[ELF] Do not create .plt.got for ppc64/ppc64le

rui314 · Oct 14, 2022 · 30dd63c · 30dd63c
1 parent 2dbbc54
commit 30dd63c
Show file tree

Hide file tree

Showing 5 changed files with 30 additions and 29 deletions.
diff --git a/elf/arch-ppc64v1.cc b/elf/arch-ppc64v1.cc
@@ -105,14 +105,10 @@ void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
   *(ub32 *)(buf + 4) = 0x4b00'0000 | (offset & 0x00ff'ffff); // b  plt0
 }
 
+// .plt.got is not necessary on PPC64 because range extension thunks
+// directly read GOT entries and jump there.
 template <>
-void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
-  // No one uses .got.plt at runtime because all calls to .got.plt are
-  // made via range extension thunks. Range extension thunks directly
-  // calls the final destination by reading a .got entry. Here, we just
-  // set a dummy instruction.
-  *(ub32 *)buf = 0x6000'0000; // nop
-}
+void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {}
 
 template <>
 void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,

diff --git a/elf/arch-ppc64v2.cc b/elf/arch-ppc64v2.cc
@@ -121,17 +121,10 @@ void write_plt_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
   *(ul32 *)buf |= (ctx.plt->shdr.sh_addr - sym.get_plt_addr(ctx)) & 0x00ff'ffff;
 }
 
+// .plt.got is not necessary on PPC64 because range extension thunks
+// directly read GOT entries and jump there.
 template <>
-void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {
-  // No one uses .got.plt at runtime because all calls to .got.plt are
-  // made via range extension thunks. Range extension thunks directly
-  // calls the final destination by reading a .got entry. Here, we just
-  // set a dummy instruction.
-  //
-  // I believe we can completely elimnate .got.plt, but saving 4 bytes
-  // for each GOTPLT entry doesn't seem to be worth its complexity.
-  *(ul32 *)buf = 0x6000'0000; // nop
-}
+void write_pltgot_entry(Context<E> &ctx, u8 *buf, Symbol<E> &sym) {}
 
 template <>
 void EhFrameSection<E>::apply_reloc(Context<E> &ctx, const ElfRel<E> &rel,

diff --git a/elf/arch-s390x.cc b/elf/arch-s390x.cc
@@ -14,14 +14,26 @@
 // aligned to 2 bytes boundaries. Despite unfamiliarty, I found that it
 // just feels like a 64-bit i386 in a parallel universe.
 //
-// Its psABI reserves %r0 and %r1 as scratch registers so we can use them
-// in our PLT. %r2-%r6 are used for parameter passing. %r2 is also used to
-// return a value. In position independent code, %r12 usually contains the
-// address of GOT. %r14 usually contains a return address. %r15 is a stack
-// pointer. Access registers %a0 and %a1 contain the upper 32 bits and
-// the lower 32 bits of the thread pointer, respectively.
+// Here is the register usage in this ABI:
 //
-// https://uclibc.org/docs/psABI-s390x.pdf
+//   r0-r1: reserved as scratch registers so we can use them in our PLT
+//   r2:    parameter passing and return values
+//   r3-r6: parameter passing
+//   r12:   address of GOT if position-independent code
+//   r14:   return address
+//   r15:   stack pointer
+//   a1:    upper 32 bits of TP (thread pointer)
+//   a2:    lower 32 bits of TP (thread pointer)
+//
+// TLS is supported on s390x in the same way as it is on other targets
+// with one exeption. On other targets, __tls_get_addr is used to get an
+// address of a thread-local variable. On s390x, __tls_get_offset is used
+// instead. The difference is __tls_get_offset returns an address of a
+// thread-local variable as an offset from TP. So we need to add TP to a
+// return value before use. I don't know why it is different, but that is
+// the way it is.
+//
+// https://github.com/IBM/s390x-abi/releases/download/v1.6/lzsabi_s390x.pdf
 
 #include "mold.h"
 
@@ -33,7 +45,7 @@ template <>
 void write_plt_header(Context<E> &ctx, u8 *buf) {
   static u8 insn[] = {
     0xe3, 0x00, 0xf0, 0x38, 0x00, 0x24, // stg   %r0, 56(%r15)
-    0xc0, 0x10, 0, 0, 0, 0,             // larl  %r1, GOT_OFFSET
+    0xc0, 0x10, 0, 0, 0, 0,             // larl  %r1, GOTPLT_OFFSET
     0xd2, 0x07, 0xf0, 0x30, 0x10, 0x08, // mvc   48(8, %r15), 8(%r1)
     0xe3, 0x10, 0x10, 0x10, 0x00, 0x04, // lg    %r1, 16(%r1)
     0x07, 0xf1,                         // br    %r1

diff --git a/elf/elf.h b/elf/elf.h
@@ -2669,7 +2669,7 @@ struct PPC64V1 {
   static constexpr u32 e_machine = EM_PPC64;
   static constexpr u32 plt_hdr_size = 52;
   static constexpr u32 plt_size = 8;
-  static constexpr u32 pltgot_size = 4;
+  static constexpr u32 pltgot_size = 0;
   static constexpr u32 tls_dtv_offset = 0x8000;
   static constexpr u32 thunk_hdr_size = 0;
   static constexpr u32 thunk_size = 28;
@@ -2706,7 +2706,7 @@ struct PPC64V2 {
   static constexpr u32 e_machine = EM_PPC64;
   static constexpr u32 plt_hdr_size = 60;
   static constexpr u32 plt_size = 4;
-  static constexpr u32 pltgot_size = 4;
+  static constexpr u32 pltgot_size = 0;
   static constexpr u32 tls_dtv_offset = 0x8000;
   static constexpr u32 thunk_hdr_size = 0;
   static constexpr u32 thunk_size = 20;

diff --git a/elf/output-chunks.cc b/elf/output-chunks.cc
@@ -1496,9 +1496,9 @@ void PltGotSection<E>::add_symbol(Context<E> &ctx, Symbol<E> *sym) {
   assert(!sym->has_plt(ctx));
   assert(sym->has_got(ctx));
 
-  sym->set_pltgot_idx(ctx, this->shdr.sh_size / E::pltgot_size);
-  this->shdr.sh_size += E::pltgot_size;
+  sym->set_pltgot_idx(ctx, symbols.size());
   symbols.push_back(sym);
+  this->shdr.sh_size = symbols.size() * E::pltgot_size;
 }
 
 template <typename E>