mirror of
https://github.com/rui314/mold.git
synced 2024-09-22 02:20:51 +03:00
Support AARCH64 relocations
With this change, mold can now create a "Hello world" program for the AArch64 target.
This commit is contained in:
Rui Ueyama
parent
fc4aafa974
commit
41497d7990
288
arch_aarch64.cc
288
arch_aarch64.cc
@ -1,27 +1,311 @@
|
||||
#include "mold.h"
|
||||
|
||||
static void write_addr(u8 *buf, u64 val) {
|
||||
u32 hi = (val & 0x1ffffc) << 3;
|
||||
u32 lo = (val & 3) << 29;
|
||||
*(u32 *)buf = (*(u32 *)buf & 0x9f00001f) | hi | lo;
|
||||
}
|
||||
|
||||
static u64 extract(u64 val, u64 hi, u64 lo) {
|
||||
return (val >> lo) & (((u64)1 << (hi - lo + 1)) - 1);
|
||||
}
|
||||
|
||||
static u64 page(u64 val) {
|
||||
return val & ~(u64)0xfff;
|
||||
}
|
||||
|
||||
template <>
|
||||
void PltSection<AARCH64>::copy_buf(Context<AARCH64> &ctx) {
|
||||
u8 *buf = ctx.buf + this->shdr.sh_offset;
|
||||
|
||||
// Write PLT header
|
||||
static const u8 plt0[] = {
|
||||
0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]!
|
||||
0x10, 0x00, 0x00, 0x90, // adrp x16, PLTGOT[2]
|
||||
0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, PLTGOT[2]]
|
||||
0x10, 0x02, 0x00, 0x91, // add x16, x16, PLTGOT[2]
|
||||
0x20, 0x02, 0x1f, 0xd6, // br x17
|
||||
0x1f, 0x20, 0x03, 0xd5, // nop
|
||||
0x1f, 0x20, 0x03, 0xd5, // nop
|
||||
0x1f, 0x20, 0x03, 0xd5, // nop
|
||||
};
|
||||
|
||||
u64 got = ctx.got->shdr.sh_addr;
|
||||
u64 plt = ctx.plt->shdr.sh_addr;
|
||||
|
||||
memcpy(buf, plt0, sizeof(plt0));
|
||||
write_addr(buf + 4, (page(got + 16) - page(plt + 4)) >> 12);
|
||||
*(u32 *)(buf + 8) |= ((got + 16) & 0xfff) << 10;
|
||||
*(u32 *)(buf + 12) |= ((got + 16) & 0xfff) << 10;
|
||||
|
||||
// Write PLT entries
|
||||
i64 relplt_idx = 0;
|
||||
|
||||
for (Symbol<AARCH64> *sym : symbols) {
|
||||
u8 *ent = buf + sym->get_plt_idx(ctx) * AARCH64::plt_size;
|
||||
|
||||
static const u8 data[] = {
|
||||
0x10, 0x00, 0x00, 0x90, // adrp x16, PLTGOT[n]
|
||||
0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, PLTGOT[n]]
|
||||
0x10, 0x02, 0x00, 0x91, // add x16, x16, PLTGOT[n]
|
||||
0x20, 0x02, 0x1f, 0xd6, // br x17
|
||||
};
|
||||
|
||||
u64 gotplt = sym->get_gotplt_addr(ctx);
|
||||
u64 plt = sym->get_plt_addr(ctx);
|
||||
|
||||
memcpy(ent, data, sizeof(data));
|
||||
write_addr(ent, (page(gotplt) - page(plt)) >> 12);
|
||||
*(u32 *)(ent + 4) |= extract(got, 11, 3) << 10;
|
||||
*(u32 *)(ent + 8) |= (got & 0xfff) << 10;
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
void PltGotSection<AARCH64>::copy_buf(Context<AARCH64> &ctx) {
|
||||
u8 *buf = ctx.buf + this->shdr.sh_offset;
|
||||
|
||||
for (Symbol<AARCH64> *sym : symbols) {
|
||||
u8 *ent = buf + sym->get_pltgot_idx(ctx) * AARCH64::pltgot_size;
|
||||
|
||||
static const u8 data[] = {
|
||||
0x10, 0x00, 0x00, 0x90, // adrp x16, GOT[n]
|
||||
0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, GOT[n]]
|
||||
0x10, 0x02, 0x00, 0x91, // add x16, x16, GOT[n]
|
||||
0x20, 0x02, 0x1f, 0xd6, // br x17
|
||||
};
|
||||
|
||||
u64 got = sym->get_got_addr(ctx);
|
||||
u64 plt = sym->get_plt_addr(ctx);
|
||||
|
||||
memcpy(ent, data, sizeof(data));
|
||||
write_addr(ent, (page(got) - page(plt)) >> 12);
|
||||
*(u32 *)(ent + 4) |= extract(got, 11, 3) << 10;
|
||||
*(u32 *)(ent + 8) |= (got & 0xfff) << 10;
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
void EhFrameSection<AARCH64>::apply_reloc(Context<AARCH64> &ctx,
|
||||
ElfRel<AARCH64> &rel,
|
||||
u64 loc, u64 val) {
|
||||
ElfRel<AARCH64> &rel,
|
||||
u64 loc, u64 val) {
|
||||
u8 *base = ctx.buf + this->shdr.sh_offset;
|
||||
|
||||
switch (rel.r_type) {
|
||||
case R_AARCH64_PREL32:
|
||||
*(u32 *)(base + loc) = val;
|
||||
return;
|
||||
}
|
||||
unreachable(ctx);
|
||||
}
|
||||
|
||||
template <>
|
||||
void InputSection<AARCH64>::apply_reloc_alloc(Context<AARCH64> &ctx, u8 *base) {
|
||||
ElfRel<AARCH64> *dynrel = nullptr;
|
||||
std::span<ElfRel<AARCH64>> rels = get_rels(ctx);
|
||||
i64 frag_idx = 0;
|
||||
|
||||
if (ctx.reldyn)
|
||||
dynrel = (ElfRel<AARCH64> *)(ctx.buf + ctx.reldyn->shdr.sh_offset +
|
||||
file.reldyn_offset + this->reldyn_offset);
|
||||
|
||||
for (i64 i = 0; i < rels.size(); i++) {
|
||||
const ElfRel<AARCH64> &rel = rels[i];
|
||||
if (rel.r_type == R_AARCH64_NONE)
|
||||
continue;
|
||||
|
||||
Symbol<AARCH64> &sym = *file.symbols[rel.r_sym];
|
||||
u8 *loc = base + rel.r_offset;
|
||||
|
||||
const SectionFragmentRef<AARCH64> *ref = nullptr;
|
||||
if (rel_fragments && rel_fragments[frag_idx].idx == i)
|
||||
ref = &rel_fragments[frag_idx++];
|
||||
|
||||
#define S (ref ? ref->frag->get_addr(ctx) : sym.get_addr(ctx))
|
||||
#define A (ref ? ref->addend : rel.r_addend)
|
||||
#define P (output_section->shdr.sh_addr + offset + rel.r_offset)
|
||||
#define G (sym.get_got_addr(ctx) - ctx.got->shdr.sh_addr)
|
||||
#define GOT ctx.got->shdr.sh_addr
|
||||
|
||||
switch (rel_exprs[i]) {
|
||||
case R_BASEREL:
|
||||
*dynrel++ = {P, R_AARCH64_RELATIVE, 0, (i64)(S + A)};
|
||||
*(u64 *)loc = S + A;
|
||||
continue;
|
||||
case R_DYN:
|
||||
*dynrel++ = {P, R_AARCH64_ABS64, (u32)sym.get_dynsym_idx(ctx), A};
|
||||
*(u64 *)loc = A;
|
||||
continue;
|
||||
}
|
||||
|
||||
switch (rel.r_type) {
|
||||
case R_AARCH64_ABS64:
|
||||
*(u64 *)loc = S + A;
|
||||
continue;
|
||||
case R_AARCH64_LDST8_ABS_LO12_NC:
|
||||
*(u32 *)loc |= extract(S + A, 11, 0) << 10;
|
||||
continue;
|
||||
case R_AARCH64_LDST32_ABS_LO12_NC:
|
||||
*(u32 *)loc |= extract(S + A, 11, 2) << 10;
|
||||
continue;
|
||||
case R_AARCH64_LDST64_ABS_LO12_NC:
|
||||
*(u32 *)loc |= extract(S + A, 11, 3) << 10;
|
||||
continue;
|
||||
case R_AARCH64_ADD_ABS_LO12_NC:
|
||||
*(u32 *)loc |= extract(S + A, 11, 0) << 10;
|
||||
continue;
|
||||
case R_AARCH64_MOVW_UABS_G0_NC:
|
||||
*(u32 *)loc |= extract(S + A, 15, 0) << 5;
|
||||
continue;
|
||||
case R_AARCH64_MOVW_UABS_G1_NC:
|
||||
*(u32 *)loc |= extract(S + A, 31, 16) << 5;
|
||||
continue;
|
||||
case R_AARCH64_MOVW_UABS_G2_NC:
|
||||
*(u32 *)loc |= extract(S + A, 47, 32) << 5;
|
||||
continue;
|
||||
case R_AARCH64_MOVW_UABS_G3:
|
||||
*(u32 *)loc |= extract(S + A, 63, 48) << 5;
|
||||
continue;
|
||||
case R_AARCH64_ADR_GOT_PAGE:
|
||||
write_addr(loc, extract(page(GOT + A) - page(P), 32, 12));
|
||||
continue;
|
||||
case R_AARCH64_ADR_PREL_PG_HI21:
|
||||
write_addr(loc, extract(page(S + A) - page(P), 32, 12));
|
||||
continue;
|
||||
case R_AARCH64_CALL26:
|
||||
case R_AARCH64_JUMP26:
|
||||
if (sym.esym().is_undef_weak())
|
||||
*(u32 *)loc |= 1;
|
||||
else
|
||||
*(u32 *)loc |= ((S + A - P) >> 2) & 0x3ffffff;
|
||||
continue;
|
||||
case R_AARCH64_PREL32:
|
||||
*(u32 *)loc = S + A - P;
|
||||
continue;
|
||||
case R_AARCH64_LD64_GOT_LO12_NC:
|
||||
*(u32 *)loc |= extract(G + GOT + A, 11, 3) << 10;
|
||||
continue;
|
||||
case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21:
|
||||
write_addr(loc, (page(sym.get_gottp_addr(ctx) + A) - page(P)) >> 12);
|
||||
continue;
|
||||
case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC:
|
||||
*(u32 *)loc |= extract(sym.get_gottp_addr(ctx) + A, 11, 3) << 10;
|
||||
continue;
|
||||
case R_AARCH64_TLSLE_ADD_TPREL_HI12:
|
||||
*(u32 *)loc |= extract(S + A - ctx.tls_begin + 16, 23, 12) << 10;
|
||||
continue;
|
||||
case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
|
||||
*(u32 *)loc |= extract(S + A - ctx.tls_begin + 16, 11, 0) << 10;
|
||||
continue;
|
||||
default:
|
||||
unreachable(ctx);
|
||||
}
|
||||
|
||||
#undef S
|
||||
#undef A
|
||||
#undef P
|
||||
#undef G
|
||||
#undef GOT
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
void InputSection<AARCH64>::apply_reloc_nonalloc(Context<AARCH64> &ctx, u8 *base) {
|
||||
std::span<ElfRel<AARCH64>> rels = get_rels(ctx);
|
||||
i64 frag_idx = 0;
|
||||
|
||||
for (i64 i = 0; i < rels.size(); i++) {
|
||||
const ElfRel<AARCH64> &rel = rels[i];
|
||||
if (rel.r_type == R_AARCH64_NONE)
|
||||
continue;
|
||||
|
||||
Symbol<AARCH64> &sym = *file.symbols[rel.r_sym];
|
||||
u8 *loc = base + rel.r_offset;
|
||||
|
||||
if (!sym.file) {
|
||||
report_undef(ctx, sym);
|
||||
continue;
|
||||
}
|
||||
|
||||
const SectionFragmentRef<AARCH64> *ref = nullptr;
|
||||
if (rel_fragments && rel_fragments[frag_idx].idx == i)
|
||||
ref = &rel_fragments[frag_idx++];
|
||||
|
||||
switch (rel.r_type) {
|
||||
// TODO
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
void InputSection<AARCH64>::scan_relocations(Context<AARCH64> &ctx) {
|
||||
assert(shdr.sh_flags & SHF_ALLOC);
|
||||
|
||||
this->reldyn_offset = file.num_dynrel * sizeof(ElfRel<AARCH64>);
|
||||
std::span<ElfRel<AARCH64>> rels = get_rels(ctx);
|
||||
bool is_writable = (shdr.sh_flags & SHF_WRITE);
|
||||
|
||||
// Scan relocations
|
||||
for (i64 i = 0; i < rels.size(); i++) {
|
||||
const ElfRel<AARCH64> &rel = rels[i];
|
||||
if (rel.r_type == R_AARCH64_NONE)
|
||||
continue;
|
||||
|
||||
Symbol<AARCH64> &sym = *file.symbols[rel.r_sym];
|
||||
u8 *loc = (u8 *)(contents.data() + rel.r_offset);
|
||||
|
||||
if (!sym.file) {
|
||||
report_undef(ctx, sym);
|
||||
continue;
|
||||
}
|
||||
|
||||
if (sym.get_type() == STT_GNU_IFUNC) {
|
||||
sym.flags |= NEEDS_GOT;
|
||||
sym.flags |= NEEDS_PLT;
|
||||
}
|
||||
|
||||
switch (rel.r_type) {
|
||||
case R_AARCH64_ABS64: {
|
||||
Action table[][4] = {
|
||||
// Absolute Local Imported data Imported code
|
||||
{ NONE, BASEREL, DYNREL, DYNREL }, // DSO
|
||||
{ NONE, BASEREL, DYNREL, DYNREL }, // PIE
|
||||
{ NONE, NONE, DYNREL, PLT }, // PDE
|
||||
};
|
||||
dispatch(ctx, table, i);
|
||||
break;
|
||||
}
|
||||
case R_AARCH64_ADR_GOT_PAGE:
|
||||
sym.flags |= NEEDS_GOT;
|
||||
break;
|
||||
case R_AARCH64_CALL26:
|
||||
case R_AARCH64_JUMP26:
|
||||
if (sym.is_imported)
|
||||
sym.flags |= NEEDS_PLT;
|
||||
break;
|
||||
case R_AARCH64_LD64_GOT_LO12_NC:
|
||||
sym.flags |= NEEDS_GOT;
|
||||
break;
|
||||
case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21:
|
||||
case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC:
|
||||
sym.flags |= NEEDS_GOTTP;
|
||||
break;
|
||||
case R_AARCH64_ADD_ABS_LO12_NC:
|
||||
case R_AARCH64_ADR_PREL_PG_HI21:
|
||||
case R_AARCH64_LDST64_ABS_LO12_NC:
|
||||
case R_AARCH64_LDST32_ABS_LO12_NC:
|
||||
case R_AARCH64_LDST8_ABS_LO12_NC:
|
||||
case R_AARCH64_MOVW_UABS_G0_NC:
|
||||
case R_AARCH64_MOVW_UABS_G1_NC:
|
||||
case R_AARCH64_MOVW_UABS_G2_NC:
|
||||
case R_AARCH64_MOVW_UABS_G3:
|
||||
case R_AARCH64_PREL32:
|
||||
case R_AARCH64_TLSLE_ADD_TPREL_HI12:
|
||||
case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC:
|
||||
break;
|
||||
default:
|
||||
Error(ctx) << *this << ": unknown relocation: "
|
||||
<< rel_to_string<AARCH64>(rel.r_type);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
10
elf.h
10
elf.h
@ -860,16 +860,16 @@ struct AARCH64 {
|
||||
static constexpr u32 R_JUMP_SLOT = R_AARCH64_JUMP_SLOT;
|
||||
static constexpr u32 R_RELATIVE = R_AARCH64_RELATIVE;
|
||||
static constexpr u32 R_IRELATIVE = R_AARCH64_IRELATIVE;
|
||||
static constexpr u32 R_DTPOFF = R_AARCH64_NONE;
|
||||
static constexpr u32 R_TPOFF = R_AARCH64_NONE;
|
||||
static constexpr u32 R_DTPMOD = R_AARCH64_NONE;
|
||||
static constexpr u32 R_TLSDESC = R_AARCH64_NONE;
|
||||
static constexpr u32 R_DTPOFF = R_AARCH64_TLS_DTPREL64;
|
||||
static constexpr u32 R_TPOFF = R_AARCH64_TLS_TPREL64;
|
||||
static constexpr u32 R_DTPMOD = R_AARCH64_TLS_DTPMOD64;
|
||||
static constexpr u32 R_TLSDESC = R_AARCH64_TLSDESC;
|
||||
|
||||
static constexpr u32 wordsize = 8;
|
||||
static constexpr u32 e_machine = EM_AARCH64;
|
||||
static constexpr u32 plt_hdr_size = 32;
|
||||
static constexpr u32 plt_size = 16;
|
||||
static constexpr u32 pltgot_size = 8;
|
||||
static constexpr u32 pltgot_size = 16;
|
||||
static constexpr bool is_rel = false;
|
||||
static constexpr bool is_le = true;
|
||||
};
|
||||
|
||||
@ -640,7 +640,8 @@ void ObjectFile<E>::initialize_mergeable_sections(Context<E> &ctx) {
|
||||
|
||||
for (i64 i = 0; i < sections.size(); i++) {
|
||||
std::unique_ptr<InputSection<E>> &isec = sections[i];
|
||||
if (isec && isec->is_alive && (isec->shdr.sh_flags & SHF_MERGE)) {
|
||||
if (isec && isec->is_alive && (isec->shdr.sh_flags & SHF_MERGE) &&
|
||||
isec->relsec_idx == -1) {
|
||||
mergeable_sections[i] = split_section(ctx, *isec);
|
||||
isec->is_alive = false;
|
||||
}
|
||||
|
||||
@ -826,11 +826,15 @@ void GotSection<E>::copy_buf(Context<E> &ctx) {
|
||||
sym->get_dynsym_idx(ctx));
|
||||
|
||||
for (Symbol<E> *sym : gottp_syms) {
|
||||
if (sym->is_imported)
|
||||
if (sym->is_imported) {
|
||||
*rel++ = reloc<E>(sym->get_gottp_addr(ctx), E::R_TPOFF,
|
||||
sym->get_dynsym_idx(ctx));
|
||||
else
|
||||
} else if (E::e_machine == EM_386 || E::e_machine == EM_X86_64) {
|
||||
buf[sym->get_gottp_idx(ctx)] = sym->get_addr(ctx) - ctx.tls_end;
|
||||
} else {
|
||||
assert(E::e_machine == EM_AARCH64);
|
||||
buf[sym->get_gottp_idx(ctx)] = sym->get_addr(ctx) - ctx.tls_begin + 16;
|
||||
}
|
||||
}
|
||||
|
||||
if (tlsld_idx != -1)
|
||||
|
||||
@ -5,7 +5,7 @@ echo -n "Testing $(basename -s .sh $0) ... "
|
||||
t=$(pwd)/tmp/$(basename -s .sh $0)
|
||||
mkdir -p $t
|
||||
|
||||
cat <<EOF | aarch64-linux-gnu-gcc-10 -o $t/a.o -c -xc -
|
||||
cat <<EOF | aarch64-linux-gnu-gcc -o $t/a.o -c -xc -
|
||||
#include <stdio.h>
|
||||
|
||||
int main() {
|
||||
@ -14,11 +14,12 @@ int main() {
|
||||
}
|
||||
EOF
|
||||
|
||||
aarch64-linux-gnu-gcc-10 -B`pwd`/.. -o $t/exe $t/a.o -static
|
||||
aarch64-linux-gnu-gcc -B`pwd`/.. -o $t/exe $t/a.o -static
|
||||
|
||||
readelf -p .comment $t/exe | grep -qw mold
|
||||
|
||||
readelf -a $t/exe > $t/log
|
||||
grep -Pq 'Machine:\s+AArch64' $t/log
|
||||
|
||||
# Generated executable can't run yet
|
||||
# qemu-aarch64 -L /usr/aarch64-linux-gnu $t/exe | grep -q 'Hello world'
|
||||
qemu-aarch64 -L /usr/aarch64-linux-gnu $t/exe | grep -q 'Hello world'
|
||||
|
||||
echo OK
|
||||
|
||||
Loading…
Reference in New Issue
Block a user