public inbox for binutils@sourceware.org
 help / color / mirror / Atom feed
From: Nelson Chu <nelson@rivosinc.com>
To: binutils@sourceware.org, jim.wilson.gcc@gmail.com,
	palmer@dabbelt.com, lifang_xia@linux.alibaba.com
Cc: nelson@rivosinc.com
Subject: [PATCH] RISC-V: Optimize relaxation of gp with max_alignment.
Date: Mon, 17 Apr 2023 17:54:43 +0800	[thread overview]
Message-ID: <20230417095443.73581-1-nelson@rivosinc.com> (raw)

From: Lifang Xia <lifang_xia@linux.alibaba.com>

This should be the first related issue, which posted in riscv-gnu-toolchain,
https://github.com/riscv-collab/riscv-gnu-toolchain/issues/497

If the output sections are not between gp and the symbol, then their alignments
shouldn't affect the gp relaxation.  However, this patch improves this idea
even more, it limits the range to the gp+-2k, which means only the output
section which are in the [gp-2K, gp+2K) range need to be considered.

Even if the output section candidates may be different for each relax passes,
the symbol that can be relaxed ar this round will not be truncated at next
round.  That is because this round you can do relax to indicate that the
section where the symbol is located is within the [gp-2K, gp+2K) range, so all
the output section alignments between them should be considered.  In other
words, if the alignments between them may cause truncated, then we should
already preserve the size and won't do the gp relaxation this time.

This patch can resolve the github issue which mentioned above, and also passed
all gcc/binutils regressions of riscv-gnu-toolchain, so should be worth and
safe enough to commit.

Originally, this patch also do the same optimization for the call relaxations,
https://sourceware.org/pipermail/binutils/2022-October/123918.html
But just in case there is something that has not been considered, we only
deal with the gp relaxation at this time.

bfd/
	* elfnn-riscv.c (riscv_elf_link_hash_table): Added new bfd_vma,
	max_alignment_for_gp.  It is used to record the maximum alignment of
	the output sections, which are in the [gp-2K, gp+2k) range.
	(riscv_elf_link_hash_table_create): Init max_alignment_for_gp to -1.
	(_bfd_riscv_get_max_alignment): Added new parameter, gp.  If gp is
	zero, then all the output section alignments are possible candidates;
	Otherwise, only the output sections which are in the [gp-2K, gp+2K)
	range need to be considered.
	(_bfd_riscv_relax_lui): Called _bfd_riscv_get_max_alignment with the
	non-zero gp if the max_alignment_for_gp is -1.
	(_bfd_riscv_relax_pc): Likewise.
	(_bfd_riscv_relax_section): Record the first input section, so that
	we can reset the max_alignment_for_gp for each repeated relax passes.
ld/
	* testsuite/ld-riscv-elf/ld-riscv-elf.exp: Updated.
	* testsuite/ld-riscv-elf/relax-max-align-gp.*: New testcase.  It fails
	without this patch.

Co-developed-by: Nelson Chu <nelson@rivosinc.com>
---
 bfd/elfnn-riscv.c                             | 103 +++++++++++++-----
 ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp    |   1 +
 .../ld-riscv-elf/relax-max-align-gp.d         |  46 ++++++++
 .../ld-riscv-elf/relax-max-align-gp.s         |  28 +++++
 4 files changed, 149 insertions(+), 29 deletions(-)
 create mode 100644 ld/testsuite/ld-riscv-elf/relax-max-align-gp.d
 create mode 100644 ld/testsuite/ld-riscv-elf/relax-max-align-gp.s

diff --git a/bfd/elfnn-riscv.c b/bfd/elfnn-riscv.c
index 0dd9b27c8ae..e90e36b58bb 100644
--- a/bfd/elfnn-riscv.c
+++ b/bfd/elfnn-riscv.c
@@ -200,6 +200,9 @@ struct riscv_elf_link_hash_table
   /* The max alignment of output sections.  */
   bfd_vma max_alignment;
 
+  /* The max alignment of output sections in [gp-2K, gp+2K) range.  */
+  bfd_vma max_alignment_for_gp;
+
   /* Used by local STT_GNU_IFUNC symbols.  */
   htab_t loc_hash_table;
   void * loc_hash_memory;
@@ -488,6 +491,7 @@ riscv_elf_link_hash_table_create (bfd *abfd)
     }
 
   ret->max_alignment = (bfd_vma) -1;
+  ret->max_alignment_for_gp = (bfd_vma) -1;
 
   /* Create hash table for local ifunc.  */
   ret->loc_hash_table = htab_try_create (1024,
@@ -4460,17 +4464,27 @@ _bfd_riscv_relax_call (bfd *abfd, asection *sec, asection *sym_sec,
 				   link_info, pcgp_relocs, rel + 1);
 }
 
-/* Traverse all output sections and return the max alignment.  */
+/* Traverse all output sections and return the max alignment.
+
+   If gp is zero, then all the output section alignments are
+   possible candidates;  Otherwise, only the output sections
+   which are in the [gp-2K, gp+2K) range need to be considered.  */
 
 static bfd_vma
-_bfd_riscv_get_max_alignment (asection *sec)
+_bfd_riscv_get_max_alignment (asection *sec, bfd_vma gp)
 {
   unsigned int max_alignment_power = 0;
   asection *o;
 
   for (o = sec->output_section->owner->sections; o != NULL; o = o->next)
     {
-      if (o->alignment_power > max_alignment_power)
+      bool valid = true;
+      if (gp
+	  && !(VALID_ITYPE_IMM (sec_addr (o) - gp)
+	       || VALID_ITYPE_IMM (sec_addr (o) + o->size - gp)))
+	valid = false;
+
+      if (valid && o->alignment_power > max_alignment_power)
 	max_alignment_power = o->alignment_power;
     }
 
@@ -4492,15 +4506,16 @@ _bfd_riscv_relax_lui (bfd *abfd,
 		      riscv_pcgp_relocs *pcgp_relocs,
 		      bool undefined_weak)
 {
+  struct riscv_elf_link_hash_table *htab = riscv_elf_hash_table (link_info);
   bfd_byte *contents = elf_section_data (sec)->this_hdr.contents;
-  bfd_vma gp = riscv_elf_hash_table (link_info)->params->relax_gp
-		   ? riscv_global_pointer_value (link_info)
-		   : 0;
+  bfd_vma gp = htab->params->relax_gp
+	       ? riscv_global_pointer_value (link_info)
+	       : 0;
   int use_rvc = elf_elfheader (abfd)->e_flags & EF_RISCV_RVC;
 
   BFD_ASSERT (rel->r_offset + 4 <= sec->size);
 
-  if (gp)
+  if (!undefined_weak && gp)
     {
       /* If gp and the symbol are in the same output section, which is not the
 	 abs section, then consider only that output section's alignment.  */
@@ -4510,16 +4525,28 @@ _bfd_riscv_relax_lui (bfd *abfd,
       if (h->u.def.section->output_section == sym_sec->output_section
 	  && sym_sec->output_section != bfd_abs_section_ptr)
 	max_alignment = (bfd_vma) 1 << sym_sec->output_section->alignment_power;
+      else
+	{
+	  /* Consider output section alignments which are in [gp-2K, gp+2K). */
+	  max_alignment = htab->max_alignment_for_gp;
+	  if (max_alignment == (bfd_vma) -1)
+	    {
+	      max_alignment = _bfd_riscv_get_max_alignment (sec, gp);
+	      htab->max_alignment_for_gp = max_alignment;
+	    }
+	}
     }
 
   /* Is the reference in range of x0 or gp?
-     Valid gp range conservatively because of alignment issue.  */
+     Valid gp range conservatively because of alignment issue.
+
+     Should we also consider the alignment issue for x0 base?  */
   if (undefined_weak
-      || (VALID_ITYPE_IMM (symval)
-	  || (symval >= gp
-	      && VALID_ITYPE_IMM (symval - gp + max_alignment + reserve_size))
-	  || (symval < gp
-	      && VALID_ITYPE_IMM (symval - gp - max_alignment - reserve_size))))
+      || VALID_ITYPE_IMM (symval)
+      || (symval >= gp
+	  && VALID_ITYPE_IMM (symval - gp + max_alignment + reserve_size))
+      || (symval < gp
+	  && VALID_ITYPE_IMM (symval - gp - max_alignment - reserve_size)))
     {
       unsigned sym = ELFNN_R_SYM (rel->r_info);
       switch (ELFNN_R_TYPE (rel->r_info))
@@ -4709,6 +4736,7 @@ _bfd_riscv_relax_pc (bfd *abfd ATTRIBUTE_UNUSED,
 		     riscv_pcgp_relocs *pcgp_relocs,
 		     bool undefined_weak)
 {
+  struct riscv_elf_link_hash_table *htab = riscv_elf_hash_table (link_info);
   bfd_byte *contents = elf_section_data (sec)->this_hdr.contents;
   bfd_vma gp = riscv_global_pointer_value (link_info);
 
@@ -4765,7 +4793,7 @@ _bfd_riscv_relax_pc (bfd *abfd ATTRIBUTE_UNUSED,
       abort ();
     }
 
-  if (gp)
+  if (!undefined_weak && gp)
     {
       /* If gp and the symbol are in the same output section, which is not the
 	 abs section, then consider only that output section's alignment.  */
@@ -4775,16 +4803,28 @@ _bfd_riscv_relax_pc (bfd *abfd ATTRIBUTE_UNUSED,
       if (h->u.def.section->output_section == sym_sec->output_section
 	  && sym_sec->output_section != bfd_abs_section_ptr)
 	max_alignment = (bfd_vma) 1 << sym_sec->output_section->alignment_power;
+      else
+	{
+	  /* Consider output section alignments which are in [gp-2K, gp+2K). */
+	  max_alignment = htab->max_alignment_for_gp;
+	  if (max_alignment == (bfd_vma) -1)
+	    {
+	      max_alignment = _bfd_riscv_get_max_alignment (sec, gp);
+	      htab->max_alignment_for_gp = max_alignment;
+	    }
+	}
     }
 
   /* Is the reference in range of x0 or gp?
-     Valid gp range conservatively because of alignment issue.  */
+     Valid gp range conservatively because of alignment issue.
+
+     Should we also consider the alignment issue for x0 base?  */
   if (undefined_weak
-      || (VALID_ITYPE_IMM (symval)
-	  || (symval >= gp
-	      && VALID_ITYPE_IMM (symval - gp + max_alignment + reserve_size))
-	  || (symval < gp
-	      && VALID_ITYPE_IMM (symval - gp - max_alignment - reserve_size))))
+      || VALID_ITYPE_IMM (symval)
+      || (symval >= gp
+	  && VALID_ITYPE_IMM (symval - gp + max_alignment + reserve_size))
+      || (symval < gp
+	  && VALID_ITYPE_IMM (symval - gp - max_alignment - reserve_size)))
     {
       unsigned sym = hi_reloc.hi_sym;
       switch (ELFNN_R_TYPE (rel->r_info))
@@ -4877,6 +4917,7 @@ _bfd_riscv_relax_section (bfd *abfd, asection *sec,
   unsigned int i;
   bfd_vma max_alignment, reserve_size = 0;
   riscv_pcgp_relocs pcgp_relocs;
+  static asection *first_section = NULL;
 
   *again = false;
 
@@ -4892,6 +4933,13 @@ _bfd_riscv_relax_section (bfd *abfd, asection *sec,
       || *(htab->data_segment_phase) == 4)
     return true;
 
+  /* Record the first relax section, so that we can reset the
+     max_alignment_for_gp for the repeated relax passes.  */
+  if (first_section == NULL)
+    first_section = sec;
+  else if (first_section == sec)
+    htab->max_alignment_for_gp = -1;
+
   riscv_init_pcgp_relocs (&pcgp_relocs);
 
   /* Read this BFD's relocs if we haven't done so already.  */
@@ -4901,17 +4949,14 @@ _bfd_riscv_relax_section (bfd *abfd, asection *sec,
 						 info->keep_memory)))
     goto fail;
 
-  if (htab)
+  /* Estimate the maximum alignment for all output sections once time
+     should be enough.  */
+  max_alignment = htab->max_alignment;
+  if (max_alignment == (bfd_vma) -1)
     {
-      max_alignment = htab->max_alignment;
-      if (max_alignment == (bfd_vma) -1)
-	{
-	  max_alignment = _bfd_riscv_get_max_alignment (sec);
-	  htab->max_alignment = max_alignment;
-	}
+      max_alignment = _bfd_riscv_get_max_alignment (sec, 0/* gp */);
+      htab->max_alignment = max_alignment;
     }
-  else
-    max_alignment = _bfd_riscv_get_max_alignment (sec);
 
   /* Examine and consider relaxing each reloc.  */
   for (i = 0; i < sec->reloc_count; i++)
diff --git a/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp b/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
index 43572c5286b..9e103b283f7 100644
--- a/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
+++ b/ld/testsuite/ld-riscv-elf/ld-riscv-elf.exp
@@ -171,6 +171,7 @@ if [istarget "riscv*-*-*"] {
     run_dump_test "attr-merge-priv-spec-failed-05"
     run_dump_test "attr-merge-priv-spec-failed-06"
     run_dump_test "attr-phdr"
+    run_dump_test "relax-max-align-gp"
     run_ld_link_tests [list \
 	[list "Weak reference 32" "-T weakref.ld -m[riscv_choose_ilp32_emul]" "" \
 	    "-march=rv32i -mabi=ilp32" {weakref32.s} \
diff --git a/ld/testsuite/ld-riscv-elf/relax-max-align-gp.d b/ld/testsuite/ld-riscv-elf/relax-max-align-gp.d
new file mode 100644
index 00000000000..637de426ee4
--- /dev/null
+++ b/ld/testsuite/ld-riscv-elf/relax-max-align-gp.d
@@ -0,0 +1,46 @@
+#source: relax-max-align-gp.s
+#ld:
+#objdump: -d
+
+.*:[ 	]+file format .*
+
+
+Disassembly of section .text:
+
+0+[0-9a-f]+ <_start>:
+.*:[ 	]+[0-9a-f]+[ 	]+add[ 	]+.*<gdata>
+.*:[ 	]+[0-9a-f]+[ 	]+jal[ 	]+.*
+.*:[ 	]+[0-9a-f]+[ 	]+j[ 	]+.*
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+.*:[ 	]+[0-9a-f]+[ 	]+nop
+
+0+[0-9a-f]+ <func>:
+.*:[ 	]+[0-9a-f]+[ 	]+ret
+[ 	]+...
diff --git a/ld/testsuite/ld-riscv-elf/relax-max-align-gp.s b/ld/testsuite/ld-riscv-elf/relax-max-align-gp.s
new file mode 100644
index 00000000000..ce3da21e7f4
--- /dev/null
+++ b/ld/testsuite/ld-riscv-elf/relax-max-align-gp.s
@@ -0,0 +1,28 @@
+
+.global _start
+_start:
+	lui   a0, %hi(gdata)
+	addi  a0, a0, %lo(gdata)
+	call func
+	j .
+	.size _start, . - _start
+
+.global func
+.align 7
+func:
+	ret
+	.size func, . - func
+
+.data
+padding:
+	.long 0
+	.long 0
+	.long 0
+	.long 0
+	.size padding, . - padding
+
+.global gdata
+.type gdata, object
+gdata:
+	.zero 4
+	.size gdata, . - gdata
-- 
2.39.2 (Apple Git-143)


             reply	other threads:[~2023-04-17  9:54 UTC|newest]

Thread overview: 2+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2023-04-17  9:54 Nelson Chu [this message]
2023-04-21  7:50 ` Nelson Chu

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=20230417095443.73581-1-nelson@rivosinc.com \
    --to=nelson@rivosinc.com \
    --cc=binutils@sourceware.org \
    --cc=jim.wilson.gcc@gmail.com \
    --cc=lifang_xia@linux.alibaba.com \
    --cc=palmer@dabbelt.com \
    /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).