Refactor Cortex-A8 erratum workaround in preparation
authorThomas Preud'homme <thomas.preudhomme@arm.com>
Tue, 10 May 2016 14:28:10 +0000 (15:28 +0100)
committerThomas Preud'homme <thomas.preudhomme@arm.com>
Tue, 10 May 2016 14:28:10 +0000 (15:28 +0100)
2016-05-10  Thomas Preud'homme  <thomas.preudhomme@arm.com>

bfd/
* elf32-arm.c (enum elf32_arm_stub_type): Delete
arm_stub_a8_veneer_lwm enumerator.
(arm_stub_a8_veneer_lwm): New unsigned constant to replace
aforementioned enumerator.
(struct elf32_arm_stub_hash_entry): Delete target_addend
field and add source_value.
(struct a8_erratum_fix): Delete addend field and add target_offset.
(stub_hash_newfunc): Initialize source_value field amd remove
initialization for target_addend.
(arm_build_one_stub): Stop special casing Thumb relocations: promote
the else to being always executed, moving the
arm_stub_a8_veneer_b_cond specific code in it.  Remove
stub_entry->target_addend from points_to computation.
(cortex_a8_erratum_scan): Store in a8_erratum_fix structure the offset
to target symbol from start of section rather than the offset from the
stub address.
(elf32_arm_size_stubs): Set stub_entry's source_value and target_value
fields from struct a8_erratum_fix's offset and target_offset
respectively.
(make_branch_to_a8_stub): Rename target variable to loc.  Compute
veneered_insn_loc and loc using stub_entry's source_value.

bfd/ChangeLog
bfd/elf32-arm.c

index c885dd09d57b4be329d73d706daf49388206f08b..0e10d3aef4e5a968f453c3405846ac837df93066 100644 (file)
@@ -1,3 +1,27 @@
+2016-05-10  Thomas Preud'homme  <thomas.preudhomme@arm.com>
+
+       * elf32-arm.c (enum elf32_arm_stub_type): Delete
+       arm_stub_a8_veneer_lwm enumerator.
+       (arm_stub_a8_veneer_lwm): New unsigned constant to replace
+       aforementioned enumerator.
+       (struct elf32_arm_stub_hash_entry): Delete target_addend
+       field and add source_value.
+       (struct a8_erratum_fix): Delete addend field and add target_offset.
+       (stub_hash_newfunc): Initialize source_value field amd remove
+       initialization for target_addend.
+       (arm_build_one_stub): Stop special casing Thumb relocations: promote
+       the else to being always executed, moving the
+       arm_stub_a8_veneer_b_cond specific code in it.  Remove
+       stub_entry->target_addend from points_to computation.
+       (cortex_a8_erratum_scan): Store in a8_erratum_fix structure the offset
+       to target symbol from start of section rather than the offset from the
+       stub address.
+       (elf32_arm_size_stubs): Set stub_entry's source_value and target_value
+       fields from struct a8_erratum_fix's offset and target_offset
+       respectively.
+       (make_branch_to_a8_stub): Rename target variable to loc.  Compute
+       veneered_insn_loc and loc using stub_entry's source_value.
+
 2016-05-10  Thomas Preud'homme  <thomas.preudhomme@arm.com>
 
        (elf32_arm_merge_eabi_attributes): Add merging logic for
index fa0d74fdc4777c5cfa00d34166fd57048b623baa..5f6941dbacc166dba354502cebfc10cf1bf2db5d 100644 (file)
@@ -2632,11 +2632,12 @@ enum elf32_arm_stub_type
 {
   arm_stub_none,
   DEF_STUBS
-  /* Note the first a8_veneer type.  */
-  arm_stub_a8_veneer_lwm = arm_stub_a8_veneer_b_cond
 };
 #undef DEF_STUB
 
+/* Note the first a8_veneer type.  */
+const unsigned arm_stub_a8_veneer_lwm = arm_stub_a8_veneer_b_cond;
+
 typedef struct
 {
   const insn_sequence* template_sequence;
@@ -2666,8 +2667,12 @@ struct elf32_arm_stub_hash_entry
   bfd_vma target_value;
   asection *target_section;
 
-  /* Offset to apply to relocation referencing target_value.  */
-  bfd_vma target_addend;
+  /* Same as above but for the source of the branch to the stub.  Used for
+     Cortex-A8 erratum workaround to patch it to branch to the stub.  As
+     such, source section does not need to be recorded since Cortex-A8 erratum
+     workaround stubs are only generated when both source and target are in the
+     same section.  */
+  bfd_vma source_value;
 
   /* The instruction which caused this stub to be generated (only valid for
      Cortex-A8 erratum workaround stubs at present).  */
@@ -2836,7 +2841,7 @@ struct a8_erratum_fix
   bfd *input_bfd;
   asection *section;
   bfd_vma offset;
-  bfd_vma addend;
+  bfd_vma target_offset;
   unsigned long orig_insn;
   char *stub_name;
   enum elf32_arm_stub_type stub_type;
@@ -3413,9 +3418,9 @@ stub_hash_newfunc (struct bfd_hash_entry *entry,
       eh = (struct elf32_arm_stub_hash_entry *) entry;
       eh->stub_sec = NULL;
       eh->stub_offset = 0;
+      eh->source_value = 0;
       eh->target_value = 0;
       eh->target_section = NULL;
-      eh->target_addend = 0;
       eh->orig_insn = 0;
       eh->stub_type = arm_stub_none;
       eh->stub_size = 0;
@@ -4450,65 +4455,36 @@ arm_build_one_stub (struct bfd_hash_entry *gen_entry,
   BFD_ASSERT (nrelocs != 0 && nrelocs <= MAXRELOCS);
 
   for (i = 0; i < nrelocs; i++)
-    if (template_sequence[stub_reloc_idx[i]].r_type == R_ARM_THM_JUMP24
-       || template_sequence[stub_reloc_idx[i]].r_type == R_ARM_THM_JUMP19
-       || template_sequence[stub_reloc_idx[i]].r_type == R_ARM_THM_CALL
-       || template_sequence[stub_reloc_idx[i]].r_type == R_ARM_THM_XPC22)
-      {
-       Elf_Internal_Rela rel;
-       bfd_boolean unresolved_reloc;
-       char *error_message;
-       enum arm_st_branch_type branch_type
-         = (template_sequence[stub_reloc_idx[i]].r_type != R_ARM_THM_XPC22
-            ? ST_BRANCH_TO_THUMB : ST_BRANCH_TO_ARM);
-       bfd_vma points_to = sym_value + stub_entry->target_addend;
-
-       rel.r_offset = stub_entry->stub_offset + stub_reloc_offset[i];
-       rel.r_info = ELF32_R_INFO (0,
-                                  template_sequence[stub_reloc_idx[i]].r_type);
-       rel.r_addend = template_sequence[stub_reloc_idx[i]].reloc_addend;
-
-       if (stub_entry->stub_type == arm_stub_a8_veneer_b_cond && i == 0)
-         /* The first relocation in the elf32_arm_stub_a8_veneer_b_cond[]
-            template should refer back to the instruction after the original
-            branch.  */
-         points_to = sym_value;
-
-       /* There may be unintended consequences if this is not true.  */
-       BFD_ASSERT (stub_entry->h == NULL);
-
-       /* Note: _bfd_final_link_relocate doesn't handle these relocations
-          properly.  We should probably use this function unconditionally,
-          rather than only for certain relocations listed in the enclosing
-          conditional, for the sake of consistency.  */
-       elf32_arm_final_link_relocate (elf32_arm_howto_from_type
-           (template_sequence[stub_reloc_idx[i]].r_type),
-         stub_bfd, info->output_bfd, stub_sec, stub_sec->contents, &rel,
-         points_to, info, stub_entry->target_section, "", STT_FUNC,
-         branch_type, (struct elf_link_hash_entry *) stub_entry->h,
-         &unresolved_reloc, &error_message);
-      }
-    else
-      {
-       Elf_Internal_Rela rel;
-       bfd_boolean unresolved_reloc;
-       char *error_message;
-       bfd_vma points_to = sym_value + stub_entry->target_addend
-         + template_sequence[stub_reloc_idx[i]].reloc_addend;
-
-       rel.r_offset = stub_entry->stub_offset + stub_reloc_offset[i];
-       rel.r_info = ELF32_R_INFO (0,
-                                  template_sequence[stub_reloc_idx[i]].r_type);
-       rel.r_addend = 0;
-
-       elf32_arm_final_link_relocate (elf32_arm_howto_from_type
-           (template_sequence[stub_reloc_idx[i]].r_type),
-         stub_bfd, info->output_bfd, stub_sec, stub_sec->contents, &rel,
-         points_to, info, stub_entry->target_section, "", STT_FUNC,
-         stub_entry->branch_type,
-         (struct elf_link_hash_entry *) stub_entry->h, &unresolved_reloc,
-         &error_message);
-      }
+    {
+      Elf_Internal_Rela rel;
+      bfd_boolean unresolved_reloc;
+      char *error_message;
+      bfd_vma points_to =
+       sym_value + template_sequence[stub_reloc_idx[i]].reloc_addend;
+
+      rel.r_offset = stub_entry->stub_offset + stub_reloc_offset[i];
+      rel.r_info = ELF32_R_INFO (0,
+                                template_sequence[stub_reloc_idx[i]].r_type);
+      rel.r_addend = 0;
+
+      if (stub_entry->stub_type == arm_stub_a8_veneer_b_cond && i == 0)
+       /* The first relocation in the elf32_arm_stub_a8_veneer_b_cond[]
+          template should refer back to the instruction after the original
+          branch.  We use target_section as Cortex-A8 erratum workaround stubs
+          are only generated when both source and target are in the same
+          section.  */
+       points_to = stub_entry->target_section->output_section->vma
+                   + stub_entry->target_section->output_offset
+                   + stub_entry->source_value;
+
+      elf32_arm_final_link_relocate (elf32_arm_howto_from_type
+         (template_sequence[stub_reloc_idx[i]].r_type),
+          stub_bfd, info->output_bfd, stub_sec, stub_sec->contents, &rel,
+          points_to, info, stub_entry->target_section, "", STT_FUNC,
+          stub_entry->branch_type,
+          (struct elf_link_hash_entry *) stub_entry->h, &unresolved_reloc,
+          &error_message);
+    }
 
   return TRUE;
 #undef MAXRELOCS
@@ -5101,7 +5077,8 @@ cortex_a8_erratum_scan (bfd *input_bfd,
                          a8_fixes[num_a8_fixes].input_bfd = input_bfd;
                          a8_fixes[num_a8_fixes].section = section;
                          a8_fixes[num_a8_fixes].offset = i;
-                         a8_fixes[num_a8_fixes].addend = offset;
+                         a8_fixes[num_a8_fixes].target_offset =
+                           target - base_vma;
                          a8_fixes[num_a8_fixes].orig_insn = insn;
                          a8_fixes[num_a8_fixes].stub_name = stub_name;
                          a8_fixes[num_a8_fixes].stub_type = stub_type;
@@ -5680,9 +5657,9 @@ elf32_arm_size_stubs (bfd *output_bfd,
          stub_entry->stub_offset = 0;
          stub_entry->id_sec = link_sec;
          stub_entry->stub_type = a8_fixes[i].stub_type;
+         stub_entry->source_value = a8_fixes[i].offset;
          stub_entry->target_section = a8_fixes[i].section;
-         stub_entry->target_value = a8_fixes[i].offset;
-         stub_entry->target_addend = a8_fixes[i].addend;
+         stub_entry->target_value = a8_fixes[i].target_offset;
          stub_entry->orig_insn = a8_fixes[i].orig_insn;
          stub_entry->branch_type = a8_fixes[i].branch_type;
 
@@ -16277,7 +16254,7 @@ make_branch_to_a8_stub (struct bfd_hash_entry *gen_entry,
   bfd_vma veneered_insn_loc, veneer_entry_loc;
   bfd_signed_vma branch_offset;
   bfd *abfd;
-  unsigned int target;
+  unsigned int loc;
 
   stub_entry = (struct elf32_arm_stub_hash_entry *) gen_entry;
   data = (struct a8_branch_to_stub_data *) in_arg;
@@ -16288,9 +16265,11 @@ make_branch_to_a8_stub (struct bfd_hash_entry *gen_entry,
 
   contents = data->contents;
 
+  /* We use target_section as Cortex-A8 erratum workaround stubs are only
+     generated when both source and target are in the same section.  */
   veneered_insn_loc = stub_entry->target_section->output_section->vma
                      + stub_entry->target_section->output_offset
-                     + stub_entry->target_value;
+                     + stub_entry->source_value;
 
   veneer_entry_loc = stub_entry->stub_sec->output_section->vma
                     + stub_entry->stub_sec->output_offset
@@ -16302,7 +16281,7 @@ make_branch_to_a8_stub (struct bfd_hash_entry *gen_entry,
   branch_offset = veneer_entry_loc - veneered_insn_loc - 4;
 
   abfd = stub_entry->target_section->owner;
-  target = stub_entry->target_value;
+  loc = stub_entry->source_value;
 
   /* We attempt to avoid this condition by setting stubs_always_after_branch
      in elf32_arm_size_stubs if we've enabled the Cortex-A8 erratum workaround.
@@ -16363,8 +16342,8 @@ make_branch_to_a8_stub (struct bfd_hash_entry *gen_entry,
       return FALSE;
     }
 
-  bfd_put_16 (abfd, (branch_insn >> 16) & 0xffff, &contents[target]);
-  bfd_put_16 (abfd, branch_insn & 0xffff, &contents[target + 2]);
+  bfd_put_16 (abfd, (branch_insn >> 16) & 0xffff, &contents[loc]);
+  bfd_put_16 (abfd, branch_insn & 0xffff, &contents[loc + 2]);
 
   return TRUE;
 }
This page took 0.036779 seconds and 4 git commands to generate.