On Tue, Sep 9, 2025 at 9:04 AM Alexander Monakov <amona...@ispras.ru> wrote: > > > On Tue, 9 Sep 2025, H.J. Lu wrote: > > > > > to restore assert if not compiling for shared library. When compiling > > > > for shared library, the recomputed model may be weaker. > > > > > > I don't understand why (for shared libraries), can you explain? > > > > Another shared library or executable may override any global > > symbols with the default visibility in a shared library. > > I know that symbols with default visibility can be interposed, but that > doesn't > answer the question. Do you have a testcase that demonstrates the issue?
TLS tests were submitted at https://patchwork.sourceware.org/project/gcc/patch/20250908221002.3223418-1-hjl.to...@gmail.com/ The key point is that compiler doesn't know if PIC code will be in shared library or executable. If there is a programmer error, linker will issue an error: [hjl@gnu-cfl-3 gcc]$ cat /tmp/x.c __attribute__ ((tls_model ("local-exec"))) __thread int i; int * foo (void) { return &i; } [hjl@gnu-cfl-3 gcc]$ ./xgcc -B./ -fPIC -O2 /tmp/x.c -c [hjl@gnu-cfl-3 gcc]$ ld -shared x.o ld: x.o: relocation R_X86_64_TPOFF32 against symbol `i' can not be used when making a shared object; recompile with -fPIC ld: failed to set dynamic section sizes: bad value [hjl@gnu-cfl-3 gcc]$ In this case, linker suggestion isn't very clear. The issue is the wrong TLS model, not PIC. > Otherwise, let's say we are talking about > > extern __thread int tls_undef; > static __thread int tls_local; > /****/ __thread int tls_global; > > With -fpic GCC initially selects local-dynamic for tls_local, global-dynamic > for tls_undef and tls_global. Why would recomputed model for tls_local be > weaker? > > > __thread int x; > > > > __attribute__((common)) > > __thread int x; > > > > int *f() > > { > > return &x; > > } > > [hjl@gnu-zen4-1 gcc]$ ./xgcc -B./ -O2 -std=c11 -fno-common -S /tmp/x.c > > -fdump-ipa-whole-program > > [hjl@gnu-zen4-1 gcc]$ cat x.c.084i.whole-program > > > > Marking local functions: > > > > > > Marking externally visible functions: f/3 > > > > > > Marking externally visible variables: x/1 > > > > Clearing variable flags: > > > > Reclaiming functions: > > Reclaiming variables: > > Clearing address taken flags: > > Symbol table: > > > > f/3 (f) > > Type: function definition analyzed > > Visibility: externally_visible semantic_interposition public > > References: x/1 (addr) > > Referring: > > Availability: available > > Function flags: count:1073741824 (estimated locally) body > > Called by: > > Calls: > > x/1 (x) > > Type: variable definition analyzed > > Visibility: externally_visible semantic_interposition public common > > References: > > Referring: f/3 (addr) > > Availability: overwritable > > Varpool flags: tls-local-exec > > Thanks. Unrelated to this particular patch, but this is questionable: we have > local-exec for an interposable (Availability: overwritable) symbol. Perhaps Definition in executable isn't interposable. > it's better to error out on the redeclaration, when the previous declaration > set a stronger TLS model? > > Alexander -- H.J.
