public inbox for binutils@sourceware.org
 help / color / mirror / Atom feed
From: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
To: Lulu Cai <cailulu@loongson.cn>
Cc: binutils@sourceware.org, xuchenghua@loongson.cn,
	chenglulu@loongson.cn, liuzhensong@loongson.cn,
	mengqinggang@loongson.cn, xry111@xry111.site,
	i.swmail@xen0n.name, maskray@google.com, luweining@loongson.cn,
	wanglei@loongson.cn, hejinyang@loongson.cn
Subject: Re: [PATCH v5 2/5] LoongArch: Add support for TLSDESC in ld.
Date: Thu, 28 Dec 2023 23:54:52 +0900	[thread overview]
Message-ID: <7DE21EC8-207D-41AA-B82E-0950D482F07B@gmail.com> (raw)
In-Reply-To: <20231222114243.1836112-3-cailulu@loongson.cn>

Hi, author of RISC-V TLSDESC patches here. I recently learned that LoongArch is working on a similar TLSDESC extension too. I’ll leave some of my thoughts on the patch set.

> On Dec 22, 2023, at 20:42, Lulu Cai <cailulu@loongson.cn> wrote:
> 
> 1.The linker for each DESC generates a R_LARCH_TLS_DESC64 dynamic
>  relocation, which relocation is placed at .rela.dyn.
>  TLSDESC always allocates two GOT slots and one dynamic relocation
>  space to TLSDESC.
> 2. When using multiple ways to access the same TLS variable, a
>   maximum of 5 GOT slots are used. For example, using GD, TLSDESC,
>   and IE to access the same TLS variable, GD always uses the first
>   two of the five GOT, TLSDESC uses the third and fourth, and IE
>   uses the last.
> ---
> bfd/elfnn-loongarch.c | 168 ++++++++++++++++++++++++++++++++++++------
> 1 file changed, 146 insertions(+), 22 deletions(-)
> 
> diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
> index faffe27294a..95a39148f73 100644
> --- a/bfd/elfnn-loongarch.c
> +++ b/bfd/elfnn-loongarch.c
> @@ -48,6 +48,12 @@ struct loongarch_elf_link_hash_entry
> #define GOT_TLS_GD  2
> #define GOT_TLS_IE  4
> #define GOT_TLS_LE  8
> +#define GOT_TLS_GDESC 16
> +
> +#define GOT_TLS_GD_BOTH_P(tls_type) \
> +  ((tls_type & GOT_TLS_GD) && (tls_type & GOT_TLS_GDESC))
> +#define GOT_TLS_GD_ANY_P(tls_type) \
> +  ((tls_type & GOT_TLS_GD) || (tls_type & GOT_TLS_GDESC))
>   char tls_type;
> };
> 
> @@ -563,6 +569,7 @@ loongarch_elf_record_tls_and_got_reference (bfd *abfd,
>     case GOT_NORMAL:
>     case GOT_TLS_GD:
>     case GOT_TLS_IE:
> +    case GOT_TLS_GDESC:
>       /* Need GOT.  */
>       if (htab->elf.sgot == NULL
> 	  && !loongarch_elf_create_got_section (htab->elf.dynobj, info))
> @@ -750,6 +757,14 @@ loongarch_elf_check_relocs (bfd *abfd, struct bfd_link_info *info,
> 	    return false;
> 	  break;
> 
> +	case R_LARCH_TLS_DESC_PC_HI20:
> +	case R_LARCH_TLS_DESC_HI20:
> +	  if (!loongarch_elf_record_tls_and_got_reference (abfd, info, h,
> +							   r_symndx,
> +							   GOT_TLS_GDESC))
> +	    return false;
> +	  break;
> +
> 	case R_LARCH_ABS_HI20:
> 	case R_LARCH_SOP_PUSH_ABSOLUTE:
> 	  if (h != NULL)
> @@ -1130,7 +1145,7 @@ allocate_dynrelocs (struct elf_link_hash_entry *h, void *inf)
> 
>       s = htab->elf.sgot;
>       h->got.offset = s->size;
> -      if (tls_type & (GOT_TLS_GD | GOT_TLS_IE))
> +      if (tls_type & (GOT_TLS_GD | GOT_TLS_IE | GOT_TLS_GDESC))
> 	{
> 	  /* TLS_GD needs two dynamic relocs and two GOT slots.  */
> 	  if (tls_type & GOT_TLS_GD)
> @@ -1167,7 +1182,15 @@ allocate_dynrelocs (struct elf_link_hash_entry *h, void *inf)
> 		  htab->elf.srelgot->size += sizeof (ElfNN_External_Rela);
> 		}
> 	    }
> +
> +	  /* TLS_DESC needs one dynamic reloc and two GOT slot.  */
> +	  if (tls_type & GOT_TLS_GDESC)
> +	    {
> +	      s->size += GOT_ENTRY_SIZE * 2;
> +	      htab->elf.srelgot->size += sizeof (ElfNN_External_Rela);
> +	    }
> 	}
> +
>       else
> 	{
> 	  s->size += GOT_ENTRY_SIZE;
> @@ -1670,19 +1693,34 @@ loongarch_elf_size_dynamic_sections (bfd *output_bfd,
> 	  if (0 < *local_got)
> 	    {
> 	      *local_got = s->size;
> +	      if (*local_tls_type & (GOT_TLS_GD | GOT_TLS_IE | GOT_TLS_GDESC))
> +		{
> +		  /* TLS gd use two got.  */
> +		  if (*local_tls_type & GOT_TLS_GD)
> +		    {
> +		      s->size += 2 * GOT_ENTRY_SIZE;
> +		      if (!bfd_link_executable (info))
> +			srel->size += sizeof (ElfNN_External_Rela);
> +		    }
> 
> -	      /* TLS gd use two got.  */
> -	      if (*local_tls_type & GOT_TLS_GD)
> -		s->size += GOT_ENTRY_SIZE * 2;
> -	      else
> -		/* Normal got, tls ie/ld use one got.  */
> -		s->size += GOT_ENTRY_SIZE;
> +		  /* TLS_DESC use two got.  */
> +		  if (*local_tls_type & GOT_TLS_GDESC)
> +		    {
> +		      s->size += 2 * GOT_ENTRY_SIZE;
> +		      srel->size += sizeof (ElfNN_External_Rela);
> +		    }
> 
> -	      if (bfd_link_executable (info)
> -		  && (*local_tls_type & (GOT_TLS_GD| GOT_TLS_IE)))
> -		;/* Do nothing.  */
> +		  /* TLS ie and use one got.  */
> +		  if (*local_tls_type & GOT_TLS_IE)
> +		    {
> +		      s->size += GOT_ENTRY_SIZE;
> +		      if (!bfd_link_executable (info))
> +			srel->size += sizeof (ElfNN_External_Rela);
> +		    }
> +		}
> 	      else
> 		{
> +		  s->size += GOT_ENTRY_SIZE;
> 		  srel->size += sizeof (ElfNN_External_Rela);
> 		}
> 	    }
> @@ -2126,6 +2164,15 @@ perform_relocation (const Elf_Internal_Rela *rel, asection *input_section,
>     case R_LARCH_TLS_GD_HI20:
>     case R_LARCH_PCREL20_S2:
>     case R_LARCH_CALL36:
> +    case R_LARCH_TLS_DESC_PC_HI20:
> +    case R_LARCH_TLS_DESC_PC_LO12:
> +    case R_LARCH_TLS_DESC64_PC_LO20:
> +    case R_LARCH_TLS_DESC64_PC_HI12:
> +    case R_LARCH_TLS_DESC_HI20:
> +    case R_LARCH_TLS_DESC_LO12:
> +    case R_LARCH_TLS_DESC64_LO20:
> +    case R_LARCH_TLS_DESC64_HI12:
> +
>       r = loongarch_check_offset (rel, input_section);
>       if (r != bfd_reloc_ok)
> 	break;
> @@ -2135,6 +2182,11 @@ perform_relocation (const Elf_Internal_Rela *rel, asection *input_section,
> 					    contents, value);
>       break;
> 
> +    case R_LARCH_TLS_DESC_LD:
> +    case R_LARCH_TLS_DESC_CALL:
> +      r = bfd_reloc_ok;
> +      break;
> +
>     case R_LARCH_RELAX:
>       break;
> 
> @@ -2383,10 +2435,10 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
>       struct elf_link_hash_entry *h = NULL;
>       const char *name;
>       bfd_reloc_status_type r = bfd_reloc_ok;
> -      bool is_ie, is_undefweak, unresolved_reloc, defined_local;
> +      bool is_ie, is_desc, is_undefweak, unresolved_reloc, defined_local;
>       bool resolved_local, resolved_dynly, resolved_to_const;
>       char tls_type;
> -      bfd_vma relocation, off, ie_off;
> +      bfd_vma relocation, off, ie_off, desc_off;
>       int i, j;
> 
>       howto = loongarch_elf_rtype_to_howto (input_bfd, r_type);
> @@ -2515,6 +2567,7 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 
>       BFD_ASSERT (!resolved_local || defined_local);
> 
> +      is_desc = false;
>       is_ie = false;
>       switch (r_type)
> 	{
> @@ -3405,6 +3458,8 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 	case R_LARCH_TLS_LD_HI20:
> 	case R_LARCH_TLS_GD_PC_HI20:
> 	case R_LARCH_TLS_GD_HI20:
> +	case R_LARCH_TLS_DESC_PC_HI20:
> +	case R_LARCH_TLS_DESC_HI20:
> 	  BFD_ASSERT (rel->r_addend == 0);
> 	  unresolved_reloc = false;
> 
> @@ -3412,6 +3467,10 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 	      || r_type == R_LARCH_TLS_IE_HI20)
> 	    is_ie = true;
> 
> +	  if (r_type == R_LARCH_TLS_DESC_PC_HI20
> +	      || r_type == R_LARCH_TLS_DESC_HI20)
> +	    is_desc = true;
> +

I see that the got_off assignment logic is duplicated for R_LARCH_SOP_PUSH_TLS_{GOT, GD}, but that path was not modified in this patch. Is this OK? Ideally, the logic should not be duplicated at all.

> 	  bfd_vma got_off = 0;
> 	  if (h != NULL)
> 	    {
> @@ -3426,9 +3485,19 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 
> 	  BFD_ASSERT (got_off != MINUS_ONE);
> 
> -	  ie_off = 0;
> 	  tls_type = _bfd_loongarch_elf_tls_type (input_bfd, h, r_symndx);
> -	  if ((tls_type & GOT_TLS_GD) && (tls_type & GOT_TLS_IE))
> +
> +	  /* If a tls variable is accessed in multiple ways, GD uses
> +	     the first two slots of GOT, desc follows with two slots,
> +	     and IE uses one slot at the end.  */
> +	  desc_off = 0;
> +	  if (GOT_TLS_GD_BOTH_P (tls_type))
> +	    desc_off = 2 * GOT_ENTRY_SIZE;
> +
> +	  ie_off = 0;
> +	  if (GOT_TLS_GD_BOTH_P (tls_type) && (tls_type & GOT_TLS_IE))
> +	    ie_off = 4 * GOT_ENTRY_SIZE;
> +	  else if (GOT_TLS_GD_ANY_P (tls_type) && (tls_type & GOT_TLS_IE))

I think the calculation should be done additively, instead of checking for every possible combination. That is,

off = 0;
if (tls gd)
  off += gd size;
desc_off = off;
if (tls ie)
  off += desc size;
ie_off = off;

would be much simpler.

> 	    ie_off = 2 * GOT_ENTRY_SIZE;
> 
> 	  if ((got_off & 1) == 0)
> @@ -3477,6 +3546,21 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 		      loongarch_elf_append_rela (output_bfd, relgot, &rela);
> 		    }
> 		}
> +	      if (tls_type & GOT_TLS_GDESC)
> +		{
> +		  /* Unless it is a static link, DESC always emits a
> +		     dynamic relocation.  */
> +		  int indx = h && h->dynindx != -1 ? h->dynindx : 0;
> +		  rela.r_offset = sec_addr (got) + got_off + desc_off;
> +		  rela.r_addend = 0;
> +		  if (indx == 0)
> +		    rela.r_addend = relocation - elf_hash_table (info)->tls_sec->vma;
> +
> +		  rela.r_info = ELFNN_R_INFO (indx, R_LARCH_TLS_DESCNN);
> +		  loongarch_elf_append_rela (output_bfd, relgot, &rela);
> +		  bfd_put_NN (output_bfd, 0,
> +			      got->contents + got_off + desc_off);
> +		}
> 	      if (tls_type & GOT_TLS_IE)
> 		{
> 		  rela.r_offset = sec_addr (got) + got_off + ie_off;
> @@ -3504,16 +3588,52 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 		    }
> 		}
> 	    }
> -	  relocation = (got_off & (~(bfd_vma)1)) + sec_addr (got)
> -			+ (is_ie ? ie_off : 0);
> +	  relocation = (got_off & (~(bfd_vma)1)) + sec_addr (got);
> +	  if (is_desc)
> +	    relocation += desc_off;
> +	  else if (is_ie)
> +	    relocation += ie_off;
> 
> 	  if (r_type == R_LARCH_TLS_LD_PC_HI20
> 	      || r_type == R_LARCH_TLS_GD_PC_HI20
> -	      || r_type == R_LARCH_TLS_IE_PC_HI20)
> +	      || r_type == R_LARCH_TLS_IE_PC_HI20
> +	      || r_type == R_LARCH_TLS_DESC_PC_HI20)
> 	    RELOCATE_CALC_PC32_HI20 (relocation, pc);
> 
> 	  break;
> 
> +	case R_LARCH_TLS_DESC_PC_LO12:
> +	case R_LARCH_TLS_DESC64_PC_LO20:
> +	case R_LARCH_TLS_DESC64_PC_HI12:
> +	case R_LARCH_TLS_DESC_LO12:
> +	case R_LARCH_TLS_DESC64_LO20:
> +	case R_LARCH_TLS_DESC64_HI12:
> +	  {
> +	    unresolved_reloc = false;
> +
> +	    if (h)
> +	      relocation = sec_addr (got) + (h->got.offset & (~(bfd_vma)1));
> +	    else
> +	      relocation = sec_addr (got)
> +			   + (local_got_offsets[r_symndx] & (~(bfd_vma)1));
> +
> +	    tls_type = _bfd_loongarch_elf_tls_type (input_bfd, h, r_symndx);
> +	    /* Use both TLS_GD and TLS_DESC.  */
> +	    if ((tls_type & GOT_TLS_GD) && (tls_type & GOT_TLS_GDESC))
> +	      relocation += 2 * GOT_ENTRY_SIZE;

For simplicity, I think the ie_off and desc_off calculation should be moved out of the switch. Then you can just reuse the desc/ie_off calculated there.

> +	  }
> +
> +	    if (r_type == R_LARCH_TLS_DESC64_PC_LO20
> +		|| r_type == R_LARCH_TLS_DESC64_PC_HI12)
> +	      RELOCATE_CALC_PC64_HI32 (relocation, pc);
> +
> +	    break;
> +
> +	case R_LARCH_TLS_DESC_LD:
> +	case R_LARCH_TLS_DESC_CALL:
> +	  unresolved_reloc = false;
> +	  break;
> +
> 	case R_LARCH_TLS_IE_PC_LO12:
> 	case R_LARCH_TLS_IE64_PC_LO20:
> 	case R_LARCH_TLS_IE64_PC_HI12:
> @@ -3523,14 +3643,17 @@ loongarch_elf_relocate_section (bfd *output_bfd, struct bfd_link_info *info,
> 	  unresolved_reloc = false;
> 
> 	  if (h)
> -	    relocation = sec_addr (got) + (h->got.offset & (~(bfd_vma)3));
> +	    relocation = sec_addr (got) + (h->got.offset & (~(bfd_vma)1));
> 	  else
> 	    relocation = sec_addr (got)
> -	      + (local_got_offsets[r_symndx] & (~(bfd_vma)3));
> +	      + (local_got_offsets[r_symndx] & (~(bfd_vma)1));
> 
> 	  tls_type = _bfd_loongarch_elf_tls_type (input_bfd, h, r_symndx);
> -	  /* Use both TLS_GD and TLS_IE.  */
> -	  if ((tls_type & GOT_TLS_GD) && (tls_type & GOT_TLS_IE))
> +	  /* Use TLS_GD TLS_DESC and TLS_IE.  */
> +	  if (GOT_TLS_GD_BOTH_P (tls_type) && (tls_type & GOT_TLS_IE))
> +	    relocation += 4 * GOT_ENTRY_SIZE;
> +	  /* Use GOT_TLS_GD_ANY_P (tls_type) and TLS_IE.  */
> +	  else if (GOT_TLS_GD_ANY_P (tls_type) && (tls_type & GOT_TLS_IE))
> 	    relocation += 2 * GOT_ENTRY_SIZE;
> 
> 	  if (r_type == R_LARCH_TLS_IE64_PC_LO20
> @@ -4174,7 +4297,8 @@ loongarch_elf_finish_dynamic_symbol (bfd *output_bfd,
> 
>   if (h->got.offset != MINUS_ONE
>       /* TLS got entry have been handled in elf_relocate_section.  */
> -      && !(loongarch_elf_hash_entry (h)->tls_type & (GOT_TLS_GD | GOT_TLS_IE))
> +      && !(loongarch_elf_hash_entry (h)->tls_type
> +	   & (GOT_TLS_GD | GOT_TLS_IE | GOT_TLS_GDESC))
>       /* Have allocated got entry but not allocated rela before.  */
>       && !UNDEFWEAK_NO_DYNAMIC_RELOC (info, h))
>     {
> -- 
> 2.43.0
> 
> 


  reply	other threads:[~2023-12-28 14:55 UTC|newest]

Thread overview: 19+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-12-22 11:42 [PATCH v5 0/5] Add support for TLS Descriptors (TLSDESC) Lulu Cai
2023-12-22 11:42 ` [PATCH v5 1/5] LoongArch: Add new relocs and macro for TLSDESC Lulu Cai
2023-12-22 11:42 ` [PATCH v5 2/5] LoongArch: Add support for TLSDESC in ld Lulu Cai
2023-12-28 14:54   ` Tatsuyuki Ishi [this message]
2023-12-29 10:24     ` Lulu Cai
2023-12-22 11:42 ` [PATCH v5 3/5] LoongArch: Add tls transition support Lulu Cai
2023-12-28 14:42   ` Tatsuyuki Ishi
2023-12-29 10:31     ` Lulu Cai
2023-12-22 11:42 ` [PATCH v5 4/5] LoongArch: Add support for TLS LD/GD/DESC relaxation Lulu Cai
2023-12-28 14:38   ` Tatsuyuki Ishi
2023-12-29 10:36     ` Lulu Cai
2023-12-29 10:45     ` Lulu Cai
2024-01-07 23:00       ` Tatsuyuki Ishi
2024-01-08  0:22         ` Fangrui Song
     [not found]         ` <DS7PR12MB5765D250CD96B4F6EF6674BACB6B2@DS7PR12MB5765.namprd12.prod.outlook.com>
2024-01-08  6:00           ` Lulu Cai
2024-01-08  7:59             ` Tatsuyuki Ishi
2024-01-27  2:54 ` [PATCH v5 0/5] Add support for TLS Descriptors (TLSDESC) Fangrui Song
     [not found] ` <DS7PR12MB5765A6FBD6297BAEDB57FF6CCB782@DS7PR12MB5765.namprd12.prod.outlook.com>
2024-01-27  7:26   ` Lulu Cai
2024-01-27 19:18     ` Fangrui Song

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=7DE21EC8-207D-41AA-B82E-0950D482F07B@gmail.com \
    --to=ishitatsuyuki@gmail.com \
    --cc=binutils@sourceware.org \
    --cc=cailulu@loongson.cn \
    --cc=chenglulu@loongson.cn \
    --cc=hejinyang@loongson.cn \
    --cc=i.swmail@xen0n.name \
    --cc=liuzhensong@loongson.cn \
    --cc=luweining@loongson.cn \
    --cc=maskray@google.com \
    --cc=mengqinggang@loongson.cn \
    --cc=wanglei@loongson.cn \
    --cc=xry111@xry111.site \
    --cc=xuchenghua@loongson.cn \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).