TLS bits

tccelf.c:
- set_elf_sym()
    check NOBITS instead of just .bss
- sort_sections()
    must have common alignment for tls sections
    fix BSD relro

tccpe.c:
- sort .tbss after .tdata

tccrun.c:
- TLS not supported with -run
- arm64: FlushInstructionCache on win32 too

lib/bt-exe.c:
- use less unused bss memory (from global vars in tcc.h)

arm64-gen.c:
- arm64_sym():
  direct access (no GOT) for 'static' vars
  relocate with 'addend' unless GOT
- reserve X16/x17 for internal usage (specifically to avoid
  their usage as value register passed to store())

arm64-gen.c

@@ -77,14 +77,10 @@ ST_DATA const int reg_classes[NB_REGS] = {
   RC_INT | RC_R(13),
   RC_INT | RC_R(14),
   RC_INT | RC_R(15),
-  RC_INT | RC_R(16),
-  RC_INT | RC_R(17),
-#ifdef TCC_TARGET_PE
-  RC_R(18), /* (x18 reserved on Windows) */
-#else
-  RC_INT | RC_R(18),
-#endif
-  RC_R30, // not in RC_INT as we make special use of x30
+  RC_R(16), /* reserved for generator (arm64_sym()) */
+  RC_R(17), /* reserved for generator (arm64_sym()) */
+  RC_R(18), /* reserved on Windows/OSX */
+  RC_R30,   /* not in RC_INT as we make special use of x30 */
   RC_FLOAT | RC_F(0),
   RC_FLOAT | RC_F(1),
   RC_FLOAT | RC_F(2),
@@ -298,6 +294,7 @@ static void arm64_spoff(int reg, uint64_t off)
 
 static uint64_t arm64_check_offset(int sz_, uint64_t off)
 {
+#ifndef TCC_TARGET_PE
     uint32_t sz = sz_;
     uint64_t scaled_mask = 0xffful << sz;
     if (!(off & ~scaled_mask) || (off < 256 || -off <= 256))
@@ -306,6 +303,7 @@ static uint64_t arm64_check_offset(int sz_, uint64_t off)
         return scaled_mask;
     if (off & 0x1fful)
         return 0x1fful;
+#endif
     return 0;
 }
 
@@ -463,12 +461,17 @@ static void arm64_strv(int sz_, int dst, int bas, uint64_t off)
     }
 }
 
+static void arm64_reloca(Sym *sym, int reloc_type, uint64_t addend)
+{
+    greloca(cur_text_section, sym, ind, reloc_type, addend);
+}
+
 static void arm64_sym(int r, Sym *sym, addr_t addend)
 {
     if (sym->type.t & VT_TLS) {
-#if TCC_TARGET_PE
+#ifdef TCC_TARGET_PE
         Sym *s2 = external_global_sym(TOK___tls_index, &int_type);
-        int r2 = get_reg(RC_INT);
+        int r2 = get_reg(RC_R(16));
         arm64_sym(30, s2, 0);
         o(0xb94003de); // ldr w30, [x30]
         o(0xf9402e40 | r2); // ldr r2, [x18, #88]
@@ -477,46 +480,35 @@ static void arm64_sym(int r, Sym *sym, addr_t addend)
 #else
         o(0xd53bd040 | r); /* mrs xr, tpidr_el0 */
 #endif
-        greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_HI12, 0);
-        /* add xr, xr, #0, lsl #12 */
-        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) | ARM64_RN(r) | ARM64_RD(r));
-        greloca(cur_text_section, sym, ind, R_AARCH64_TLSLE_ADD_TPREL_LO12, 0);
-        /* add xr, xr, #0 */
-        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | ARM64_RD(r));
-        goto add_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
-add_addend:
-    if (addend) {
-        // add xr, xr, #addend
-	if (addend & 0xffful)
-           o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | r |
-             (addend & 0xfff) << 10);
-        if (addend > 0xffful) {
-            // add xr, xr, #addend, lsl #12
-	    if (addend & 0xfff000ul)
-                o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) |
-                  ARM64_RN(r) | r | ((addend >> 12) & 0xfff) << 10);
-            if (addend > 0xfffffful) {
-		/* very unlikely */
-		int t = r ? 0 : 1;
-		o(ARM64_STR_X_PRE | 0x001F0FE0U | t); /* str xt, [sp, #-16]! */
-		arm64_movimm(t, addend & ~0xffffffull); // use xt for addent
+        arm64_reloca(sym, R_AARCH64_TLSLE_ADD_TPREL_HI12, addend);
+        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1) | ARM64_RN(r) | ARM64_RD(r)); /* add xr, xr, #0, lsl #12 */
+        arm64_reloca(sym, R_AARCH64_TLSLE_ADD_TPREL_LO12, addend);
+        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | ARM64_RD(r)); /* add xr, xr, #0 */
+#ifndef TCC_TARGET_PE /* no elf-style GOT on Windows */
+    } else if (!(sym->type.t & VT_STATIC)) {
+        arm64_reloca(sym, R_AARCH64_ADR_GOT_PAGE, 0);
+        o(ARM64_ADRP | r); /* adrp xr, #sym */
+        arm64_reloca(sym, R_AARCH64_LD64_GOT_LO12_NC, 0);
+        o(ARM64_LDR_X | ARM64_RN(r) | r); /* ld xr,[xr, #sym] */
+        if (addend > 0xffffff) {
+            int t = get_reg(RC_R(16));
+            arm64_movimm(t, addend); // use xt for addent
             o(ARM64_ADD_REG | ARM64_SF(1) | ARM64_RM(t) | ARM64_RN(r) | r); /* add xr, xr, xt */
-		o(ARM64_LDR_X_POST | 0x000107E0U | t); /* ldr xt, [sp], #16 */
-	    }
+        } else {
+           if (addend & 0xfff) /* add xr, xr, #addend */
+                o(ARM64_ADD_IMM | ARM64_SF(1)
+                    | ARM64_RN(r) | r | (addend & 0xfff) << 10);
+            addend >>= 12;
+            if (addend & 0xfff) /* add xr, xr, #addend, lsl #12 */
+                o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_SH(1)
+                    | ARM64_RN(r) | r | (addend & 0xfff) << 10);
         }
+#endif
+    } else {
+        arm64_reloca(sym, R_AARCH64_ADR_PREL_PG_HI21, addend);
+        o(ARM64_ADRP | r); /* adrp xr, #sym */
+        arm64_reloca(sym, R_AARCH64_ADD_ABS_LO12_NC, addend);
+        o(ARM64_ADD_IMM | ARM64_SF(1) | ARM64_RN(r) | r); /* add xr, xr, #sym */
     }
 }
 
@@ -678,8 +670,8 @@ ST_FUNC void store(int r, SValue *sv)
 
 static void arm64_gen_bl_or_b(int b)
 {
-    if ((vtop->r & (VT_VALMASK | VT_LVAL)) == VT_CONST && (vtop->r & VT_SYM)) {
-	greloca(cur_text_section, vtop->sym, ind,
+    if ((vtop->r & (VT_VALMASK | VT_LVAL | VT_SYM)) == (VT_CONST | VT_SYM)) {
+	arm64_reloca(vtop->sym,
                 b ? R_AARCH64_JUMP26 :  R_AARCH64_CALL26, 0);
 	o(b ? ARM64_B : ARM64_BL); // b/bl .
     }
@@ -691,16 +683,16 @@ static void arm64_gen_bl_or_b(int b)
     }
 }
 
-#if defined(CONFIG_TCC_BCHECK)
-
-static void gen_bounds_call(int v)
+#if defined CONFIG_TCC_BCHECK || defined TCC_TARGET_PE
+static void gen_static_call(int v)
 {
     Sym *sym = external_helper_sym(v);
-
-    greloca(cur_text_section, sym, ind, R_AARCH64_CALL26, 0);
+    arm64_reloca(sym, R_AARCH64_CALL26, 0);
     o(ARM64_BL); // bl
 }
+#endif
 
+#if defined(CONFIG_TCC_BCHECK)
 static void gen_bounds_prolog(void)
 {
     /* leave some room for bound checking code */
@@ -735,7 +727,7 @@ static void gen_bounds_epilog(void)
         saved_ind = ind;
         ind = func_bound_ind;
         arm64_sym(0, sym_data, 0);
-        gen_bounds_call(TOK___bound_local_new);
+        gen_static_call(TOK___bound_local_new);
         ind = saved_ind;
     }
 
@@ -743,7 +735,7 @@ static void gen_bounds_epilog(void)
     o(0xa9bf07e0); /* stp x0, x1, [sp, #-16]! */
     o(0x3c9f0fe0); /* str q0, [sp, #-16]! */
     arm64_sym(0, sym_data, 0);
-    gen_bounds_call(TOK___bound_local_delete);
+    gen_static_call(TOK___bound_local_delete);
     o(0x3cc107e0); /* ldr q0, [sp], #16 */
     o(0xa8c107e0); /* ldp x0, x1, [sp], #16 */
 }
@@ -1004,11 +996,8 @@ static void arm64_sub_sp(uint64_t diff)
         return;
 #ifdef TCC_TARGET_PE
     if (diff >= 4096) {
-        Sym *sym = external_helper_sym(TOK___chkstk);
-
         arm64_movimm(15, diff >> 4);
-        greloca(cur_text_section, sym, ind, R_AARCH64_CALL26, 0);
-        o(ARM64_BL); // bl __chkstk
+        gen_static_call(TOK___chkstk);
         o(0xcb2f73ff); // sub sp,sp,x15,lsl #4
         return;
     }

lib/bt-exe.c

@@ -3,7 +3,7 @@
    from tccrun.c into executables. */
 
 #define CONFIG_TCC_BACKTRACE_ONLY
-#define ONE_SOURCE 1
+#define ONE_SOURCE 0
 #define pstrcpy tcc_pstrcpy
 #include "../tccrun.c"
 

tccelf.c

@@ -693,6 +693,14 @@ version_add (TCCState *s1)
 }
 #endif /* ndef ELF_OBJ_ONLY */
 
+/* catch .tbss also */
+static int IS_BSS(TCCState *s1, int ndx)
+{
+    return ndx == SHN_COMMON
+     || (ndx < s1->nb_sections
+         && s1->sections[ndx]->sh_type == SHT_NOBITS);
+}
+
 /* add an elf symbol : check if it is already defined and patch
    it. Return symbol index. NOTE that sh_num can be SHN_UNDEF. */
 ST_FUNC int set_elf_sym(Section *s, addr_t value, unsigned long size,
@@ -744,13 +752,10 @@ ST_FUNC int set_elf_sym(Section *s, addr_t value, unsigned long size,
                 /* ignore hidden symbols after */
             } else if (s->sh_flags & SHF_DYNSYM) {
                 /* we accept that two DLL define the same symbol */
-            } else if ((esym->st_shndx == SHN_COMMON
-                            || esym->st_shndx == bss_section->sh_num)
-                        && (shndx < SHN_LORESERVE
-                            && shndx != bss_section->sh_num)) {
+            } else if (!IS_BSS(s1, shndx) && IS_BSS(s1, esym->st_shndx)) {
                 /* data symbol gets precedence over common/bss */
                 goto do_patch;
-            } else if (shndx == SHN_COMMON || shndx == bss_section->sh_num) {
+            } else if (IS_BSS(s1, shndx)) {
                 /* data symbol keeps precedence over common/bss */
 	    } else if (esym->st_other & ST_ASM_SET) {
 		/* If the existing symbol came from an asm .set
@@ -2259,6 +2264,7 @@ static int sort_sections(TCCState *s1, int *sec_order, struct dyn_inf *d)
     int i, j, k, f, f0, n;
     int nb_sections = s1->nb_sections;
     int *sec_cls = sec_order + nb_sections;
+    int tls_align = 0; /* need common alignment for all tls sections */
 
     for (i = 1; i < nb_sections; i++) {
         s = s1->sections[i];
@@ -2294,6 +2300,8 @@ static int sort_sections(TCCState *s1, int *sec_order, struct dyn_inf *d)
             k = 0x60;
         /* RELRO sections --> */
         } else if (s->sh_flags & SHF_TLS) {
+            if (s->sh_addralign > tls_align)
+                tls_align = s->sh_addralign;
             k = 0x40 + (s->sh_type == SHT_NOBITS);
         } else if (s->sh_type == SHT_PREINIT_ARRAY) {
             k = 0x42;
@@ -2302,11 +2310,11 @@ static int sort_sections(TCCState *s1, int *sec_order, struct dyn_inf *d)
         } else if (s->sh_type == SHT_FINI_ARRAY) {
             k = 0x44;
         } else if (s->sh_type == SHT_DYNAMIC) {
-            k = 0x46;
+            k = 0x48;
         } else if (s == s1->got) {
-            k = 0x47; /* .got as RELRO needs BIND_NOW in DT_FLAGS */
+            k = 0x49; /* .got as RELRO needs BIND_NOW in DT_FLAGS */
         } else if (s->reloc && (s->reloc->sh_flags & SHF_ALLOC) && j == 0x100) {
-            k = 0x45;
+            k = 0x45; /* +1 for data.ro */
         /* <-- */
         } else if (s->sh_type == SHT_NOTE) {
             k = 0x08, d->notes = 1;
@@ -2349,15 +2357,18 @@ static int sort_sections(TCCState *s1, int *sec_order, struct dyn_inf *d)
 	    /* NetBSD only supports 2 PT_LOAD sections.
 	       See: https://blog.netbsd.org/tnf/entry/the_first_report_on_lld */
 	    if ((f & SHF_WRITE) == 0)
-                f |= SHF_EXECINSTR, k = 0; /* no relro */
+                f |= SHF_EXECINSTR;
+            k = 0; /* no relro */
 #endif
+            if ((k & 0xfff0) == 0x240) /* RELRO sections */
+                d->relro = 1, f |= SHFX_RELRO;
             /* start new header when flags changed, but avoid zero memsz */
             if (f != f0 && s->sh_size)
                 f0 = f, ++n, f |= SHFX_NEWPH;
-            if ((s->sh_flags & SHF_TLS) && s->sh_size)
+            if ((s->sh_flags & SHF_TLS) && s->sh_size) {
+                s->sh_addralign = tls_align;
                 d->tls = 1, f |= SHF_TLS;
-            if ((k & 0xfff0) == 0x240) /* RELRO sections */
-                d->relro = 1, f |= SHFX_RELRO;
+            }
         }
         sec_cls[i] = f;
         //printf("ph %d sec %02d : %3X %3X  %8.2X  %04X  %s\n", (f>0) * n, i, f, k, s->sh_type, (int)s->sh_size, s->name);
@@ -2484,7 +2495,7 @@ static int layout_sections(TCCState *s1, int *sec_order, struct dyn_inf *d)
         s->sh_offset = file_offset;
         s->sh_addr = addr;
 
-        //printf("%d : %08x %08x %04x %03x %s\n", (int)(ph - d->phdr), (int)file_offset, (int)addr, (int)s->sh_size, s->sh_type, s->name);
+        //printf("%d : %08x %08x %04x %03x %s %d\n", n, (int)file_offset, (int)addr, (int)s->sh_size, s->sh_type, s->name, align + 1);
         addr += s->sh_size;
         if (s->sh_type != SHT_NOBITS)
             file_offset += s->sh_size;
@@ -2516,8 +2527,6 @@ static int layout_sections(TCCState *s1, int *sec_order, struct dyn_inf *d)
         }
         if (f & SHF_TLS) {
             ph2 = update_phdr(&d->phdr[d->tls], PT_TLS, s, addr, file_offset);
-            if (s->sh_addralign > ph2->p_align)
-                ph2->p_align = s->sh_addralign;
             if (s->sh_type == SHT_NOBITS)
                 addr -= s->sh_size;
             /* for xxx-link.c:relocate() */

tccpe.c

@@ -351,7 +351,8 @@ enum {
     sec_bss ,
     sec_idata ,
     sec_pdata ,
-    sec_tls ,
+    sec_tdata ,
+    sec_tbss ,
     sec_other ,
     sec_rsrc ,
     sec_debug ,
@@ -1197,6 +1198,7 @@ static void pe_build_tls(struct pe_info *pe, Section *s)
     int c, n;
 
     if (0 == s) {
+        pe->tls_size = sizeof (IMAGE_TLS_DIRECTORY);
         pe->tls_dir = section_add(pe->thunk, pe->tls_size, 16);
         pe->tls_data = section_add(data_section, PTR_SIZE * (1+3), 16);
         /* put relocations on entries */
@@ -1209,6 +1211,7 @@ static void pe_build_tls(struct pe_info *pe, Section *s)
                 0, data_section->sh_num, "__tls_index");
         return;
     }
+
     if (0 == s1->tls_start)
         s1->tls_start = s->sh_addr;
     d = (void*)(pe->thunk->data + pe->tls_dir);
@@ -1235,7 +1238,7 @@ static int pe_section_class(Section *s)
         return sec_debug;
     } else if (flags & SHF_ALLOC) {
         if (flags & SHF_TLS)
-            return sec_tls;
+            return sec_tdata + (type == SHT_NOBITS);
         if (type == SHT_PROGBITS
          || type == SHT_INIT_ARRAY
          || type == SHT_FINI_ARRAY) {
@@ -1273,6 +1276,7 @@ static int pe_assign_addresses (struct pe_info *pe)
     if (PE_DLL == pe->type
         || (s1->pe_dll_characteristics & IMAGE_DLLCHARACTERISTICS_DYNAMIC_BASE))
         pe->reloc = new_section(s1, ".reloc", SHT_PROGBITS, 0);
+    pe->thunk = rodata_section;
     //pe->thunk = new_section(s1, ".iedat", SHT_PROGBITS, SHF_ALLOC);
 
     nbs = s1->nb_sections;
@@ -1284,35 +1288,30 @@ static int pe_assign_addresses (struct pe_info *pe)
         for (n = i; n > 1 && k < (c = sec_cls[n - 1]); --n)
             sec_cls[n] = c, sec_order[n] = sec_order[n - 1];
         sec_cls[n] = k, sec_order[n] = i;
-        if (k == sec_tls)
-            pe->tls_size = sizeof (IMAGE_TLS_DIRECTORY);
+        if ((s->sh_flags & SHF_TLS) && 0 == pe->tls_size)
+            pe_build_tls(pe, NULL);
     }
     si = NULL;
     addr = pe->imagebase + 1;
-
     for (i = 1; (c = sec_cls[i]) < sec_last; ++i) {
         s = s1->sections[sec_order[i]];
 
         if (PE_MERGE_DATA && c == sec_bss)
             c = sec_data;
+        if (c == sec_tbss)
+            c = sec_tdata;
 
         if (si && c == si->cls && c != sec_debug) {
             /* merge with previous section */
-            s->sh_addr = addr = ((addr - 1) | (16 - 1)) + 1;
+            s->sh_addr = addr = ((addr - 1) | (s->sh_addralign - 1)) + 1;
         } else {
             si = NULL;
             s->sh_addr = addr = pe_virtual_align(pe, addr);
         }
 
-        if (NULL == pe->thunk
-            && c == (data_section == rodata_section ? sec_data : sec_rdata))
-            pe->thunk = s;
-
         if (s == pe->thunk) {
             pe_build_imports(pe);
             pe_build_exports(pe);
-            if (pe->tls_size)
-                pe_build_tls(pe, NULL);
         }
 
         if (s == pe->reloc)
@@ -1367,7 +1366,7 @@ add_section:
         Section *s = s1->sections[sec_order[i]];
         int type = s->sh_type;
         int flags = s->sh_flags;
-        printf("section %-16s %-10s %p %04x %s,%s,%s\n",
+        printf("section %-16s %-10s %p %04x %2d %s,%s,%s\n",
             s->name,
             type == SHT_PROGBITS ? "progbits" :
             type == SHT_INIT_ARRAY ? "initarr" :
@@ -1378,6 +1377,7 @@ add_section:
             type == SHT_RELX ? "rel" : "???",
             s->sh_addr,
             (unsigned)s->data_offset,
+            s->sh_addralign,
             flags & SHF_ALLOC ? "alloc" : "",
             flags & SHF_WRITE ? "write" : "",
             flags & SHF_EXECINSTR ? "exec" : ""

tccrun.c

@@ -385,6 +385,9 @@ redo:
                 continue;
             }
 
+            if ((s->sh_flags & SHF_TLS) && length)
+                return tcc_error_noabort("thread-local storage not supported with -run");
+
             align = s->sh_addralign;
             if (++n == 1) {
 #if defined TCC_TARGET_I386 || defined TCC_TARGET_X86_64
@@ -486,13 +489,18 @@ static int protect_pages(void *ptr, unsigned long length, int mode)
         };
     if (mprotect(ptr, length, protect[mode]))
         return -1;
+#endif
 /* XXX: BSD sometimes dump core with bad system call */
-# if (defined TCC_TARGET_ARM && !TARGETOS_BSD) || defined TCC_TARGET_ARM64 || defined TCC_TARGET_RISCV64
+#if (defined TCC_TARGET_ARM && !TARGETOS_BSD) \
+    || defined TCC_TARGET_ARM64 || defined TCC_TARGET_RISCV64
     if (mode == 0 || mode == 3) {
+#ifdef _WIN32
+        FlushInstructionCache(GetCurrentProcess(), ptr, length);
+#else
         void __clear_cache(void *beginning, void *end);
         __clear_cache(ptr, (char *)ptr + length);
+#endif
     }
-# endif
 #endif
     return 0;
 }