binutils: fix the ARC TLS initial-exec addend for 2.47

Linking a PIE with initial-exec TLS produces a relocation whose addend repeats
the symbol's offset in the TLS block:

  dynsym:  bar   st_value = 4
  reloc:   R_ARC_TLS_TPOFF   bar + 4

The loader adds both -- uClibc-ng and glibc alike, see glibc's
sysdeps/arc/dl-machine.h -- so every TLS variable except the one at offset 0
resolves four bytes (or however far in it sits) past its own storage.  On
uClibc-ng's suite that is tst-tls1-pie and tst-tls2-pie, which read the TCB
instead of "bar" and were the last two failures on the new arc-archs target.

bfd/arc-got.h takes the GOT slot as the addend for initial-exec, and for a
symbol the linker can resolve locally that slot already holds the offset.
Only a relocation without a symbol needs the value there, because the loader
then has nothing to look up, so restrict it to that case.

Verified by building 2.47 with the patch: the relocation becomes "bar + 0",
which is also what Synopsys' own toolchain (binutils 2.38.50) emits, and both
tests pass with an unmodified loader.  Mainline has read the slot since
08759e0fc8b ("Fixes done to TLS.", 2016), and 2.41 behaves like 2.47.

Signed-off-by: Ramin Moussavi <lordrasmus@gmail.com>
Signed-off-by: Waldemar Brodkorb <wbx@openadk.org>

toolchain/binutils/patches/2.47/arc-tls-tpoff-addend.patch

@@ -0,0 +1,40 @@
+arc: do not fold the GOT slot into the addend of R_ARC_TLS_TPOFF
+
+The dynamic loader computes the thread pointer offset as
+l_tls_offset + st_value + r_addend -- uClibc-ng and glibc do it the same way,
+see glibc's sysdeps/arc/dl-machine.h.  Passing the GOT slot as the addend on
+top of a relocation that names the symbol makes the loader count the symbol's
+offset twice, because for a symbol that is not preemptible -- in a PIE, say --
+the linker has already resolved that offset into the slot.
+
+Every TLS variable except the one at offset 0 then reads from the wrong place:
+uClibc-ng's tst-tls1-pie and tst-tls2-pie fail on arc for exactly that reason,
+"bar" being four bytes into the block.  Synopsys' own toolchain (binutils
+2.38.50) emits addend 0 here and is unaffected; mainline has read the slot
+since 08759e0fc8b in 2016.
+
+A relocation without a symbol still needs the value, because the loader then
+has nothing to look up, so keep it for that case.
+
+diff --git a/bfd/arc-got.h b/bfd/arc-got.h
+index 5e9e48f7eb5..17be7a20122 100644
+--- a/bfd/arc-got.h
++++ b/bfd/arc-got.h
+@@ -514,8 +514,16 @@ GOT_OFFSET = %#lx, GOT_VMA = %#lx, INDEX = %ld, ADDEND = 0x0\n",
+       if (e == TLS_GOT_MOD_AND_OFF || e == TLS_GOT_OFF)
+ 	{
+ 	  bfd_vma addend = 0;
+-	  if (list->type == GOT_TLS_IE)
++	  if (list->type == GOT_TLS_IE && dynindx == 0)
+ 	  {
++	    /* The dynamic loader computes the thread pointer offset as
++	       l_tls_offset + st_value + r_addend (uClibc-ng and glibc both do,
++	       see sysdeps/arc/dl-machine.h).  Passing the GOT slot -- which
++	       holds the link time resolved offset whenever the symbol is not
++	       preemptible, as in a PIE -- as the addend on top of a relocation
++	       that names the symbol makes it count that offset twice.  Only a
++	       relocation without a symbol needs the value here, because then
++	       the loader has nothing to look up.  */
+ 	    addend = bfd_get_32 (output_bfd,
+ 				 htab->sgot->contents + got_offset);
+ 	  }