Native ELF TLS for the Vita target

cmake/recipes/BuildHostDependencies.cmake

@@ -15,7 +15,9 @@ function(toolchain_deps toolchain_deps_dir toolchain_install_dir toolchain_suffi
         "${PROJECT_SOURCE_DIR}/patches/binutils/0002-fix-broken-reloc.patch|1"
         "${PROJECT_SOURCE_DIR}/patches/binutils/0003-fix-elf-vaddr.patch|1"
         "${PROJECT_SOURCE_DIR}/patches/binutils/0004-fix-interworking-veneers.patch|1"
-        "${PROJECT_SOURCE_DIR}/patches/binutils/0005-genscripts-mkdir.patch|1")
+        "${PROJECT_SOURCE_DIR}/patches/binutils/0005-genscripts-mkdir.patch|1"
+        "${PROJECT_SOURCE_DIR}/patches/binutils/0006-tls-in-rodata-segment.patch|1"
+        "${PROJECT_SOURCE_DIR}/patches/binutils/0007-tls-tcb-size-per-emulation.patch|1")
     set(gdb_patch_series
         "${PROJECT_SOURCE_DIR}/patches/gdb.patch|1"
         "${PROJECT_SOURCE_DIR}/patches/gdb-zlib.patch|1"

cmake/recipes/BuildSdkComponents.cmake

@@ -47,6 +47,7 @@ ExternalProject_Add(newlib
     # Multibyte/UTF-8: sin esto setlocale queda clavado en US-ASCII y toda la
     # familia mbrtowc/wcrtomb es inoperante (descubierto con musl libc-test)
     --enable-newlib-mb
+    --enable-newlib-reent-thread-local
     BUILD_COMMAND ${compiler_flags} ${toolchain_tools} ${wrapper_command} $(MAKE)
     INSTALL_COMMAND $(MAKE) install DESTDIR=${CMAKE_INSTALL_PREFIX}
     # Save the commit id for tracking purposes

cmake/recipes/GccCommonArgs.cmake

@@ -13,7 +13,8 @@ set(gcc_patch_series
     "${PROJECT_SOURCE_DIR}/patches/gcc/0001-vita-target.patch|1"
     "${PROJECT_SOURCE_DIR}/patches/gcc/0002-vita-driver.patch|1"
     "${PROJECT_SOURCE_DIR}/patches/gcc/0003-libgomp-vita.patch|1"
-    "${PROJECT_SOURCE_DIR}/patches/gcc/0004-host-compat.patch|1")
+    "${PROJECT_SOURCE_DIR}/patches/gcc/0004-host-compat.patch|1"
+    "${PROJECT_SOURCE_DIR}/patches/gcc/0005-libstdcxx-cxa-thread-atexit.patch|1")
 list(JOIN gcc_patch_series "^" gcc_patch_series_arg)
 if("${CMAKE_CXX_COMPILER_ID}" MATCHES "Clang" AND ${CMAKE_SYSTEM_NAME} STREQUAL "Darwin")
     # GCC on OSX (Clang in diguise) needs more bracket nesting depth to compile gcc
@@ -36,7 +37,7 @@ set(common_gcc_configure_args
     --disable-libstdcxx-pch
     --disable-nls
     --disable-shared
-    --disable-tls
+    --enable-tls
     --with-gnu-as
     --with-gnu-ld
     --with-newlib

patches/binutils/0006-tls-in-rodata-segment.patch

@@ -0,0 +1,64 @@
+diff --git a/ld/scripttempl/elf.sc b/ld/scripttempl/elf.sc
+--- a/ld/scripttempl/elf.sc
++++ b/ld/scripttempl/elf.sc
+@@ -729,6 +729,27 @@
+   ${TEXT_PLT+${PLT_NEXT_DATA+${PLT} ${OTHER_PLT_SECTIONS}}}
+ EOF
+
++if [ -n "${TLS_IN_RODATA}" ]; then
++cat <<EOF
++  /* Thread Local Storage sections. Kept in the read-only segment (not with
++     .data below): the loader only ever reads .tdata to build each thread's
++     TLS block, it's never written in place, and this target's module
++     metadata encodes tls_start as an offset relative to the same segment
++     as the entry point/module info - keeping them together avoids ever
++     needing an offset that crosses into a different segment.
++     (READONLY) drops SHF_WRITE from these output sections (bfd/elf.c
++     only sets it when SEC_READONLY is absent) so the enclosing PT_LOAD
++     doesn't inherit a spurious write flag from otherwise-never-written
++     TLS template data. Set by TLS_IN_RODATA in the emulparams file. */
++  .tdata	${RELOCATING-0} (READONLY) :
++   {
++     ${RELOCATING+${CREATE_SHLIB-PROVIDE_HIDDEN ($(def_symbol "__tdata_start"));}}
++     *(.tdata${RELOCATING+ .tdata.* .gnu.linkonce.td.*})
++   }
++  .tbss		${RELOCATING-0} (READONLY) : { *(.tbss${RELOCATING+ .tbss.* .gnu.linkonce.tb.*})${RELOCATING+ *(.tcommon)} }
++EOF
++fi
++
+ test -n "${RELOCATING}" && cat <<EOF
+   /* Various note sections.  Placed here so that they are always included
+      in the read-only segment and not treated as orphan sections.  The
+@@ -802,7 +823,10 @@
+   .gnu_extab    ${RELOCATING-0} : ONLY_IF_RW { *(.gnu_extab) }
+   .gcc_except_table ${RELOCATING-0} : ONLY_IF_RW { *(.gcc_except_table${RELOCATING+ .gcc_except_table.*}) }
+   .exception_ranges ${RELOCATING-0} : ONLY_IF_RW { *(.exception_ranges${RELOCATING+*}) }
++EOF
+
++if [ -z "${TLS_IN_RODATA}" ]; then
++cat <<EOF
+   /* Thread Local Storage sections.  */
+   .tdata	${RELOCATING-0} :
+    {
+@@ -810,7 +834,10 @@
+      *(.tdata${RELOCATING+ .tdata.* .gnu.linkonce.td.*})
+    }
+   .tbss		${RELOCATING-0} : { *(.tbss${RELOCATING+ .tbss.* .gnu.linkonce.tb.*})${RELOCATING+ *(.tcommon)} }
++EOF
++fi
+
++cat <<EOF
+   ${RELOCATING+$(emit_init_fini_arrays)}
+
+   ${SMALL_DATA_CTOR-${RELOCATING+${CTOR}}}
+diff --git a/ld/emulparams/armvita.sh b/ld/emulparams/armvita.sh
+--- a/ld/emulparams/armvita.sh
++++ b/ld/emulparams/armvita.sh
+@@ -3,3 +3,7 @@
+ TEXT_START_ADDR=0x81000000
+ MAXPAGESIZE="0x10000"
+ COMMONPAGESIZE="0x10000"
++# module_info's tls_start is an offset relative to the same segment as the
++# entry point (see vita-toolchain's sce-elf.c) - keep .tdata/.tbss in the
++# read-only segment, alongside it, instead of the default writable one.
++TLS_IN_RODATA=yes

patches/binutils/0007-tls-tcb-size-per-emulation.patch

@@ -0,0 +1,100 @@
+diff --git a/bfd/elf32-arm.h b/bfd/elf32-arm.h
+--- a/bfd/elf32-arm.h
++++ b/bfd/elf32-arm.h
+@@ -84,6 +84,12 @@
+   int merge_exidx_entries;
+   int cmse_implib;
+   bfd *in_implib_bfd;
++  /* Size of the reserved thread control block ahead of TP in the Local
++     Exec TLS model. 8 on every mainstream ARM EABI target (glibc/newlib's
++     usual "Variant 1" layout); some bare-metal targets place their TLS
++     block directly at the thread pointer instead and need this to be 0.
++     Per-emulation, set via ld/emulparams/<emu>.sh's TCB_SIZE. */
++  int tcb_size;
+ };
+
+ void bfd_elf32_arm_set_target_params
+diff --git a/bfd/elf32-arm.c b/bfd/elf32-arm.c
+--- a/bfd/elf32-arm.c
++++ b/bfd/elf32-arm.c
+@@ -3046,8 +3046,6 @@
+   bool non_a8_stub;
+ };
+
+-/* The size of the thread control block.  */
+-#define TCB_SIZE	8
+
+ /* ARM-specific information about a PLT entry, over and above the usual
+    gotplt_union.  */
+@@ -3385,6 +3383,12 @@
+   /* True if the target system uses FDPIC. */
+   int fdpic_p;
+
++  /* Size of the reserved TCB ahead of the thread pointer for the Local
++     Exec TLS model - see struct elf32_arm_params::tcb_size. Defaults to
++     the standard ARM EABI value of 8; set from ld via
++     bfd_elf32_arm_set_target_params. */
++  bfd_vma tcb_size;
++
+   /* Fixup section. Used for FDPIC.  */
+   asection *srofixup;
+ };
+@@ -4057,6 +4061,7 @@
+   ret->use_rel = true;
+   ret->obfd = abfd;
+   ret->fdpic_p = 0;
++  ret->tcb_size = 8;
+
+   if (!bfd_hash_table_init (&ret->stub_hash_table, stub_hash_newfunc,
+ 			    sizeof (struct elf32_arm_stub_hash_entry)))
+@@ -9030,6 +9035,7 @@
+   globals->fix_arm1176 = params->fix_arm1176;
+   globals->cmse_implib = params->cmse_implib;
+   globals->in_implib_bfd = params->in_implib_bfd;
++  globals->tcb_size = (bfd_vma) params->tcb_size;
+
+   BFD_ASSERT (is_arm_elf (output_bfd));
+   elf_arm_tdata (output_bfd)->no_enum_size_warning
+@@ -9940,12 +9946,13 @@
+ tpoff (struct bfd_link_info *info, bfd_vma address)
+ {
+   struct elf_link_hash_table *htab = elf_hash_table (info);
++  struct elf32_arm_link_hash_table *arm_htab = elf32_arm_hash_table (info);
+   bfd_vma base;
+
+   /* If tls_sec is NULL, we should have signalled an error already.  */
+   if (htab->tls_sec == NULL)
+     return 0;
+-  base = align_power ((bfd_vma) TCB_SIZE, htab->tls_sec->alignment_power);
++  base = align_power (arm_htab->tcb_size, htab->tls_sec->alignment_power);
+   return address - htab->tls_sec->vma + base;
+ }
+
+diff --git a/ld/emultempl/armelf.em b/ld/emultempl/armelf.em
+--- a/ld/emultempl/armelf.em
++++ b/ld/emultempl/armelf.em
+@@ -46,7 +46,8 @@
+   1,				/* fix_arm1176 */
+   -1,				/* merge_exidx_entries */
+   0,				/* cmse_implib */
+-  NULL				/* in_implib_bfd */
++  NULL,				/* in_implib_bfd */
++  ${TCB_SIZE-8}			/* tcb_size */
+ };
+ static char *in_implib_filename = NULL;
+
+diff --git a/ld/emulparams/armvita.sh b/ld/emulparams/armvita.sh
+--- a/ld/emulparams/armvita.sh
++++ b/ld/emulparams/armvita.sh
+@@ -3,6 +3,11 @@
+ TEXT_START_ADDR=0x81000000
+ MAXPAGESIZE="0x10000"
+ COMMONPAGESIZE="0x10000"
++# Vita's TPIDRURO points directly at .tdata, no reserved TCB header ahead of
++# it (confirmed on real hardware: the kernel copies the TLS template to
++# TPIDRURO+0, not the ARM EABI "Variant 1" default of TPIDRURO+8). See
++# elf32_arm_params::tcb_size in bfd/elf32-arm.h.
++TCB_SIZE=0
+ # module_info's tls_start is an offset relative to the same segment as the
+ # entry point (see vita-toolchain's sce-elf.c) - keep .tdata/.tbss in the
+ # read-only segment, alongside it, instead of the default writable one.

patches/gcc/0001-vita-target.patch

@@ -15,7 +15,19 @@ diff --git a/gcc/config/arm/arm.h b/gcc/config/arm/arm.h
 index 30e1d6dc9..bd5f7b19c 100644
 --- a/gcc/config/arm/arm.h
 +++ b/gcc/config/arm/arm.h
-@@ -731,6 +731,10 @@ extern const int arm_arch_cde_coproc_bits[];
+@@ -87,8 +87,10 @@ extern tree arm_bf16_ptr_type_node;
+ #undef  CPP_SPEC
+ #define CPP_SPEC "%(subtarget_cpp_spec)"
+
+ #ifndef CC1_SPEC
+-#define CC1_SPEC ""
++/* The loader hands each thread its TLS block at a fixed offset from
++   TPIDRURO; every model but local-exec wants a GOT it never fills.  */
++#define CC1_SPEC "%{!ftls-model=*:-ftls-model=local-exec}"
+ #endif
+
+ /* This macro defines names of additional specifications to put in the specs
+@@ -731,6 +733,10 @@ extern const int arm_arch_cde_coproc_bits[];
  #define WCHAR_TYPE_SIZE BITS_PER_WORD
  #endif
 
@@ -26,7 +38,7 @@ index 30e1d6dc9..bd5f7b19c 100644
  /* Sized for fixed-point types.  */
 
  #define SHORT_FRACT_TYPE_SIZE 8
-@@ -1996,7 +2000,7 @@ enum arm_auto_incmodes
+@@ -1996,7 +2002,7 @@ enum arm_auto_incmodes
  /* signed 'char' is most compatible, but RISC OS wants it unsigned.
     unsigned is probably best, but may break some code.  */
  #ifndef DEFAULT_SIGNED_CHAR

patches/gcc/0005-libstdcxx-cxa-thread-atexit.patch

@@ -0,0 +1,29 @@
+--- a/libstdc++-v3/configure.ac
++++ b/libstdc++-v3/configure.ac
+@@ -339,6 +339,11 @@
+   # on a hosted environment.
+   if test "x${with_newlib}" = "xyes"; then
+     os_include_dir="os/newlib"
++    case "${host}" in
++      arm-vita-eabi*)
++        AC_DEFINE(HAVE___CXA_THREAD_ATEXIT)
++        ;;
++    esac
+     AC_DEFINE(HAVE_HYPOT)
+
+     # GLIBCXX_CHECK_STDLIB_SUPPORT
+--- a/libstdc++-v3/configure
++++ b/libstdc++-v3/configure
+@@ -27343,6 +27343,12 @@
+   # on a hosted environment.
+   if test "x${with_newlib}" = "xyes"; then
+     os_include_dir="os/newlib"
++    case "${host}" in
++      arm-vita-eabi*)
++        $as_echo "#define HAVE___CXA_THREAD_ATEXIT 1" >>confdefs.h
++
++        ;;
++    esac
+     $as_echo "#define HAVE_HYPOT 1" >>confdefs.h
+
+