[PATCH v2 0/6] Fix some bugs of relaxation

[PATCH v2 0/6] Fix some bugs of relaxation

[mengqinggang]

These patches mainly fix the shared objects can't be relaxed bug.
The "elf_seg_map (info->output_bfd) == NULL" condition prevent the
shared object to be relaxed. Ld can perform up to 10 relaxation tries.
In first try, both execute file and shared object do not meet this condition.
But the execute file can trigger a second try because of the program header
size change due to the number of segement change (execute file add 
extra INTERP segemnt).

Changelog:

v1 -> v2:
  1. Delete useless .fill 0x4000.
  2. Change Fixme to FIXME.
  3. Modify some comments.

mengqinggang (6):
  LoongArch: Fix ld --no-relax bug
  LoongArch: Directly delete relaxed instuctions in first relaxation
    pass
  LoongArch: Multiple relax_trip in one relax_pass
  LoongArch: Remove "elf_seg_map (info->output_bfd) == NULL" relaxation
    condition
  LoongArch: Modify link_info.relax_pass from 3 to 2
  LoongArch: Add more relaxation testcases

 bfd/elfnn-loongarch.c                         | 54 ++++++++-----
 ld/emultempl/loongarchelf.em                  | 20 ++++-
 .../ld-loongarch-elf/relax-segment-max.s      | 12 +++
 .../ld-loongarch-elf/relax-segment-min.s      | 12 +++
 ld/testsuite/ld-loongarch-elf/relax-so.s      |  4 +
 ld/testsuite/ld-loongarch-elf/relax.exp       | 80 +++++++++++++++++--
 6 files changed, 157 insertions(+), 25 deletions(-)
 create mode 100644 ld/testsuite/ld-loongarch-elf/relax-segment-max.s
 create mode 100644 ld/testsuite/ld-loongarch-elf/relax-segment-min.s
 create mode 100644 ld/testsuite/ld-loongarch-elf/relax-so.s

-- 
2.36.0

[PATCH v2 1/6] LoongArch: Fix ld --no-relax bug

[mengqinggang]

When calling ld with --no-relax, pcalau12i + ld.d still can be relaxed.
This patch fix this bug and pcalau12i + ld.d can be relaxed with --relax.
---
 bfd/elfnn-loongarch.c | 19 ++++++-------------
 1 file changed, 6 insertions(+), 13 deletions(-)

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index 18ad3cc91ca..1162cb35cd6 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -3996,29 +3996,22 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
 	    loongarch_relax_align (abfd, sec, sym_sec, info, rel, symval);
 	  break;
 	case R_LARCH_DELETE:
-	  if (info->relax_pass == 1)
+	  if (1 == info->relax_pass)
 	    {
 	      loongarch_relax_delete_bytes (abfd, sec, rel->r_offset, 4, info);
 	      rel->r_info = ELFNN_R_INFO (0, R_LARCH_NONE);
 	    }
 	  break;
 	case R_LARCH_PCALA_HI20:
-	  if (info->relax_pass == 0)
-	    {
-	      if (i + 4 > sec->reloc_count)
-		break;
-	      loongarch_relax_pcala_addi (abfd, sec, rel, symval);
-	    }
+	  if (0 == info->relax_pass && (i + 4) <= sec->reloc_count)
+	    loongarch_relax_pcala_addi (abfd, sec, rel, symval);
 	  break;
 	case R_LARCH_GOT_PC_HI20:
-	  if (local_got)
+	  if (local_got && 0 == info->relax_pass
+	      && (i + 4) <= sec->reloc_count)
 	    {
-	      if (i + 4 > sec->reloc_count)
-		break;
 	      if (loongarch_relax_pcala_ld (abfd, sec, rel))
-		{
-		  loongarch_relax_pcala_addi (abfd, sec, rel, symval);
-		}
+		loongarch_relax_pcala_addi (abfd, sec, rel, symval);
 	    }
 	  break;
 	default:
-- 
2.36.0

[PATCH v2 2/6] LoongArch: Directly delete relaxed instuctions in first relaxation pass

[mengqinggang]

Directly delete relaxed instuctions in first relaxation pass, not use
R_LARCH_DELETE relocation. If not, the PC-relative offset may increase.
---
 bfd/elfnn-loongarch.c | 12 +++++++-----
 1 file changed, 7 insertions(+), 5 deletions(-)

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index 1162cb35cd6..9d4ea7e4ae7 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -3739,7 +3739,8 @@ loongarch_relax_delete_bytes (bfd *abfd,
 /* Relax pcalau12i,addi.d => pcaddi.  */
 static bool
 loongarch_relax_pcala_addi (bfd *abfd, asection *sec,
-		       Elf_Internal_Rela *rel_hi, bfd_vma symval)
+		       Elf_Internal_Rela *rel_hi, bfd_vma symval,
+		       struct bfd_link_info *info)
 {
   bfd_byte *contents = elf_section_data (sec)->this_hdr.contents;
   Elf_Internal_Rela *rel_lo = rel_hi + 2;
@@ -3771,8 +3772,9 @@ loongarch_relax_pcala_addi (bfd *abfd, asection *sec,
   /* Adjust relocations.  */
   rel_hi->r_info = ELFNN_R_INFO (ELFNN_R_SYM (rel_hi->r_info),
 				 R_LARCH_PCREL20_S2);
-  rel_lo->r_info = ELFNN_R_INFO (ELFNN_R_SYM (rel_hi->r_info),
-				 R_LARCH_DELETE);
+  rel_lo->r_info = ELFNN_R_INFO (0, R_LARCH_NONE);
+
+  loongarch_relax_delete_bytes (abfd, sec, rel_lo->r_offset, 4, info);
 
   return true;
 }
@@ -4004,14 +4006,14 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
 	  break;
 	case R_LARCH_PCALA_HI20:
 	  if (0 == info->relax_pass && (i + 4) <= sec->reloc_count)
-	    loongarch_relax_pcala_addi (abfd, sec, rel, symval);
+	    loongarch_relax_pcala_addi (abfd, sec, rel, symval, info);
 	  break;
 	case R_LARCH_GOT_PC_HI20:
 	  if (local_got && 0 == info->relax_pass
 	      && (i + 4) <= sec->reloc_count)
 	    {
 	      if (loongarch_relax_pcala_ld (abfd, sec, rel))
-		loongarch_relax_pcala_addi (abfd, sec, rel, symval);
+		loongarch_relax_pcala_addi (abfd, sec, rel, symval, info);
 	    }
 	  break;
 	default:
-- 
2.36.0

[PATCH v2 3/6] LoongArch: Multiple relax_trip in one relax_pass

[mengqinggang]

If deleting instructions in one relax_trip, set again to true to start the
next relax_trip.
---
 bfd/elfnn-loongarch.c | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index 9d4ea7e4ae7..7436a14441f 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -3740,7 +3740,7 @@ loongarch_relax_delete_bytes (bfd *abfd,
 static bool
 loongarch_relax_pcala_addi (bfd *abfd, asection *sec,
 		       Elf_Internal_Rela *rel_hi, bfd_vma symval,
-		       struct bfd_link_info *info)
+		       struct bfd_link_info *info, bool *again)
 {
   bfd_byte *contents = elf_section_data (sec)->this_hdr.contents;
   Elf_Internal_Rela *rel_lo = rel_hi + 2;
@@ -3766,6 +3766,9 @@ loongarch_relax_pcala_addi (bfd *abfd, asection *sec,
       || ((bfd_signed_vma)(symval - pc) > (bfd_signed_vma)(int32_t)0x1ffffc))
     return false;
 
+  /* Continue next relax trip.  */
+  *again = true;
+
   pca = pcaddi | rd;
   bfd_put (32, abfd, pca, contents + rel_hi->r_offset);
 
@@ -4006,14 +4009,15 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
 	  break;
 	case R_LARCH_PCALA_HI20:
 	  if (0 == info->relax_pass && (i + 4) <= sec->reloc_count)
-	    loongarch_relax_pcala_addi (abfd, sec, rel, symval, info);
+	    loongarch_relax_pcala_addi (abfd, sec, rel, symval, info, again);
 	  break;
 	case R_LARCH_GOT_PC_HI20:
 	  if (local_got && 0 == info->relax_pass
 	      && (i + 4) <= sec->reloc_count)
 	    {
 	      if (loongarch_relax_pcala_ld (abfd, sec, rel))
-		loongarch_relax_pcala_addi (abfd, sec, rel, symval, info);
+		loongarch_relax_pcala_addi (abfd, sec, rel, symval,
+					    info, again);
 	    }
 	  break;
 	default:
-- 
2.36.0

[PATCH v2 4/6] LoongArch: Remove "elf_seg_map (info->output_bfd) == NULL" relaxation condition

[mengqinggang]

Previously the condition prevented shared objects from being relaxed.
To remove the limitation, we need to update program header size and
.eh_frame_hdr size before relaxation.
---
 bfd/elfnn-loongarch.c        | 25 +++++++++++++++++++++----
 ld/emultempl/loongarchelf.em | 18 ++++++++++++++++++
 2 files changed, 39 insertions(+), 4 deletions(-)

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index 7436a14441f..987e5b98bc0 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -3738,7 +3738,7 @@ loongarch_relax_delete_bytes (bfd *abfd,
 
 /* Relax pcalau12i,addi.d => pcaddi.  */
 static bool
-loongarch_relax_pcala_addi (bfd *abfd, asection *sec,
+loongarch_relax_pcala_addi (bfd *abfd, asection *sec, asection *sym_sec,
 		       Elf_Internal_Rela *rel_hi, bfd_vma symval,
 		       struct bfd_link_info *info, bool *again)
 {
@@ -3747,7 +3747,24 @@ loongarch_relax_pcala_addi (bfd *abfd, asection *sec,
   uint32_t pca = bfd_get (32, abfd, contents + rel_hi->r_offset);
   uint32_t add = bfd_get (32, abfd, contents + rel_lo->r_offset);
   uint32_t rd = pca & 0x1f;
+
+  /* This section's output_offset need to subtract the bytes of instructions
+     relaxed by the previous sections, so it needs to be updated beforehand.
+     size_input_section already took care of updating it after relaxation,
+     so we additionally update once here.  */
+  sec->output_offset = sec->output_section->size;
   bfd_vma pc = sec_addr (sec) + rel_hi->r_offset;
+
+  /* If pc and symbol not in the same segment, add/sub segment alignment.
+     FIXME: if there are multiple readonly segments?  */
+  if (!(sym_sec->flags & SEC_READONLY))
+    {
+      if (symval > pc)
+	pc -= info->maxpagesize;
+      else if (symval < pc)
+	pc += info->maxpagesize;
+    }
+
   const uint32_t addi_d = 0x02c00000;
   const uint32_t pcaddi = 0x18000000;
 
@@ -3889,7 +3906,6 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
       || sec->sec_flg0
       || (sec->flags & SEC_RELOC) == 0
       || sec->reloc_count == 0
-      || elf_seg_map (info->output_bfd) == NULL
       || (info->disable_target_specific_optimizations
 	  && info->relax_pass == 0)
       /* The exp_seg_relro_adjust is enum phase_enum (0x4),
@@ -4009,14 +4025,15 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
 	  break;
 	case R_LARCH_PCALA_HI20:
 	  if (0 == info->relax_pass && (i + 4) <= sec->reloc_count)
-	    loongarch_relax_pcala_addi (abfd, sec, rel, symval, info, again);
+	    loongarch_relax_pcala_addi (abfd, sec, sym_sec, rel, symval,
+					info, again);
 	  break;
 	case R_LARCH_GOT_PC_HI20:
 	  if (local_got && 0 == info->relax_pass
 	      && (i + 4) <= sec->reloc_count)
 	    {
 	      if (loongarch_relax_pcala_ld (abfd, sec, rel))
-		loongarch_relax_pcala_addi (abfd, sec, rel, symval,
+		loongarch_relax_pcala_addi (abfd, sec, sym_sec, rel, symval,
 					    info, again);
 	    }
 	  break;
diff --git a/ld/emultempl/loongarchelf.em b/ld/emultempl/loongarchelf.em
index 4850feb8767..d81c99da48b 100644
--- a/ld/emultempl/loongarchelf.em
+++ b/ld/emultempl/loongarchelf.em
@@ -62,6 +62,24 @@ gld${EMULATION_NAME}_after_allocation (void)
 	}
     }
 
+  /* The program header size of executable file may increase.  */
+  if (bfd_get_flavour (link_info.output_bfd) == bfd_target_elf_flavour
+      && !bfd_link_relocatable (&link_info))
+    {
+      if (lang_phdr_list == NULL)
+        elf_seg_map (link_info.output_bfd) = NULL;
+      if (!_bfd_elf_map_sections_to_segments (link_info.output_bfd,
+					      &link_info,
+					      NULL))
+        einfo (_("%F%P: map sections to segments failed: %E\n"));
+    }
+
+  /* Adjust program header size and .eh_frame_hdr size before
+     lang_relax_sections. Without it, the vma of data segment may increase.  */
+  lang_do_assignments (lang_allocating_phase_enum);
+  lang_reset_memory_regions ();
+  lang_size_sections (NULL, true);
+
   enum phase_enum *phase = &(expld.dataseg.phase);
   bfd_elf${ELFSIZE}_loongarch_set_data_segment_info (&link_info, (int *) phase);
   /* gld${EMULATION_NAME}_map_segments (need_layout); */
-- 
2.36.0

[PATCH v2 5/6] LoongArch: Modify link_info.relax_pass from 3 to 2

[mengqinggang]

The first pass handles R_LARCH_RELAX relocations, the second pass
handles R_LARCH_ALIGN relocations.
---
 bfd/elfnn-loongarch.c        | 2 +-
 ld/emultempl/loongarchelf.em | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/bfd/elfnn-loongarch.c b/bfd/elfnn-loongarch.c
index 987e5b98bc0..024c5d4cd96 100644
--- a/bfd/elfnn-loongarch.c
+++ b/bfd/elfnn-loongarch.c
@@ -4013,7 +4013,7 @@ loongarch_elf_relax_section (bfd *abfd, asection *sec,
       switch (ELFNN_R_TYPE (rel->r_info))
 	{
 	case R_LARCH_ALIGN:
-	  if (2 == info->relax_pass)
+	  if (1 == info->relax_pass)
 	    loongarch_relax_align (abfd, sec, sym_sec, info, rel, symval);
 	  break;
 	case R_LARCH_DELETE:
diff --git a/ld/emultempl/loongarchelf.em b/ld/emultempl/loongarchelf.em
index d81c99da48b..9974989489c 100644
--- a/ld/emultempl/loongarchelf.em
+++ b/ld/emultempl/loongarchelf.em
@@ -42,7 +42,7 @@ larch_elf_before_allocation (void)
 	ENABLE_RELAXATION;
     }
 
-  link_info.relax_pass = 3;
+  link_info.relax_pass = 2;
 }
 
 static void
-- 
2.36.0

[PATCH v2 6/6] LoongArch: Add more relaxation testcases

[mengqinggang]

1. .so relaxation testcase
2. ld --no-relax testcase
3. segment alignment testcase
---
 .../ld-loongarch-elf/relax-segment-max.s      | 12 +++
 .../ld-loongarch-elf/relax-segment-min.s      | 12 +++
 ld/testsuite/ld-loongarch-elf/relax-so.s      |  4 +
 ld/testsuite/ld-loongarch-elf/relax.exp       | 80 +++++++++++++++++--
 4 files changed, 103 insertions(+), 5 deletions(-)
 create mode 100644 ld/testsuite/ld-loongarch-elf/relax-segment-max.s
 create mode 100644 ld/testsuite/ld-loongarch-elf/relax-segment-min.s
 create mode 100644 ld/testsuite/ld-loongarch-elf/relax-so.s

diff --git a/ld/testsuite/ld-loongarch-elf/relax-segment-max.s b/ld/testsuite/ld-loongarch-elf/relax-segment-max.s
new file mode 100644
index 00000000000..df15d4cab09
--- /dev/null
+++ b/ld/testsuite/ld-loongarch-elf/relax-segment-max.s
@@ -0,0 +1,12 @@
+# The .align may cause overflow because deleting nops.
+  .text		      # 0x120004000
+  .align 3
+  la.local $r12, .L1
+
+#  .fill 0x1f7ffc # max fill without overflow, .data address is 0x120200000
+#  .fill 0x1f8000 # min fill with overflow, .data address is 0x120204000
+  .fill 0x1fbff4 # max fill with overflow, .data address is 0x120204000
+
+  .data
+.L1:
+  .byte 2
diff --git a/ld/testsuite/ld-loongarch-elf/relax-segment-min.s b/ld/testsuite/ld-loongarch-elf/relax-segment-min.s
new file mode 100644
index 00000000000..476b9fba78c
--- /dev/null
+++ b/ld/testsuite/ld-loongarch-elf/relax-segment-min.s
@@ -0,0 +1,12 @@
+# The .align may cause overflow because deleting nops.
+  .text		      # 0x120004000
+  .align 3
+  la.local $r12, .L1
+
+#  .fill 0x1f7ffc # max fill without overflow, .data address is 0x120200000
+  .fill 0x1f8000 # min fill with overflow, .data address is 0x120204000
+#  .fill 0x1fbff4 # max fill with overflow, .data address is 0x120204000
+
+  .data
+.L1:
+  .byte 2
diff --git a/ld/testsuite/ld-loongarch-elf/relax-so.s b/ld/testsuite/ld-loongarch-elf/relax-so.s
new file mode 100644
index 00000000000..01a404a8b47
--- /dev/null
+++ b/ld/testsuite/ld-loongarch-elf/relax-so.s
@@ -0,0 +1,4 @@
+.text
+.align 2
+.L1:
+  la.local $r12, .L1
diff --git a/ld/testsuite/ld-loongarch-elf/relax.exp b/ld/testsuite/ld-loongarch-elf/relax.exp
index 7ff876d7914..24d79ed5c20 100644
--- a/ld/testsuite/ld-loongarch-elf/relax.exp
+++ b/ld/testsuite/ld-loongarch-elf/relax.exp
@@ -22,7 +22,7 @@
 if [istarget loongarch64-*-*] {
 
   if [isbuild loongarch64-*-*] {
-    set testname "loongarch relax build"
+    set testname "loongarch relax .exe build"
     set pre_builds [list \
       [list \
 	"$testname" \
@@ -39,17 +39,87 @@ if [istarget loongarch64-*-*] {
     if [file exist "tmpdir/relax"] {
       set objdump_output [run_host_cmd "objdump" "-d tmpdir/relax"]
       if { [ regexp ".*pcaddi.*pcaddi.*" $objdump_output] } {
-	pass "loongarch relax"
+	pass "loongarch relax .exe"
       } {
-	fail "loongarch relax"
+	fail "loongarch relax .exe"
       }
     }
+
+    set testname "loongarch ld --no-relax build"
+    set pre_builds [list \
+      [list \
+	"$testname" \
+	"-Wl,--no-relax" \
+	"" \
+	{relax.s} \
+	{} \
+	"norelax" \
+      ] \
+    ]
+
+    run_cc_link_tests $pre_builds
+
+    if [file exist "tmpdir/norelax"] {
+      set objdump_output [run_host_cmd "objdump" "-d tmpdir/norelax"]
+      if { [ regexp ".*pcaddi.*" $objdump_output] } {
+	fail "loongarch ld --no-relax"
+      } {
+	pass "loongarch ld --no-relax"
+      }
+    }
+
+    run_ld_link_tests \
+	[list \
+	    [list \
+		"loongarch relax .so build" \
+		"-shared -e 0x0" "" \
+		"" \
+		{relax-so.s} \
+		{} \
+		"relax-so" \
+	    ] \
+	]
+
+      if [file exist "tmpdir/relax-so"] {
+	set objdump_output [run_host_cmd "objdump" "-d tmpdir/relax-so"]
+	if { [ regexp ".*pcaddi.*" $objdump_output] } {
+	  pass "loongarch relax .so"
+	} {
+	  fail "loongarch relax .so"
+	}
+      }
+
+    # If symbol in data segment, offset need to sub segment align to prevent
+    # overflow.
+    run_ld_link_tests \
+	[list \
+	    [list \
+		"loongarch relax segment alignment min" \
+		"-e0 -Ttext 0x120004000 -pie -z relro" "" \
+		"" \
+		{relax-segment-min.s} \
+		{} \
+		"relax-segment-min" \
+	    ] \
+	]
+
+    run_ld_link_tests \
+	[list \
+	    [list \
+		"loongarch relax segment alignment max" \
+		"-e0 -Ttext 0x120004000 -pie -z relro" "" \
+		"" \
+		{relax-segment-max.s} \
+		{} \
+		"relax-segment-max" \
+	    ] \
+	]
   }
 
   run_ld_link_tests \
       [list \
 	  [list \
-	      "relax-align" \
+	      "loongarch relax-align" \
 	      "-e 0x0 -z relro" "" \
 	      "" \
 	      {relax-align.s} \
@@ -64,7 +134,7 @@ if [istarget loongarch64-*-*] {
   run_ld_link_tests \
       [list \
 	  [list \
-	      "uleb128" \
+	      "loongarch uleb128" \
 	      "-e 0x0" "" \
 	      "" \
 	      {uleb128.s} \
-- 
2.36.0