LoongArch/GAS: Add support for branch relaxation

For the instructions of R_LARCH_B16/B21, if the immediate overflow,
add a B instruction and R_LARCH_B26 relocation.

For example:

.L1
  ...
  blt $t0, $t1, .L1
    R_LARCH_B16

change to:

.L1
  ...
  bge $t0, $t1, .L2
  b .L1
    R_LARCH_B26
.L2

gas/config/tc-loongarch.c

@@ -106,6 +106,16 @@ const char *md_shortopts = "O::g::G:";
 
 static const char default_arch[] = DEFAULT_ARCH;
 
+/* The lowest 4-bit is the bytes of instructions.  */
+#define RELAX_BRANCH_16 0xc0000014
+#define RELAX_BRANCH_21 0xc0000024
+#define RELAX_BRANCH_26 0xc0000048
+
+#define RELAX_BRANCH(x) \
+  (((x) & 0xf0000000) == 0xc0000000)
+#define RELAX_BRANCH_ENCODE(x) \
+  (BFD_RELOC_LARCH_B16 == (x) ? RELAX_BRANCH_16 : RELAX_BRANCH_21)
+
 enum options
 {
   OPTION_IGNORE = OPTION_MD_BASE,
@@ -953,11 +963,22 @@ append_fixed_insn (struct loongarch_cl_insn *insn)
   move_insn (insn, frag_now, f - frag_now->fr_literal);
 }
 
+/* Add instructions based on the worst-case scenario firstly.  */
+static void
+append_relaxed_branch_insn (struct loongarch_cl_insn *insn, int max_chars,
+	    int var, relax_substateT subtype, symbolS *symbol, offsetT offset)
+{
+  frag_grow (max_chars);
+  move_insn (insn, frag_now, frag_more (0) - frag_now->fr_literal);
+  frag_var (rs_machine_dependent, max_chars, var,
+	    subtype, symbol, offset, NULL);
+}
+
 static void
 append_fixp_and_insn (struct loongarch_cl_insn *ip)
 {
   reloc_howto_type *howto;
-  bfd_reloc_code_real_type reloc_type;
+  bfd_reloc_code_real_type r_type;
   struct reloc_info *reloc_info = ip->reloc_info;
   size_t i;
 
@@ -965,14 +986,40 @@ append_fixp_and_insn (struct loongarch_cl_insn *ip)
 
   for (i = 0; i < ip->reloc_num; i++)
     {
-      reloc_type = reloc_info[i].type;
-      howto = bfd_reloc_type_lookup (stdoutput, reloc_type);
-      if (howto == NULL)
-	as_fatal (_("no HOWTO loong relocation number %d"), reloc_type);
+      r_type = reloc_info[i].type;
 
-      ip->fixp[i] =
-	fix_new_exp (ip->frag, ip->where, bfd_get_reloc_size (howto),
-		     &reloc_info[i].value, FALSE, reloc_type);
+      if (r_type != BFD_RELOC_UNUSED)
+	{
+
+	  gas_assert (&(reloc_info[i].value));
+	  if (BFD_RELOC_LARCH_B16 == r_type || BFD_RELOC_LARCH_B21 == r_type)
+	    {
+	      int min_bytes = 4; /* One branch instruction.  */
+	      unsigned max_bytes = 8; /* Branch and jump instructions.  */
+
+	      if (now_seg == absolute_section)
+		{
+		  as_bad (_("relaxable branches not supported in absolute section"));
+		  return;
+		}
+
+	      append_relaxed_branch_insn (ip, max_bytes, min_bytes,
+					  RELAX_BRANCH_ENCODE (r_type),
+					  reloc_info[i].value.X_add_symbol,
+					  reloc_info[i].value.X_add_number);
+	      return;
+	    }
+	  else
+	    {
+	      howto = bfd_reloc_type_lookup (stdoutput, r_type);
+	      if (howto == NULL)
+		as_fatal (_("no HOWTO loong relocation number %d"), r_type);
+
+	      ip->fixp[i] = fix_new_exp (ip->frag, ip->where,
+					 bfd_get_reloc_size (howto),
+					 &reloc_info[i].value, FALSE, r_type);
+	    }
+	}
     }
 
   if (ip->insn_length < ip->relax_max_length)
@@ -1487,14 +1534,6 @@ md_apply_fix (fixS *fixP, valueT *valP, segT seg ATTRIBUTE_UNUSED)
     }
 }
 
-int
-loongarch_relax_frag (asection *sec ATTRIBUTE_UNUSED,
-		      fragS *fragp ATTRIBUTE_UNUSED,
-		      long stretch ATTRIBUTE_UNUSED)
-{
-  return 0;
-}
-
 int
 md_estimate_size_before_relax (fragS *fragp ATTRIBUTE_UNUSED,
 			       asection *segtype ATTRIBUTE_UNUSED)
@@ -1526,30 +1565,6 @@ tc_gen_reloc (asection *section ATTRIBUTE_UNUSED, fixS *fixp)
   return reloc;
 }
 
-/* Convert a machine dependent frag.  */
-void
-md_convert_frag (bfd *abfd ATTRIBUTE_UNUSED, segT asec ATTRIBUTE_UNUSED,
-		 fragS *fragp)
-{
-  expressionS exp;
-  exp.X_op = O_symbol;
-  exp.X_add_symbol = fragp->fr_symbol;
-  exp.X_add_number = fragp->fr_offset;
-  bfd_byte *buf = (bfd_byte *)fragp->fr_literal + fragp->fr_fix;
-
-  fixS *fixp = fix_new_exp (fragp, buf - (bfd_byte *)fragp->fr_literal,
-				4, &exp, false, fragp->fr_subtype);
-  buf += 4;
-
-  fixp->fx_file = fragp->fr_file;
-  fixp->fx_line = fragp->fr_line;
-  fragp->fr_fix += fragp->fr_var;
-
-  gas_assert (fragp->fr_next == NULL
-	      || (fragp->fr_next->fr_address - fragp->fr_address
-		  == fragp->fr_fix));
-}
-
 /* Standard calling conventions leave the CFA at SP on entry.  */
 void
 loongarch_cfi_frame_initial_instructions (void)
@@ -1775,3 +1790,142 @@ loongarch_elf_final_processing (void)
 {
   elf_elfheader (stdoutput)->e_flags = LARCH_opts.ase_abi;
 }
+
+/* Compute the length of a branch sequence, and adjust the stored length
+   accordingly.  If FRAGP is NULL, the worst-case length is returned.  */
+static unsigned
+loongarch_relaxed_branch_length (fragS *fragp, asection *sec, int update)
+{
+  int length = 4;
+
+  if (!fragp)
+    return 8;
+
+  if (fragp->fr_symbol != NULL
+      && S_IS_DEFINED (fragp->fr_symbol)
+      && !S_IS_WEAK (fragp->fr_symbol)
+      && sec == S_GET_SEGMENT (fragp->fr_symbol))
+    {
+      offsetT val = S_GET_VALUE (fragp->fr_symbol) + fragp->fr_offset;
+
+      val -= fragp->fr_address + fragp->fr_fix;
+
+      if (RELAX_BRANCH_16 == fragp->fr_subtype
+	  && OUT_OF_RANGE (val, 16, 2))
+	{
+	  length = 8;
+	  if (update)
+	    fragp->fr_subtype = RELAX_BRANCH_26;
+	}
+
+      if (RELAX_BRANCH_21 == fragp->fr_subtype
+	  && OUT_OF_RANGE (val, 21, 2))
+	{
+	  length = 8;
+	  if (update)
+	    fragp->fr_subtype = RELAX_BRANCH_26;
+	}
+
+      if (RELAX_BRANCH_26 == fragp->fr_subtype)
+	length = 8;
+    }
+
+  return length;
+}
+
+int
+loongarch_relax_frag (asection *sec ATTRIBUTE_UNUSED,
+		      fragS *fragp ATTRIBUTE_UNUSED,
+		      long stretch ATTRIBUTE_UNUSED)
+{
+  if (RELAX_BRANCH (fragp->fr_subtype))
+    {
+      offsetT old_var = fragp->fr_var;
+      fragp->fr_var = loongarch_relaxed_branch_length (fragp, sec, true);
+      return fragp->fr_var - old_var;
+    }
+  return 0;
+}
+
+/* Expand far branches to multi-instruction sequences.
+   Branch instructions:
+   beq, bne, blt, bgt, bltz, bgtz, ble, bge, blez, bgez
+   bltu, bgtu, bleu, bgeu
+   beqz, bnez, bceqz, bcnez.  */
+
+static void
+loongarch_convert_frag_branch (fragS *fragp)
+{
+  bfd_byte *buf;
+  expressionS exp;
+  fixS *fixp;
+  insn_t insn;
+
+  buf = (bfd_byte *)fragp->fr_literal + fragp->fr_fix;
+
+  exp.X_op = O_symbol;
+  exp.X_add_symbol = fragp->fr_symbol;
+  exp.X_add_number = fragp->fr_offset;
+
+  gas_assert ((fragp->fr_subtype & 0xf) == fragp->fr_var);
+
+  /* blt $t0, $t1, .L1
+     nop
+     change to:
+     bge $t0, $t1, .L2
+     b .L1
+   .L2:
+     nop  */
+  switch (fragp->fr_subtype)
+    {
+    case RELAX_BRANCH_26:
+      insn = bfd_getl32 (buf);
+      /* Invert the branch condition.  */
+      if (LARCH_FLOAT_BRANCH == (insn & LARCH_BRANCH_OPCODE_MASK))
+	insn ^= LARCH_FLOAT_BRANCH_INVERT_BIT;
+      else
+	insn ^= LARCH_BRANCH_INVERT_BIT;
+      insn |= ENCODE_BRANCH16_IMM (8);  /* Set target to PC + 8.  */
+      bfd_putl32 (insn, buf);
+      buf += 4;
+
+      /* Add the B instruction and jump to the original target.  */
+      bfd_putl32 (LARCH_B, buf);
+      fixp = fix_new_exp (fragp, buf - (bfd_byte *)fragp->fr_literal,
+			  4, &exp, false, BFD_RELOC_LARCH_B26);
+      buf += 4;
+      break;
+    case RELAX_BRANCH_21:
+      fixp = fix_new_exp (fragp, buf - (bfd_byte *)fragp->fr_literal,
+			  4, &exp, false, BFD_RELOC_LARCH_B21);
+      buf += 4;
+      break;
+    case RELAX_BRANCH_16:
+      fixp = fix_new_exp (fragp, buf - (bfd_byte *)fragp->fr_literal,
+			  4, &exp, false, BFD_RELOC_LARCH_B16);
+      buf += 4;
+      break;
+
+    default:
+      abort();
+    }
+
+  fixp->fx_file = fragp->fr_file;
+  fixp->fx_line = fragp->fr_line;
+
+  gas_assert (buf == (bfd_byte *)fragp->fr_literal
+	      + fragp->fr_fix + fragp->fr_var);
+
+  fragp->fr_fix += fragp->fr_var;
+}
+
+/* Relax a machine dependent frag.  This returns the amount by which
+   the current size of the frag should change.  */
+
+void
+md_convert_frag (bfd *abfd ATTRIBUTE_UNUSED, segT asec ATTRIBUTE_UNUSED,
+		 fragS *fragp)
+{
+  gas_assert (RELAX_BRANCH (fragp->fr_subtype));
+  loongarch_convert_frag_branch (fragp);
+}

gas/testsuite/gas/loongarch/la_branch_relax_1.d

@@ -0,0 +1,64 @@
+#as:
+#objdump: -dr
+
+.*:[    ]+file format .*
+
+
+Disassembly of section .text:
+
+0* <.L1>:
+[ 	]+...
+[ 	]+48d158:[ 	]+5c00098d[ 	]+bne[ 	]+\$t0, \$t1, 8[ 	]+# 48d160 <.L1\+0x48d160>
+[ 	]+48d15c:[ 	]+532ea7ed[ 	]+b[ 	]+-4772188[ 	]+# 0 <.L1>
+[ 	]+48d15c: R_LARCH_B26[ 	]+.L1
+[ 	]+48d160:[ 	]+5800098d[ 	]+beq[ 	]+\$t0, \$t1, 8[ 	]+# 48d168 <.L1\+0x48d168>
+[ 	]+48d164:[ 	]+532e9fed[ 	]+b[ 	]+-4772196[ 	]+# 0 <.L1>
+[ 	]+48d164: R_LARCH_B26[ 	]+.L1
+[ 	]+48d168:[ 	]+6400098d[ 	]+bge[ 	]+\$t0, \$t1, 8[ 	]+# 48d170 <.L1\+0x48d170>
+[ 	]+48d16c:[ 	]+532e97ed[ 	]+b[ 	]+-4772204[ 	]+# 0 <.L1>
+[ 	]+48d16c: R_LARCH_B26[ 	]+.L1
+[ 	]+48d170:[ 	]+640009ac[ 	]+bge[ 	]+\$t1, \$t0, 8[ 	]+# 48d178 <.L1\+0x48d178>
+[ 	]+48d174:[ 	]+532e8fed[ 	]+b[ 	]+-4772212[ 	]+# 0 <.L1>
+[ 	]+48d174: R_LARCH_B26[ 	]+.L1
+[ 	]+48d178:[ 	]+64000980[ 	]+bgez[ 	]+\$t0, 8[ 	]+# 48d180 <.L1\+0x48d180>
+[ 	]+48d17c:[ 	]+532e87ed[ 	]+b[ 	]+-4772220[ 	]+# 0 <.L1>
+[ 	]+48d17c: R_LARCH_B26[ 	]+.L1
+[ 	]+48d180:[ 	]+6400080c[ 	]+blez[ 	]+\$t0, 8[ 	]+# 48d188 <.L1\+0x48d188>
+[ 	]+48d184:[ 	]+532e7fed[ 	]+b[ 	]+-4772228[ 	]+# 0 <.L1>
+[ 	]+48d184: R_LARCH_B26[ 	]+.L1
+[ 	]+48d188:[ 	]+600009ac[ 	]+blt[ 	]+\$t1, \$t0, 8[ 	]+# 48d190 <.L1\+0x48d190>
+[ 	]+48d18c:[ 	]+532e77ed[ 	]+b[ 	]+-4772236[ 	]+# 0 <.L1>
+[ 	]+48d18c: R_LARCH_B26[ 	]+.L1
+[ 	]+48d190:[ 	]+6000098d[ 	]+blt[ 	]+\$t0, \$t1, 8[ 	]+# 48d198 <.L1\+0x48d198>
+[ 	]+48d194:[ 	]+532e6fed[ 	]+b[ 	]+-4772244[ 	]+# 0 <.L1>
+[ 	]+48d194: R_LARCH_B26[ 	]+.L1
+[ 	]+48d198:[ 	]+6000080c[ 	]+bgtz[ 	]+\$t0, 8[ 	]+# 48d1a0 <.L1\+0x48d1a0>
+[ 	]+48d19c:[ 	]+532e67ed[ 	]+b[ 	]+-4772252[ 	]+# 0 <.L1>
+[ 	]+48d19c: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1a0:[ 	]+60000980[ 	]+bltz[ 	]+\$t0, 8[ 	]+# 48d1a8 <.L1\+0x48d1a8>
+[ 	]+48d1a4:[ 	]+532e5fed[ 	]+b[ 	]+-4772260[ 	]+# 0 <.L1>
+[ 	]+48d1a4: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1a8:[ 	]+6c00098d[ 	]+bgeu[ 	]+\$t0, \$t1, 8[ 	]+# 48d1b0 <.L1\+0x48d1b0>
+[ 	]+48d1ac:[ 	]+532e57ed[ 	]+b[ 	]+-4772268[ 	]+# 0 <.L1>
+[ 	]+48d1ac: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1b0:[ 	]+6c0009ac[ 	]+bgeu[ 	]+\$t1, \$t0, 8[ 	]+# 48d1b8 <.L1\+0x48d1b8>
+[ 	]+48d1b4:[ 	]+532e4fed[ 	]+b[ 	]+-4772276[ 	]+# 0 <.L1>
+[ 	]+48d1b4: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1b8:[ 	]+680009ac[ 	]+bltu[ 	]+\$t1, \$t0, 8[ 	]+# 48d1c0 <.L1\+0x48d1c0>
+[ 	]+48d1bc:[ 	]+532e47ed[ 	]+b[ 	]+-4772284[ 	]+# 0 <.L1>
+[ 	]+48d1bc: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1c0:[ 	]+6800098d[ 	]+bltu[ 	]+\$t0, \$t1, 8[ 	]+# 48d1c8 <.L1\+0x48d1c8>
+[ 	]+48d1c4:[ 	]+532e3fed[ 	]+b[ 	]+-4772292[ 	]+# 0 <.L1>
+[ 	]+48d1c4: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1c8:[ 	]+44000980[ 	]+bnez[ 	]+\$t0, 8[ 	]+# 48d1d0 <.L1\+0x48d1d0>
+[ 	]+48d1cc:[ 	]+532e37ed[ 	]+b[ 	]+-4772300[ 	]+# 0 <.L1>
+[ 	]+48d1cc: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1d0:[ 	]+40000980[ 	]+beqz[ 	]+\$t0, 8[ 	]+# 48d1d8 <.L1\+0x48d1d8>
+[ 	]+48d1d4:[ 	]+532e2fed[ 	]+b[ 	]+-4772308[ 	]+# 0 <.L1>
+[ 	]+48d1d4: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1d8:[ 	]+48000900[ 	]+bcnez[ 	]+\$fcc0, 8[ 	]+# 48d1e0 <.L1\+0x48d1e0>
+[ 	]+48d1dc:[ 	]+532e27ed[ 	]+b[ 	]+-4772316[ 	]+# 0 <.L1>
+[ 	]+48d1dc: R_LARCH_B26[ 	]+.L1
+[ 	]+48d1e0:[ 	]+48000800[ 	]+bceqz[ 	]+\$fcc0, 8[ 	]+# 48d1e8 <.L1\+0x48d1e8>
+[ 	]+48d1e4:[ 	]+532e1fed[ 	]+b[ 	]+-4772324[ 	]+# 0 <.L1>
+[ 	]+48d1e4: R_LARCH_B26[ 	]+.L1

gas/testsuite/gas/loongarch/la_branch_relax_1.s

@@ -0,0 +1,33 @@
+# Instruction and Relocation generating tests
+
+.L1:
+  .fill 0x123456, 4, 0x0
+
+# R_LARCH_B16
+  beq $r12, $r13, .L1
+  bne $r12, $r13, .L1
+
+  blt $r12, $r13, .L1
+  bgt $r12, $r13, .L1
+
+  bltz $r12, .L1
+  bgtz $r12, .L1
+
+  ble $r12, $r13, .L1
+  bge $r12, $r13, .L1
+
+  blez $r12, .L1
+  bgez $r12, .L1
+
+  bltu $r12, $r13, .L1
+  bgtu $r12, $r13, .L1
+
+  bleu $r12, $r13, .L1
+  bgeu $r12, $r13, .L1
+
+# R_LARCH_B21
+  beqz $r12, .L1
+  bnez $r12, .L1
+
+  bceqz $fcc0, .L1
+  bcnez $fcc0, .L1

gas/testsuite/gas/loongarch/la_branch_relax_2.d

@@ -0,0 +1,40 @@
+#as:
+#objdump: -dr
+
+.*:[    ]+file format .*
+
+
+Disassembly of section .text:
+
+0* <.L1>:
+[ 	]+...
+[ 	]+20000:[ 	]+5a00018d[ 	]+beq[ 	]+\$t0, \$t1, -131072[ 	]+# 0 <.L1>
+[ 	]+20000: R_LARCH_B16[ 	]+.L1
+[ 	]+20004:[ 	]+5c00098d[ 	]+bne[ 	]+\$t0, \$t1, 8[ 	]+# 2000c <.L1\+0x2000c>
+[ 	]+20008:[ 	]+51fffbff[ 	]+b[ 	]+-131080[ 	]+# 0 <.L1>
+[ 	]+20008: R_LARCH_B26[ 	]+.L1
+[ 	]+2000c:[ 	]+5c00098d[ 	]+bne[ 	]+\$t0, \$t1, 8[ 	]+# 20014 <.L1\+0x20014>
+[ 	]+20010:[ 	]+52000000[ 	]+b[ 	]+131072[ 	]+# 40010 <.L2>
+[ 	]+20010: R_LARCH_B26[ 	]+.L2
+[ 	]+20014:[ 	]+59fffd8d[ 	]+beq[ 	]+\$t0, \$t1, 131068[ 	]+# 40010 <.L2>
+[ 	]+20014: R_LARCH_B16[ 	]+.L2
+[ 	]+...
+0*40010 <.L2>:
+[ 	]+...
+[ 	]+440010:[ 	]+40000190[ 	]+beqz[ 	]+\$t0, -4194304[ 	]+# 40010 <.L2>
+[ 	]+440010: R_LARCH_B21[ 	]+.L2
+[ 	]+440014:[ 	]+44000980[ 	]+bnez[ 	]+\$t0, 8[ 	]+# 44001c <.L2\+0x40000c>
+[ 	]+440018:[ 	]+53fffbef[ 	]+b[ 	]+-4194312[ 	]+# 40010 <.L2>
+[ 	]+440018: R_LARCH_B26[ 	]+.L2
+[ 	]+44001c:[ 	]+44000980[ 	]+bnez[ 	]+\$t0, 8[ 	]+# 440024 <.L2\+0x400014>
+[ 	]+440020:[ 	]+50000010[ 	]+b[ 	]+4194304[ 	]+# 840020 <.L3>
+[ 	]+440020: R_LARCH_B26[ 	]+.L3
+[ 	]+440024:[ 	]+43fffd8f[ 	]+beqz[ 	]+\$t0, 4194300[ 	]+# 840020 <.L3>
+[ 	]+440024: R_LARCH_B21[ 	]+.L3
+[ 	]+...
+0*840020 <.L3>:
+[ 	]+840020:[ 	]+5800018d[ 	]+beq[ 	]+\$t0, \$t1, 0[ 	]+# 840020 <.L3>
+[ 	]+840020: R_LARCH_B16[ 	]+.L4
+0*840024 <.L5>:
+[ 	]+840024:[ 	]+40000180[ 	]+beqz[ 	]+\$t0, 0[ 	]+# 840024 <.L5>
+[ 	]+840024: R_LARCH_B21[ 	]+.L5

gas/testsuite/gas/loongarch/la_branch_relax_2.s

@@ -0,0 +1,23 @@
+# Immediate boundary value tests
+
+.L1:
+  .fill 0x8000, 4, 0
+  beq $r12, $r13, .L1 # min imm -0x20000
+  beq $r12, $r13, .L1 # out of range
+  beq $r12, $r13, .L2 # out of range
+  beq $r12, $r13, .L2 # max imm 0x1fffc
+  .fill 0x7ffe, 4, 0
+.L2:
+  .fill 0x100000, 4, 0
+  beqz $r12, .L2 # min imm -0x400000
+  beqz $r12, .L2 # out of range
+  beqz $r12, .L3 # out of range
+  beqz $r12, .L3 # max imm 0x3ffffc
+  .fill 0xffffe, 4, 0
+.L3:
+
+# 0 imm
+.L4:
+  beq $r12, $r13, .L4
+.L5:
+  beqz $r12, .L5

include/opcode/loongarch.h

@@ -29,6 +29,18 @@ extern "C"
 #endif
 
   #define LARCH_NOP 0x03400000
+  #define LARCH_B 0x50000000
+  /* BCEQZ/BCNEZ.  */
+  #define LARCH_FLOAT_BRANCH 0x48000000
+  #define LARCH_BRANCH_OPCODE_MASK 0xfc000000
+  #define LARCH_BRANCH_INVERT_BIT 0x04000000
+  #define LARCH_FLOAT_BRANCH_INVERT_BIT 0x00000100
+
+  #define ENCODE_BRANCH16_IMM(x) (((x) >> 2) << 10)
+
+  #define OUT_OF_RANGE(value, bits, align)	\
+    ((value) < (-(1 << ((bits) - 1) << align)) 	\
+      || (value) > ((((1 << ((bits) - 1)) - 1) << align)))
 
   typedef uint32_t insn_t;