tls: fix local-exec addresses and external symbols

Preserve symbolic offsets when materializing i386 TLS addresses, generate ARM64 TLS addresses from the thread pointer, and carry RISC-V addends on both halves of the TPREL relocation pair.

Keep undefined ELF TLS symbols typed as STT_TLS so objects produced by TCC can be linked by the system linker. Extend the TLS regression coverage for narrow signed and unsigned accesses, address-taking, runtime indexing, thread isolation, large RISC-V offsets, unsupported runtimes, and external TLS definitions.

arm64-gen.c

@@ -468,20 +468,8 @@ static void arm64_strv(int sz_, int dst, int bas, uint64_t off)
     }
 }
 
-static void arm64_sym(int r, Sym *sym, unsigned long addend)
+static void arm64_add_offset(int r, unsigned long addend)
 {
-#ifdef TCC_TARGET_PE
-    /* PE links symbol addresses directly; there is no ELF-style GOT here. */
-    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_PREL_PG_HI21, 0);
-    o(ARM64_ADRP | r);            // adrp xr, #sym
-    greloca(cur_text_section, sym, ind, R_AARCH64_ADD_ABS_LO12_NC, 0);
-    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | r); // add xr, xr, #sym
-#else
-    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_GOT_PAGE, 0);
-    o(ARM64_ADRP | r);            // adrp xr, #sym
-    greloca(cur_text_section, sym, ind, R_AARCH64_LD64_GOT_LO12_NC, 0);
-    o(ARM64_LDR_X | ARM64_RN(r) | r); // ld xr,[xr, #sym]
-#endif
     if (addend) {
         // add xr, xr, #addend
 	if (addend & 0xffful)
@@ -504,17 +492,35 @@ static void arm64_sym(int r, Sym *sym, unsigned long addend)
     }
 }
 
+static void arm64_sym(int r, Sym *sym, unsigned long addend)
+{
+#ifdef TCC_TARGET_PE
+    /* PE links symbol addresses directly; there is no ELF-style GOT here. */
+    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_PREL_PG_HI21, 0);
+    o(ARM64_ADRP | r);            // adrp xr, #sym
+    greloca(cur_text_section, sym, ind, R_AARCH64_ADD_ABS_LO12_NC, 0);
+    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | r); // add xr, xr, #sym
+#else
+    greloca(cur_text_section, sym, ind, R_AARCH64_ADR_GOT_PAGE, 0);
+    o(ARM64_ADRP | r);            // adrp xr, #sym
+    greloca(cur_text_section, sym, ind, R_AARCH64_LD64_GOT_LO12_NC, 0);
+    o(ARM64_LDR_X | ARM64_RN(r) | r); // ld xr,[xr, #sym]
+#endif
+    arm64_add_offset(r, addend);
+}
+
 static void arm64_load_cmp(int r, SValue *sv);
 
-static void arm64_tls_x30(SValue *sv)
+static void arm64_tls_sym(int r, Sym *sym, unsigned long addend)
 {
-    o(0xd53bd05e); /* mrs x30, tpidr_el0 */
+    o(0xd53bd040 | r); /* mrs xr, tpidr_el0 */
-    greloca(cur_text_section, sv->sym, ind, R_AARCH64_TLSLE_ADD_TPREL_HI12, 0);
+    greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_HI12, 0);
-    /* add x30, x30, #0, lsl #12 */
+    /* add xr, xr, #0, lsl #12 */
-    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) | ARM64_RN(30) | ARM64_RD(30));
+    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) | ARM64_RN(r) | ARM64_RD(r));
-    greloca(cur_text_section, sv->sym, ind, R_AARCH64_TLSLE_ADD_TPREL_LO12, 0);
+    greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_LO12, 0);
-    /* add x30, x30, #0 */
+    /* add xr, xr, #0 */
-    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(30) | ARM64_RD(30));
+    o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | ARM64_RD(r));
+    arm64_add_offset(r, addend);
 }
 
 ST_FUNC void load(int r, SValue *sv)
@@ -537,7 +543,9 @@ ST_FUNC void load(int r, SValue *sv)
     if (svr == (VT_CONST | VT_LVAL)) {
 	uint64_t i = sv->c.i;
 
-	if (sv->sym)
+	if (sv->sym && (sv->sym->type.t & VT_TLS))
+            arm64_tls_sym(30, sv->sym, 0);
+        else if (sv->sym)
             arm64_sym(30, sv->sym, // use x30 for address
 	              arm64_check_offset(0, arm64_type_size(svtt), i));
 	else
@@ -564,7 +572,7 @@ ST_FUNC void load(int r, SValue *sv)
 
     if (svr == (VT_CONST | VT_LVAL | VT_SYM)) {
         if (sv->sym->type.t & VT_TLS)
-            arm64_tls_x30(sv);
+            arm64_tls_sym(30, sv->sym, 0);
         else
         arm64_sym(30, sv->sym, // use x30 for address
 		  arm64_check_offset(0, arm64_type_size(svtt), svcoff));
@@ -578,6 +586,9 @@ ST_FUNC void load(int r, SValue *sv)
     }
 
     if (svr == (VT_CONST | VT_SYM)) {
+        if (sv->sym->type.t & VT_TLS)
+            arm64_tls_sym(intr(r), sv->sym, svcul);
+        else
             arm64_sym(intr(r), sv->sym, svcul);
         return;
     }
@@ -659,7 +670,7 @@ ST_FUNC void store(int r, SValue *sv)
     if (svr == (VT_CONST | VT_LVAL)) {
 	uint64_t i = sv->c.i;
 	if (sv->sym && (sv->sym->type.t & VT_TLS))
-            arm64_tls_x30(sv);
+            arm64_tls_sym(30, sv->sym, 0);
         else
 	if (sv->sym)
             arm64_sym(30, sv->sym, // use x30 for address
@@ -686,15 +697,7 @@ ST_FUNC void store(int r, SValue *sv)
 
     if (svr == (VT_CONST | VT_LVAL | VT_SYM)) {
         if (sv->sym->type.t & VT_TLS) {
-            o(0xd53bd05e); /* mrs x30, tpidr_el0 */
+            arm64_tls_sym(30, sv->sym, 0);
-            greloca(cur_text_section, sv->sym, ind,
-                    R_AARCH64_TLSLE_ADD_TPREL_HI12, 0);
-            o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) |
-              ARM64_RN(30) | ARM64_RD(30)); /* add x30, x30, #0, lsl #12 */
-            greloca(cur_text_section, sv->sym, ind,
-                    R_AARCH64_TLSLE_ADD_TPREL_LO12, 0);
-            o(ARM64_ADD_IMM | ARM64_SF(1) |
-              ARM64_RN(30) | ARM64_RD(30)); /* add x30, x30, #0 */
             if (IS_FREG(r))
                 arm64_strv(arm64_type_size(svtt), fltr(r), 30, svcoff);
             else

i386-gen.c

@@ -354,7 +354,7 @@ ST_FUNC void load(int r, SValue *sv)
     } else {
         if ((fr & VT_SYM) && (sv->sym->type.t & VT_TLS)) {
             oad(0x058b65 | REG_VALUE(r) << 19, 0); /* mov gs:0,r */
-            oad(0xC081 | REG_VALUE(r) << 8, 0); /* add tpoffs,r */
+            oad(0xC081 | REG_VALUE(r) << 8, fc); /* add tpoffs,r */
             greloc(cur_text_section, sv->sym, ind - 4, R_386_TLS_LE);
 #if defined CONFIG_TCC_PIC
         } else if ((fr & (VT_VALMASK|VT_SYM)) == (VT_CONST|VT_SYM)) {

riscv64-gen.c

@@ -186,10 +186,8 @@ static int load_symofs(int r, SValue *sv, int forstore, int *new_fc)
                     R_RISCV_TPREL_HI20, sv->c.i);
             o(0x37 | (rr << 7));  // lui RR, 0 %tprel_hi(sym)
             greloca(cur_text_section, sv->sym, ind,
-                    R_RISCV_TPREL_LO12_I, 0);
+                    R_RISCV_TPREL_LO12_I, sv->c.i);
             EI(0x13, 0, rr, rr, 0); // addi RR, RR, 0 %tprel_lo(sym)
-	    if (fc)
-                EI(0x13, 0, rr, rr, fc); // addi RR, RR, fc
             ER(0x33, 0, rr, rr, 4, 0); // add RR, RR, tp
             *new_fc = 0;
             return rr;

tccelf.c

@@ -205,7 +205,9 @@ ST_FUNC void tccelf_end_file(TCCState *s1)
 #ifndef TCC_TARGET_PE
             if (sym_bind == STB_GLOBAL && s1->output_type == TCC_OUTPUT_OBJ) {
                 /* undefined symbols with STT_FUNC are confusing gnu ld when
-                   linking statically to STT_GNU_IFUNC */
+                   linking statically to STT_GNU_IFUNC.  Keep TLS typed,
+                   otherwise gnu ld rejects non-TLS refs to TLS definitions. */
+                if (sym_type != STT_TLS)
                     sym_type = STT_NOTYPE;
             }
 #endif

tests/tests2/144_tls.c

@@ -6,6 +6,11 @@
 typedef enum { tls_a, tls_b, tls_c } tls_enum_type;
 typedef struct { int tls_d, tls_e, tls_f; } tls_struct_type;
 
+typedef struct {
+    int *init;
+    signed short *element;
+} tls_address_type;
+
 __thread char tls_char = 42;
 __thread short tls_short = 43;
 __thread int tls_init = 44;
@@ -17,6 +22,41 @@ __thread long double tls_long_double = 48.0;
 __thread int *tls_ptr = (int *)49;
 __thread tls_enum_type tls_enum = tls_b;
 __thread tls_struct_type tls_struct = { 50, 51, 52 };
+__thread signed char tls_signed_char = -5;
+__thread unsigned char tls_unsigned_char = 250;
+__thread signed short tls_signed_short = -1234;
+__thread unsigned short tls_unsigned_short = 60000;
+__thread signed short tls_array[4] = { -1000, 2000, -3000, 4000 };
+
+#if defined __riscv && __riscv_xlen == 64
+typedef struct {
+    char padding[4096];
+    int value;
+} tls_large_type;
+
+__thread tls_large_type tls_large = { { 0 }, 123456 };
+#endif
+
+static int check_extra(int changed)
+{
+    volatile int index = 2;
+
+    if (changed)
+        return tls_signed_char != -7 || tls_unsigned_char != 42
+            || tls_signed_short != -2222 || tls_unsigned_short != 12345
+            || tls_array[1] != 2222 || tls_array[index] != -3333;
+
+    if (tls_signed_char != -5 || tls_unsigned_char != 250
+        || tls_signed_short != -1234 || tls_unsigned_short != 60000
+        || tls_array[1] != 2000 || tls_array[index] != -3000)
+        return 1;
+
+#if defined __riscv && __riscv_xlen == 64
+    if (tls_large.value != 123456)
+        return 1;
+#endif
+    return 0;
+}
 
 static int check(void)
 {
@@ -40,9 +80,18 @@ static int check(void)
 
 static void *thread_func(void *arg)
 {
-    (void)arg;
+    tls_address_type *main_addresses = arg;
+    int *init_address = &tls_init;
+    signed short *element_address = &tls_array[1];
+    volatile int index = 2;
     long errors = check();
 
+    if (init_address == main_addresses->init
+        || element_address == main_addresses->element
+        || *init_address != 44 || *element_address != 2000
+        || tls_array[index] != -3000 || check_extra(0))
+        errors = 1;
+
     tls_char = 10;
     tls_short = 20;
     tls_init = 30;
@@ -56,6 +105,12 @@ static void *thread_func(void *arg)
     tls_struct.tls_d = 100;
     tls_struct.tls_e = 110;
     tls_struct.tls_f = 120;
+    *element_address = 2222;
+    tls_array[index] = -3333;
+    tls_signed_char = -7;
+    tls_unsigned_char = 42;
+    tls_signed_short = -2222;
+    tls_unsigned_short = 12345;
 
     CHECK(tls_char, "%d", 10);
     CHECK(tls_short, "%d", 20);
@@ -71,6 +126,10 @@ static void *thread_func(void *arg)
     CHECK(tls_struct.tls_e, "%d", 110);
     CHECK(tls_struct.tls_f, "%d", 120);
 
+    if (*init_address != 30 || *element_address != 2222
+        || tls_array[index] != -3333 || check_extra(1))
+        errors = 1;
+
     return (void *)errors;
 }
 
@@ -79,9 +138,20 @@ int main()
     pthread_t t;
     void *ret;
     int errors = check();
+    tls_address_type addresses = { &tls_init, &tls_array[1] };
 
-    pthread_create(&t, NULL, thread_func, NULL);
+    if (*addresses.init != 44 || *addresses.element != 2000
+        || check_extra(0))
+        errors = 1;
+
+    pthread_create(&t, NULL, thread_func, &addresses);
     pthread_join(t, &ret);
 
-    return (ret ? 1 : 0) | errors | check();
+    errors |= (ret ? 1 : 0) | check();
+    if (*addresses.init != 44 || *addresses.element != 2000
+        || check_extra(0))
+        errors = 1;
+
+    printf("errors: %d\n", errors);
+    return errors;
 }

tests/tests2/144_tls.expect

@@ -50,3 +50,4 @@
 50
 51
 52
+errors: 0

tests/tests2/146+_tls_extern.c

@@ -0,0 +1,2 @@
+__thread signed short extern_tls_short = 1234;
+__thread float extern_tls_float = 1.5f;

tests/tests2/146_tls_extern.c

@@ -0,0 +1,22 @@
+#include <stdio.h>
+
+extern __thread signed short extern_tls_short;
+extern __thread float extern_tls_float;
+
+int main(void)
+{
+    int errors = 0;
+
+    if (extern_tls_short != 1234 || extern_tls_float != 1.5f)
+        errors = 1;
+
+    extern_tls_short = -2222;
+    extern_tls_float = 2.5f;
+    extern_tls_float += 1.0f;
+
+    if (extern_tls_short != -2222 || extern_tls_float != 3.5f)
+        errors = 1;
+
+    printf("extern tls errors: %d\n", errors);
+    return errors;
+}

tests/tests2/146_tls_extern.expect

@@ -0,0 +1 @@
+extern tls errors: 0

tests/tests2/Makefile

@@ -15,6 +15,9 @@ endif
 ifeq (,$(filter i386,$(ARCH)))
  SKIP += 98_al_ax_extend.test 99_fastcall.test
 endif
+ifeq (,$(filter x86_64 riscv64 arm64 i386,$(ARCH)))
+ SKIP += 146_tls_extern.test # TLS only implemented on these architectures
+endif
 ifeq (,$(filter i386 x86_64,$(ARCH)))
  SKIP += 85_asm-outside-function.test # x86 asm
  SKIP += 127_asm_goto.test    # hardcodes x86 asm
@@ -39,6 +42,9 @@ endif
 ifeq (-$(findstring gcc,$(CC))-,--)
  SKIP += $(patsubst %.expect,%.test,$(GEN-ALWAYS))
 endif
+ifeq ($(CC_NAME),tcc)
+ SKIP += 146_tls_extern.test # needs an external system linker
+endif
 ifeq (-$(CONFIG_WIN32)-$(CONFIG_i386)$(CONFIG_arm)-,--yes-)
  SKIP += 95_bitfields%.test # type_align is different on 32bit-non-windows
 endif
@@ -47,19 +53,22 @@ ifeq (-$(CONFIG_WIN32)-,-yes-)
  SKIP += 114_bound_signal.test # No pthread support
  SKIP += 117_builtins.test # win32 port doesn't define __builtins
  SKIP += 124_atomic_counter.test # No pthread support
- SKIP += 144_tls.test # No pthread support
+ SKIP += 144_tls.test 146_tls_extern.test # TLS runtime not supported
 endif
 ifneq (,$(filter OpenBSD FreeBSD NetBSD,$(TARGETOS)))
  SKIP += 106_versym.test # no pthread_condattr_setpshared
  SKIP += 114_bound_signal.test # libc problem signal/fork
  SKIP += 116_bound_setjmp2.test # No TLS_FUNC/TLS_VAR in bcheck.c
- SKIP += 144_tls.test # No tls support
+ SKIP += 144_tls.test 146_tls_extern.test # No tls support
+endif
+ifeq ($(TARGETOS),DragonFly)
+ SKIP += 144_tls.test 146_tls_extern.test # TLS runtime not supported
 endif
 ifeq ($(CONFIG_OSX),yes)
- SKIP += 144_tls.test # TLS runtime not supported on Mach-O
+ SKIP += 144_tls.test 146_tls_extern.test # TLS runtime not supported on Mach-O
 endif
 ifeq (-$(ARCH)-,-arm-)
- SKIP += 144_tls.test # doesn't work currently
+ SKIP += 144_tls.test 146_tls_extern.test # doesn't work currently
 endif
 ifneq (-$(ARCH)-,-arm64-)
  SKIP += 138_arm64_encoding.test
@@ -122,6 +131,15 @@ GEN-ALWAYS =
 144_tls.test: FLAGS += -pthread
 144_tls.test: NORUN = true
 
+# Verify that undefined TLS symbols in TCC objects work with the system linker.
+146_tls_extern.test: NORUN = true
+146_tls_extern.test: T1 = \
+	$(TCC) -c $1 -o 146_tls_extern-main.o && \
+	$(TCC) -c $(subst 146,146+,$1) -o 146_tls_extern-def.o && \
+	$(CC) -no-pie -Wl,-z,noexecstack 146_tls_extern-main.o \
+		146_tls_extern-def.o -o $(basename $@).exe && \
+	./$(basename $@).exe
+
 112_backtrace.test: FLAGS += -dt -b
 112_backtrace.test 113_btdll.test 126_bound_global.test: FILTER += \
     -e 's;[0-9A-Fa-fx]\{5,\};........;g' \
@@ -204,4 +222,4 @@ $(sort $(GEN-ALWAYS) $(UPDATE)) : force
 force:
 
 clean :
-	rm -f fred.txt *.output *.exe *.dll *.so *.def $(GEN-ALWAYS)
+	rm -f fred.txt *.output *.o *.exe *.dll *.so *.def $(GEN-ALWAYS)