[PATCH] ld: testsuite: xfail ld-elf/compress1a etc. on Solaris/sparcv9 [PR25802]

[Rainer Orth]

Hi Jan,

> On 29.09.2025 14:59, Rainer Orth wrote:
>> A couple of tests FAIL on Solaris/sparcv9:
>> 
>> FAIL: ld-elf/compress1a
>> FAIL: ld-elf/compressed1a
>> FAIL: ld-elf/eh5
>> FAIL: --gc-sections with multiple debug sections for a function section
>> 
>> The symptom is always the same:
>> 
>> tmpdir/compress1.o:(.debug_info+0x10): relocation truncated to fit: R_SPARC_UA32 against `.text'
>> 
>> and several more.  However, the tests link just fine with Solaris ld.
>> 
>> As analysed in the PR, this patch xfail's the first three and runs the
>> last with Ttext=0x0x80000000 (the address used in
>> /usr/lib/ld/sparcv9/map.below4G) to allow the link to succeed.
>
> I'm not very happy about such XFAILs, and other than Nick had requested
> in the bug report they're also entirely uncommented. Plus as per comment

sorry I missed that, probably in part because none of the existing xfails
had comments.

> 9 there (post-dating the patch submission afaict), the problem looks
> like it wants dealing with differently anyway.

Not really: I've since learned that Solaris ld linking the tests without
errors (actually only the first three ones, the last is about R_SPARC_32
instead of R_SPARC_UA32 and fails with both linkers) is a historical
oversight going 30 years back.  While I was able to mimic that behaviour
in bfd as described in the PR, Ali Bahrami strongly advised against
following Solaris ld behaviour here which violates the SPARC psABI.

To summarize the situation:

* Since .eh_frame needs to read-write on Solaris except Solaris/amd64 as
  described in gas/config/te-solaris.h, it lands in the data segment, so
  the R_SPARC_UA32 and R_SPARC_32 relocations in .eh_frame against .text
  are out of range.  Thus the "relocation truncated to fit" errors are
  benign.

* One could avoid those by linking the affected tests with
  -Ttext=0x80000000, matching Solaris /usr/lib/ld/map.below4G.

* There are two options to handle this:

** Just xfail all four affected tests on sparcv9-*-solaris2* since the
   error is correct and nothing can be done it about without special
   handling (which wouldn't be used in the wild).

** I found a way to apply the linker flag above to only those four tests
   on Solaris/sparcv9:

Define a new proc in ld-lib.exp

# Return ldflags to load 64-bit Solaris/SPARC programs below 4 GB.  This
# matches /usr/lib/ld/map.below4G.
# Applies to both sparcv9 and amd64, but only needed on SPARC.
proc sol2sparcv9_below4G_ldflags { } {
    if [istarget sparcv9*-*-solaris2*] {
	return "-Ttext=0x80000000"
    }
    return ""
}

and apply it to all for affected tests like this:

diff --git a/ld/testsuite/ld-elf/compress1a.d b/ld/testsuite/ld-elf/compress1a.d
--- a/ld/testsuite/ld-elf/compress1a.d
+++ b/ld/testsuite/ld-elf/compress1a.d
@@ -1,6 +1,7 @@
 #source: compress1.s
 #as: --compress-debug-sections
-#ld: -e func_cu2
+# See PR ld/25802 for the Solaris/sparcv9 handling.
+#ld: -e func_cu2 [sol2sparcv9_below4G_ldflags]

since you can call arbitrary Tcl procs with "#ld: ".  While this works
as expected, it seems completely overengineered to me.

Which variant should I use then?

	Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University