On 10/25/22 18:29, Ivan Tetyushkin wrote:
> ---
>   gdb/testsuite/gdb.base/print-file-var.exp | 4 +++-
>   1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/gdb/testsuite/gdb.base/print-file-var.exp b/gdb/testsuite/gdb.base/print-file-var.exp
> index 9abe87d7758..73137630fed 100644
> --- a/gdb/testsuite/gdb.base/print-file-var.exp
> +++ b/gdb/testsuite/gdb.base/print-file-var.exp
> @@ -42,6 +42,8 @@ proc test {hidden dlopen version_id_main lang} {
>       set libobj1 [standard_output_file ${lib1}$suffix.so]
>       set libobj2 [standard_output_file ${lib2}$suffix.so]
>   
> +    set runtimelibobj2 [get_runtime_file $libobj2]
> +
>       set lib_opts { debug $lang }
>       lappend lib_opts "additional_flags=-DHIDDEN=$hidden"
>   
> @@ -60,7 +62,7 @@ proc test {hidden dlopen version_id_main lang} {
>       set link_opts [list debug shlib=${libobj1}]
>   
>       if {$dlopen} {
> -	lappend main_opts "additional_flags=-DSHLIB_NAME=\"$libobj2\""
> +	lappend main_opts "additional_flags=-DSHLIB_NAME=\"$runtimelibobj2\""
>   	lappend link_opts "shlib_load"
>       } else {
>   	lappend link_opts "shlib=${libobj2}"

I get this test-case passing by avoiding to use an absolute file name:
...
diff --git a/gdb/testsuite/gdb.base/print-file-var.exp 
b/gdb/testsuite/gdb.base/print-file-
var.exp
index 9abe87d7758..338840cb05e 100644
--- a/gdb/testsuite/gdb.base/print-file-var.exp
+++ b/gdb/testsuite/gdb.base/print-file-var.exp
@@ -60,7 +60,7 @@ proc test {hidden dlopen version_id_main lang} {
      set link_opts [list debug shlib=${libobj1}]

      if {$dlopen} {
-       lappend main_opts "additional_flags=-DSHLIB_NAME=\"$libobj2\""
+       lappend main_opts "additional_flags=-DSHLIB_NAME=\"[file tail 
$libobj2]\""
         lappend link_opts "shlib_load"
      } else {
         lappend link_opts "shlib=${libobj2}"
...
which means that the file is found relative to $ORIGIN, as is the case 
for $dlopen == 0.

I'm not entirely happy with this fix, because what we'd rather want is 
to use the name as returned by remote_download, but that's done in 
gdb_load_shlib later on (which also handles shlib_target_file, so 
actually, the name as returned by this proc is the one we really want), 
so it's not available yet at the time we do the compilation.

I've thought about allowing gdb_load_shlib to be called without gdb 
instance, and for cases where there's no instance, registering the 
solib-search-path setting to be done at gdb start.  This will allow us 
to move the gdb_load_shlib calls to before the compilation, such that we 
can use the result to set -DSHLIB_NAME.  But I think it's a solution 
bound to cause confusion because things happen under the hood.

Alternatively, we can try to not pass in the name into compilation 
(working around only some of the problems related to remote host 
testing, see https://sourceware.org/bugzilla/show_bug.cgi?id=16947), but 
that also has its limitations: we need to be able to patch target memory 
or the ability to use argv.

Perhaps splitting up the functionality in gdb_load_shlib makes the most 
sense.  By default, it'll do the same as before, but by calling it 
twice, once with -only-download, and once with -no-download can we get 
the solib name before starting gdb.

I've given this a try in patch attached below.

Ivan, could you check if the patch fixes the gdb.base/print-file-var.exp 
test-case for you?

Thanks,
- Tom