sure `__executable_start' is regular:
+#
+# 10000000 T __executable_start
+#
+# not absolute:
+#
+# 10000000 A __executable_start
+
+#...
+0*10000000 T __executable_start
+#pass
Index: binutils/ld/testsuite/ld-scripts/segment-start.ld
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ binutils/ld/testsuite/ld-scripts/segment-start.ld 2018-07-19 16:05:52.174501255 +0100
@@ -0,0 +1,12 @@
+SECTIONS
+{
+ PROVIDE (__executable_start = SEGMENT_START ("text-segment", 0x10000000));
+ .text : { *(.text) }
+ .data : { *(.data) }
+ .bss : { *(.bss) }
+ .loader : { *(.loader) }
+ .symtab : { *(.symtab) }
+ .strtab : { *(.strtab) }
+ .shstrtab : { *(.shstrtab) }
+ /DISCARD/ : { *(*) }
+}
Index: binutils/ld/testsuite/ld-scripts/segment-start.s
===================================================================
--- /dev/null 1970-01-01 00:00:00.000000000 +0000
+++ binutils/ld/testsuite/ld-scripts/segment-start.s 2018-07-19 16:05:52.194658276 +0100
@@ -0,0 +1,2 @@
+ .text
+ .space 16
More information about the Binutils
mailing list