Re: [PATCH] Enable use of ARMv5t BLX for Thumb-mode calls via PLT

[Julian Brown <julian@codesourcery.com>]

[-- Attachment #1: Type: text/plain, Size: 1905 bytes --]

Daniel Jacobowitz wrote:
> On Mon, Apr 11, 2005 at 07:55:12PM +0100, Julian Brown wrote:
> 
>>With this patch, the option "--use-thumb-blx" can be passed to the 
>>linker to enable the use of the BLX instruction when the target supports 
>>it. The format of the PLT itself isn't changed by this patch: there is 
>>still a Thumb stub before each ARM PLT entry, but it is now unused. The 
>>ARM mode entry is called directly instead.
> 
> 
> The Thumb stub is never referenced from outside of the object.  So, if
> we aren't going to branch to it, no point in emitting it either.  Patch
> looks plausible to me...

This is a new version of the patch which doesn't emit Thumb stubs before 
PLT entries if BLX is being used to call them, and also renames the 
switch to --use-blx as per Paul's suggestion.

Re-tested on arm-none-eabi and arm-none-symbianelf.

OK to apply?

ChangeLog:

bfd:
   * bfd-in.h (bfd_elf32_arm_set_target_relocs): Update prototype.
   * bfd-in2.h: Regenerate.
   * elf32-arm.c (elf32_arm_link_hash_table): New field, 'use_blx'.
   (elf32_arm_link_hash_table_create): Initialise fix_v4bx, use_blx.
   (bfd_elf32_arm_set_target_relocs): Handle use_blx.
   (elf32_arm_final_link_relocate): Use Thumb BLX for R_ARM_THM_PC22
   relocations if requested to.
   (allocate_dynrelocs): Don't count size of omitted Thumb stubs based on
   use_blx rather than symbian_p.
   (elf32_arm_finish_dynamic_symbol): Don't output Thumb PLT stubs if
   use_blx is in effect.
   (elf32_arm_symbian_link_hash_table_create): Enable use_blx by default
   for SymbianOS.

ld:
   * ld.texinfo: Document --use-blx.
   * emultempl/armelf.em (use_blx): New variable.
   (arm_elf_create_output_section_statements): Communicate value of
   use_blx to bfd.
   (PARSE_AND_LIST_PROLOGUE): Add OPTION_USE_BLX.
   (PARSE_AND_LIST_OPTIONS): Add --use-blx option.
   (PARSE_AND_LIST_ARGS_CASES): Add OPTION_USE_BLX case.

[-- Attachment #2: patch-4 --]
[-- Type: text/plain, Size: 9834 bytes --]

Index: bfd/bfd-in.h
===================================================================
RCS file: /cvs/src/src/bfd/bfd-in.h,v
retrieving revision 1.97
diff -c -p -r1.97 bfd-in.h
*** bfd/bfd-in.h	30 Mar 2005 17:19:27 -0000	1.97
--- bfd/bfd-in.h	12 Apr 2005 20:57:45 -0000
*************** extern bfd_boolean bfd_elf32_arm_process
*** 816,822 ****
    (bfd *, struct bfd_link_info *, int);
  
  void bfd_elf32_arm_set_target_relocs
!   (struct bfd_link_info *, int, char *, int);
  
  extern bfd_boolean bfd_elf32_arm_get_bfd_for_interworking
    (bfd *, struct bfd_link_info *);
--- 816,822 ----
    (bfd *, struct bfd_link_info *, int);
  
  void bfd_elf32_arm_set_target_relocs
!   (struct bfd_link_info *, int, char *, int, int);
  
  extern bfd_boolean bfd_elf32_arm_get_bfd_for_interworking
    (bfd *, struct bfd_link_info *);
Index: bfd/elf32-arm.c
===================================================================
RCS file: /cvs/src/src/bfd/elf32-arm.c,v
retrieving revision 1.31
diff -c -p -r1.31 elf32-arm.c
*** bfd/elf32-arm.c	8 Apr 2005 11:47:59 -0000	1.31
--- bfd/elf32-arm.c	12 Apr 2005 20:57:45 -0000
*************** struct elf32_arm_link_hash_table
*** 1324,1329 ****
--- 1324,1332 ----
      /* Nonzero to fix BX instructions for ARMv4 targets.  */
      int fix_v4bx;
  
+     /* Nonzero if the ARM/Thumb BLX instructions are available for use.  */
+     int use_blx;
+ 
      /* The number of bytes in the initial entry in the PLT.  */
      bfd_size_type plt_header_size;
  
*************** elf32_arm_link_hash_table_create (bfd *a
*** 1559,1564 ****
--- 1562,1569 ----
    ret->plt_header_size = 20;
    ret->plt_entry_size = 12;
  #endif
+   ret->fix_v4bx = 0;
+   ret->use_blx = 0;
    ret->symbian_p = 0;
    ret->use_rel = 1;
    ret->sym_sec.abfd = NULL;
*************** void
*** 2125,2131 ****
  bfd_elf32_arm_set_target_relocs (struct bfd_link_info *link_info,
  				 int target1_is_rel,
  				 char * target2_type,
!                                  int fix_v4bx)
  {
    struct elf32_arm_link_hash_table *globals;
  
--- 2130,2137 ----
  bfd_elf32_arm_set_target_relocs (struct bfd_link_info *link_info,
  				 int target1_is_rel,
  				 char * target2_type,
!                                  int fix_v4bx,
! 				 int use_blx)
  {
    struct elf32_arm_link_hash_table *globals;
  
*************** bfd_elf32_arm_set_target_relocs (struct 
*** 2144,2149 ****
--- 2150,2156 ----
  			  target2_type);
      }
    globals->fix_v4bx = fix_v4bx;
+   globals->use_blx |= use_blx;
  }
  #endif
  
*************** elf32_arm_final_link_relocate (reloc_how
*** 2893,2898 ****
--- 2900,2906 ----
  	bfd_signed_vma reloc_signed_min = ~ reloc_signed_max;
  	bfd_vma check;
  	bfd_signed_vma signed_check;
+ 	bfd_boolean thumb_plt_call = FALSE;
  
  	/* Need to refetch the addend and squish the two 11 bit pieces
  	   together.  */
*************** elf32_arm_final_link_relocate (reloc_how
*** 2942,2949 ****
  	    value = (splt->output_section->vma
  		     + splt->output_offset
  		     + h->plt.offset);
! 	    /* Target the Thumb stub before the ARM PLT entry.  */
! 	    value -= 4;
  	    *unresolved_reloc_p = FALSE;
  	  }
  
--- 2950,2968 ----
  	    value = (splt->output_section->vma
  		     + splt->output_offset
  		     + h->plt.offset);
!  	    if (globals->use_blx)
!  	      {
!  		/* If the Thumb BLX instruction is available, convert the
! 		   BL to a BLX instruction to call the ARM-mode PLT entry.  */
!  		if ((lower_insn & (0x3 << 11)) == 0x3 << 11)
! 		  {
! 		    lower_insn = (lower_insn & ~(0x3 << 11)) | 0x1 << 11;
! 		    thumb_plt_call = TRUE;
! 		  }
!  	      }
!  	    else
!  	      /* Target the Thumb stub before the ARM PLT entry.  */
!  	      value -= PLT_THUMB_STUB_SIZE;
  	    *unresolved_reloc_p = FALSE;
  	  }
  
*************** elf32_arm_final_link_relocate (reloc_how
*** 2967,2974 ****
  	  overflow = TRUE;
  
  #ifndef OLD_ARM_ABI
! 	if (r_type == R_ARM_THM_XPC22
! 	    && ((lower_insn & 0x1800) == 0x0800))
  	  /* For a BLX instruction, make sure that the relocation is rounded up
  	     to a word boundary.  This follows the semantics of the instruction
  	     which specifies that bit 1 of the target address will come from bit
--- 2986,2994 ----
  	  overflow = TRUE;
  
  #ifndef OLD_ARM_ABI
! 	if ((r_type == R_ARM_THM_XPC22
! 	     && ((lower_insn & 0x1800) == 0x0800))
! 	    || thumb_plt_call)
  	  /* For a BLX instruction, make sure that the relocation is rounded up
  	     to a word boundary.  This follows the semantics of the instruction
  	     which specifies that bit 1 of the target address will come from bit
*************** allocate_dynrelocs (struct elf_link_hash
*** 5059,5065 ****
  
  	  /* If we will insert a Thumb trampoline before this PLT, leave room
  	     for it.  */
! 	  if (!htab->symbian_p && eh->plt_thumb_refcount > 0)
  	    {
  	      h->plt.offset += PLT_THUMB_STUB_SIZE;
  	      s->size += PLT_THUMB_STUB_SIZE;
--- 5079,5085 ----
  
  	  /* If we will insert a Thumb trampoline before this PLT, leave room
  	     for it.  */
! 	  if (!htab->use_blx && eh->plt_thumb_refcount > 0)
  	    {
  	      h->plt.offset += PLT_THUMB_STUB_SIZE;
  	      s->size += PLT_THUMB_STUB_SIZE;
*************** elf32_arm_finish_dynamic_symbol (bfd * o
*** 5632,5638 ****
  
  	  BFD_ASSERT ((got_displacement & 0xf0000000) == 0);
  
! 	  if (eh->plt_thumb_refcount > 0)
  	    {
  	      bfd_put_16 (output_bfd, elf32_arm_plt_thumb_stub[0],
  			  splt->contents + h->plt.offset - 4);
--- 5652,5658 ----
  
  	  BFD_ASSERT ((got_displacement & 0xf0000000) == 0);
  
! 	  if (!htab->use_blx && eh->plt_thumb_refcount > 0)
  	    {
  	      bfd_put_16 (output_bfd, elf32_arm_plt_thumb_stub[0],
  			  splt->contents + h->plt.offset - 4);
*************** elf32_arm_symbian_link_hash_table_create
*** 6505,6510 ****
--- 6525,6532 ----
        /* The PLT entries are each three instructions.  */
        htab->plt_entry_size = 4 * NUM_ELEM (elf32_arm_symbian_plt_entry);
        htab->symbian_p = 1;
+       /* Symbian uses armv5t or above, so use_blx is always true.  */
+       htab->use_blx = 1;
        htab->root.is_relocatable_executable = 1;
      }
    return ret;
Index: ld/ld.texinfo
===================================================================
RCS file: /cvs/src/src/ld/ld.texinfo,v
retrieving revision 1.140
diff -c -p -r1.140 ld.texinfo
*** ld/ld.texinfo	1 Feb 2005 17:31:01 -0000	1.140
--- ld/ld.texinfo	12 Apr 2005 20:57:45 -0000
*************** linker, which causes v4t @code{BX rM} in
*** 5262,5267 ****
--- 5262,5278 ----
  In the former case, the switch should not be used, and @samp{R_ARM_V4BX}
  relocations are ignored.
  
+ @cindex USE_BLX
+ @kindex --use-blx
+ The @samp{--use-blx} switch enables the linker to use ARM/Thumb
+ BLX instructions (available on ARMv5t and above) in various
+ situations. Currently it is used to perform calls via the PLT from
+ Thumb code rather than using BX and a mode-switching stub before each
+ PLT entry. This should lead to such calls executing slightly faster.
+ 
+ This option is enabled implicitly for SymbianOS, so there is no need to
+ specify it if you are using that target.
+ 
  @ifclear GENERIC
  @lowersections
  @end ifclear
Index: ld/emultempl/armelf.em
===================================================================
RCS file: /cvs/src/src/ld/emultempl/armelf.em,v
retrieving revision 1.44
diff -c -p -r1.44 armelf.em
*** ld/emultempl/armelf.em	3 Mar 2005 11:52:04 -0000	1.44
--- ld/emultempl/armelf.em	12 Apr 2005 20:57:45 -0000
*************** static int byteswap_code = 0;
*** 32,37 ****
--- 32,38 ----
  static int target1_is_rel = 0${TARGET1_IS_REL};
  static char *target2_type = "${TARGET2_TYPE}";
  static int fix_v4bx = 0;
+ static int use_blx = 0;
  
  static void
  gld${EMULATION_NAME}_before_parse (void)
*************** static void
*** 192,198 ****
  arm_elf_create_output_section_statements (void)
  {
    bfd_elf32_arm_set_target_relocs (&link_info, target1_is_rel, target2_type,
!                                    fix_v4bx);
  }
  
  EOF
--- 193,199 ----
  arm_elf_create_output_section_statements (void)
  {
    bfd_elf32_arm_set_target_relocs (&link_info, target1_is_rel, target2_type,
!                                    fix_v4bx, use_blx);
  }
  
  EOF
*************** PARSE_AND_LIST_PROLOGUE='
*** 206,212 ****
  #define OPTION_TARGET1_REL		303
  #define OPTION_TARGET1_ABS		304
  #define OPTION_TARGET2			305
! #define OPTION_FIX_V4BX                 306
  '
  
  PARSE_AND_LIST_SHORTOPTS=p
--- 207,214 ----
  #define OPTION_TARGET1_REL		303
  #define OPTION_TARGET1_ABS		304
  #define OPTION_TARGET2			305
! #define OPTION_FIX_V4BX			306
! #define OPTION_USE_BLX			307
  '
  
  PARSE_AND_LIST_SHORTOPTS=p
*************** PARSE_AND_LIST_LONGOPTS='
*** 219,224 ****
--- 221,227 ----
    { "target1-abs", no_argument, NULL, OPTION_TARGET1_ABS},
    { "target2", required_argument, NULL, OPTION_TARGET2},
    { "fix-v4bx", no_argument, NULL, OPTION_FIX_V4BX},
+   { "use-blx", no_argument, NULL, OPTION_USE_BLX},
  '
  
  PARSE_AND_LIST_OPTIONS='
*************** PARSE_AND_LIST_OPTIONS='
*** 228,233 ****
--- 231,237 ----
    fprintf (file, _("     --target1=abs            Interpret R_ARM_TARGET1 as R_ARM_ABS32\n"));
    fprintf (file, _("     --target2=<type>         Specify definition of R_ARM_TARGET2\n"));
    fprintf (file, _("     --fix-v4bx               Rewrite BX rn as MOV pc, rn for ARMv4\n"));
+   fprintf (file, _("     --use-blx                Enable use of BLX instructions\n"));
  '
  
  PARSE_AND_LIST_ARGS_CASES='
*************** PARSE_AND_LIST_ARGS_CASES='
*** 258,263 ****
--- 262,271 ----
      case OPTION_FIX_V4BX:
        fix_v4bx = 1;
        break;
+ 
+     case OPTION_USE_BLX:
+       use_blx = 1;
+       break;
  '
  
  # We have our own after_open and before_allocation functions, but they call