feat(arm64): add TLS Local Exec code generation and linker relocations

arm64-gen.c

@@ -552,6 +552,23 @@ ST_FUNC void load(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 */
+            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_ldrv(arm64_type_size(svtt), fltr(r), 30, svcoff);
+            else
+                arm64_ldrx(!(svtt&VT_UNSIGNED), arm64_type_size(svtt),
+                           intr(r), 30, svcoff);
+            return;
+        }
         arm64_sym(30, sv->sym, // use x30 for address
 		  arm64_check_offset(0, arm64_type_size(svtt), svcoff));
         if (IS_FREG(r))
@@ -645,6 +662,22 @@ 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)) {
+            o(0xd53bd05e);
+            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));
+            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));
+            if (IS_FREG(r))
+                arm64_strv(arm64_type_size(svtt), fltr(r), 30, i);
+            else
+                arm64_strx(arm64_type_size(svtt), intr(r), 30, i);
+            return;
+        }
 	if (sv->sym)
             arm64_sym(30, sv->sym, // use x30 for address
 		      arm64_check_offset(0, arm64_type_size(svtt), i));
@@ -668,6 +701,22 @@ 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 */
+            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
+                arm64_strx(arm64_type_size(svtt), intr(r), 30, svcoff);
+            return;
+        }
         arm64_sym(30, sv->sym, // use x30 for address
 		  arm64_check_offset(0, arm64_type_size(svtt), svcoff));
         if (IS_FREG(r))

arm64-link.c

@@ -43,6 +43,8 @@ ST_FUNC int code_reloc (int reloc_type)
         case R_AARCH64_LDST32_ABS_LO12_NC:
         case R_AARCH64_LDST16_ABS_LO12_NC:
         case R_AARCH64_LDST8_ABS_LO12_NC:
+        case R_AARCH64_TLSLE_ADD_TPREL_HI12:
+        case R_AARCH64_TLSLE_ADD_TPREL_LO12:
         case R_AARCH64_GLOB_DAT:
         case R_AARCH64_COPY:
             return 0;
@@ -80,6 +82,8 @@ ST_FUNC int gotplt_entry_type (int reloc_type)
         case R_AARCH64_COPY:
         case R_AARCH64_CONDBR19:
         case R_AARCH64_TSTBR14:
+        case R_AARCH64_TLSLE_ADD_TPREL_HI12:
+        case R_AARCH64_TLSLE_ADD_TPREL_LO12:
             return NO_GOTPLT_ENTRY;
 
         case R_AARCH64_ABS32:
@@ -365,6 +369,27 @@ ST_FUNC void relocate(TCCState *s1, ElfW_Rel *rel, int type, unsigned char *ptr,
 #endif
             write64le(ptr, val - rel->r_addend);
             return;
+        case R_AARCH64_TLSLE_ADD_TPREL_HI12:
+        case R_AARCH64_TLSLE_ADD_TPREL_LO12: {
+            addr_t tls_start = 0;
+            int i;
+            for (i = 1; i < s1->nb_sections; i++) {
+                Section *s = s1->sections[i];
+                if (s->sh_flags & SHF_TLS && s->sh_size) {
+                    if (!tls_start || s->sh_addr < tls_start)
+                        tls_start = s->sh_addr;
+                }
+            }
+            /* glibc arm64: tp points to tcbhead_t (DTV), TLS data starts after it */
+            int64_t tp_offset = val - tls_start + 16;
+            int64_t imm;
+            if (type == R_AARCH64_TLSLE_ADD_TPREL_HI12)
+                imm = (tp_offset >> 12) & 0xfff;
+            else
+                imm = tp_offset & 0xfff;
+            write32le(ptr, ((read32le(ptr) & 0xffc003ff) | (imm << 10)));
+            return;
+        }
         case R_AARCH64_RELATIVE:
 #ifdef TCC_TARGET_PE
             add32le(ptr, val - s1->pe_imagebase);

tests/tests2/Makefile

@@ -19,7 +19,7 @@ ifeq (,$(filter i386 x86_64,$(ARCH)))
  SKIP += 85_asm-outside-function.test # x86 asm
  SKIP += 127_asm_goto.test    # hardcodes x86 asm
 endif
-ifeq (,$(filter x86_64 riscv64,$(ARCH)))
+ifeq (,$(filter x86_64 riscv64 arm64,$(ARCH)))
  SKIP += 144_tls.test # TLS only implemented on x86_64 so far
 endif
 ifeq ($(CONFIG_backtrace),no)