[02/18] x86: parse VEX and alike specifiers for .insn

Message ID 3bab7c05-755b-f494-8d7a-0183ab14563a@suse.com
State Accepted
Headers
Series x86: new .insn directive |

Checks

Context Check Description
snail/binutils-gdb-check success Github commit url

Commit Message

Jan Beulich March 3, 2023, 12:57 p.m. UTC
  All encoding spaces can be used this way; there's a certain risk that
the bits presently reserved could be used for other purposes down the
road, but people using .insn are expected to know what they're doing
anyway. Plus this way there's at least _some_ way to have those bits
set.

For now this will only allow operand-less insns to be encoded this way.
---
For now only numeric parts of the specifiers are handled in a case-
insensitive manner. All other parts have to use upper case letters.
  

Patch

--- a/gas/config/tc-i386.c
+++ b/gas/config/tc-i386.c
@@ -307,6 +307,9 @@  struct _i386_insn
     unsigned int prefixes;
     unsigned char prefix[MAX_PREFIXES];
 
+    /* .insn allows for reserved opcode spaces.  */
+    unsigned char insn_opcode_space;
+
     /* Register is in low 3 bits of opcode.  */
     bool short_form;
 
@@ -568,6 +571,9 @@  static expressionS im_expressions[MAX_IM
 /* Current operand we are working on.  */
 static int this_operand = -1;
 
+/* Are we processing a .insn directive?  */
+#define dot_insn() (i.tm.mnem_off == MN__insn)
+
 /* We support four different modes.  FLAG_CODE variable is used to distinguish
    these.  */
 
@@ -3648,6 +3654,8 @@  build_vex_prefix (const insn_template *t
     vector_length = avxscalar;
   else if (i.tm.opcode_modifier.vex == VEX256)
     vector_length = 1;
+  else if (dot_insn () && i.tm.opcode_modifier.vex == VEX128)
+    vector_length = 0;
   else
     {
       unsigned int op;
@@ -3715,7 +3723,9 @@  build_vex_prefix (const insn_template *t
 
       /* The high 3 bits of the second VEX byte are 1's compliment
 	 of RXB bits from REX.  */
-      i.vex.bytes[1] = (~i.rex & 0x7) << 5 | i.tm.opcode_space;
+      i.vex.bytes[1] = ((~i.rex & 7) << 5)
+		       | (!dot_insn () ? i.tm.opcode_space
+				       : i.insn_opcode_space);
 
       i.vex.bytes[2] = (w << 7
 			| register_specifier << 3
@@ -3852,7 +3862,9 @@  build_evex_prefix (void)
      bits from REX.  */
   gas_assert (i.tm.opcode_space >= SPACE_0F);
   gas_assert (i.tm.opcode_space <= SPACE_EVEXMAP6);
-  i.vex.bytes[1] = (~i.rex & 0x7) << 5 | i.tm.opcode_space;
+  i.vex.bytes[1] = ((~i.rex & 7) << 5)
+		   | (!dot_insn () ? i.tm.opcode_space
+				   : i.insn_opcode_space);
 
   /* The fifth bit of the second EVEX byte is 1's compliment of the
      REX_R bit in VREX.  */
@@ -3968,6 +3980,13 @@  build_evex_prefix (void)
 	case EVEX512:
 	  vec_length = 2 << 5;
 	  break;
+	case EVEX_L3:
+	  if (dot_insn ())
+	    {
+	      vec_length = 3 << 5;
+	      break;
+	    }
+	  /* Fall through.  */
 	default:
 	  abort ();
 	  break;
@@ -10739,6 +10758,7 @@  s_insn (int dummy ATTRIBUTE_UNUSED)
   bad:
       *saved_ilp = saved_char;
       ignore_rest_of_line ();
+      i.tm.mnem_off = 0;
       return;
     }
   line += end - line;
@@ -10761,6 +10781,9 @@  s_insn (int dummy ATTRIBUTE_UNUSED)
 	  && (*e == '.' || is_space_char (*e)))
 	{
 	  xop = true;
+	  /* Arrange for build_vex_prefix() to emit 0x8f.  */
+	  i.tm.opcode_space = SPACE_XOP08;
+	  i.insn_opcode_space = n;
 	  line = e;
 	}
     }
@@ -10784,6 +10807,188 @@  s_insn (int dummy ATTRIBUTE_UNUSED)
 
   if (line > end && *line == '.')
     {
+      /* Length specifier (VEX.L, XOP.L, EVEX.L'L).  */
+      switch (line[1])
+	{
+	case 'L':
+	  switch (line[2])
+	    {
+	    case '0':
+	      if (evex)
+		i.tm.opcode_modifier.evex = EVEX128;
+	      else
+		i.tm.opcode_modifier.vex = VEX128;
+	      break;
+
+	    case '1':
+	      if (evex)
+		i.tm.opcode_modifier.evex = EVEX256;
+	      else
+		i.tm.opcode_modifier.vex = VEX256;
+	      break;
+
+	    case '2':
+	      if (evex)
+		i.tm.opcode_modifier.evex = EVEX512;
+	      break;
+
+	    case '3':
+	      if (evex)
+		i.tm.opcode_modifier.evex = EVEX_L3;
+	      break;
+
+	    case 'I':
+	      if (line[3] == 'G')
+		{
+		  if (evex)
+		    i.tm.opcode_modifier.evex = EVEXLIG;
+		  else
+		    i.tm.opcode_modifier.vex = VEXScalar; /* LIG */
+		  ++line;
+		}
+	      break;
+	    }
+
+	  if (i.tm.opcode_modifier.vex || i.tm.opcode_modifier.evex)
+	    line += 3;
+	  break;
+
+	case '1':
+	  if (line[2] == '2' && line[3] == '8')
+	    {
+	      if (evex)
+		i.tm.opcode_modifier.evex = EVEX128;
+	      else
+		i.tm.opcode_modifier.vex = VEX128;
+	      line += 4;
+	    }
+	  break;
+
+	case '2':
+	  if (line[2] == '5' && line[3] == '6')
+	    {
+	      if (evex)
+		i.tm.opcode_modifier.evex = EVEX256;
+	      else
+		i.tm.opcode_modifier.vex = VEX256;
+	      line += 4;
+	    }
+	  break;
+
+	case '5':
+	  if (evex && line[2] == '1' && line[3] == '2')
+	    {
+	      i.tm.opcode_modifier.evex = EVEX512;
+	      line += 4;
+	    }
+	  break;
+	}
+    }
+
+  if (line > end && *line == '.')
+    {
+      /* embedded prefix (VEX.pp, XOP.pp, EVEX.pp).  */
+      switch (line[1])
+	{
+	case 'N':
+	  if (line[2] == 'P')
+	    line += 3;
+	  break;
+
+	case '6':
+	  if (line[2] == '6')
+	    {
+	      i.tm.opcode_modifier.opcodeprefix = PREFIX_0X66;
+	      line += 3;
+	    }
+	  break;
+
+	case 'F': case 'f':
+	  if (line[2] == '3')
+	    {
+	      i.tm.opcode_modifier.opcodeprefix = PREFIX_0XF3;
+	      line += 3;
+	    }
+	  else if (line[2] == '2')
+	    {
+	      i.tm.opcode_modifier.opcodeprefix = PREFIX_0XF2;
+	      line += 3;
+	    }
+	  break;
+	}
+    }
+
+  if (line > end && !xop && *line == '.')
+    {
+      /* Encoding space (VEX.mmmmm, EVEX.mmmm).  */
+      switch (line[1])
+	{
+	case '0':
+	  if (TOUPPER (line[2]) != 'F')
+	    break;
+	  if (line[3] == '.' || is_space_char (line[3]))
+	    {
+	      i.insn_opcode_space = SPACE_0F;
+	      line += 3;
+	    }
+	  else if (line[3] == '3'
+		   && (line[4] == '8' || TOUPPER (line[4]) == 'A')
+		   && (line[5] == '.' || is_space_char (line[5])))
+	    {
+	      i.insn_opcode_space = line[4] == '8' ? SPACE_0F38 : SPACE_0F3A;
+	      line += 5;
+	    }
+	  break;
+
+	case 'M':
+	  if (ISDIGIT (line[2]) && line[2] != '0')
+	    {
+	      char *e;
+	      unsigned long n = strtoul (line + 2, &e, 10);
+
+	      if (n <= (evex ? 15 : 31)
+		  && (*e == '.' || is_space_char (*e)))
+		{
+		  i.insn_opcode_space = n;
+		  line = e;
+		}
+	    }
+	  break;
+	}
+    }
+
+  if (line > end && *line == '.' && line[1] == 'W')
+    {
+      /* VEX.W, XOP.W, EVEX.W  */
+      switch (line[2])
+	{
+	case '0':
+	  i.tm.opcode_modifier.vexw = VEXW0;
+	  break;
+
+	case '1':
+	  i.tm.opcode_modifier.vexw = VEXW1;
+	  break;
+
+	case 'I':
+	  if (line[3] == 'G')
+	    {
+	      i.tm.opcode_modifier.vexw = VEXWIG;
+	      ++line;
+	    }
+	  break;
+	}
+
+      if (i.tm.opcode_modifier.vexw)
+	line += 3;
+    }
+
+  if (line > end && *line && !is_space_char (*line))
+    {
+      /* Improve diagnostic a little.  */
+      if (*line == '.' && line[1] && !is_space_char (line[1]))
+	++line;
+      goto done;
     }
 
   input_line_pointer = line;
@@ -10812,24 +11017,30 @@  s_insn (int dummy ATTRIBUTE_UNUSED)
     }
 
   /* Trim off encoding space.  */
-  if (j > 1 && !i.tm.opcode_space && (val >> ((j - 1) * 8)) == 0x0f)
+  if (j > 1 && !i.insn_opcode_space && (val >> ((j - 1) * 8)) == 0x0f)
     {
       uint8_t byte = val >> ((--j - 1) * 8);
 
-      i.tm.opcode_space = SPACE_0F;
+      i.insn_opcode_space = SPACE_0F;
       switch (byte & -(j > 1))
 	{
 	case 0x38:
-	  i.tm.opcode_space = SPACE_0F38;
+	  i.insn_opcode_space = SPACE_0F38;
 	  --j;
 	  break;
 	case 0x3a:
-	  i.tm.opcode_space = SPACE_0F3A;
+	  i.insn_opcode_space = SPACE_0F3A;
 	  --j;
 	  break;
 	}
+      i.tm.opcode_space = i.insn_opcode_space;
       val &= ((uint64_t)1 << (j * 8)) - 1;
     }
+  if (!i.tm.opcode_space && (vex || evex))
+    /* Arrange for build_vex_prefix() to properly emit 0xC4/0xC5.
+       Also avoid hitting abort() there or in build_evex_prefix().  */
+    i.tm.opcode_space = i.insn_opcode_space == SPACE_0F ? SPACE_0F
+						   : SPACE_0F38;
 
   if (j > 2)
     {
@@ -10839,12 +11050,33 @@  s_insn (int dummy ATTRIBUTE_UNUSED)
   i.opcode_length = j;
   i.tm.base_opcode = val;
 
+  if (vex || xop)
+    {
+      if (!i.tm.opcode_modifier.vex)
+	i.tm.opcode_modifier.vex = VEXScalar; /* LIG */
+
+      build_vex_prefix (NULL);
+      i.rex &= REX_OPCODE;
+    }
+  else if (evex)
+    {
+      if (!i.tm.opcode_modifier.evex)
+	i.tm.opcode_modifier.evex = EVEXLIG;
+
+      build_evex_prefix ();
+      i.rex &= REX_OPCODE;
+    }
+
   output_insn ();
 
+ done:
   *saved_ilp = saved_char;
   input_line_pointer = line;
 
   demand_empty_rest_of_line ();
+
+  /* Make sure dot_insn() won't yield "true" anymore.  */
+  i.tm.mnem_off = 0;
 }
 
 #ifdef TE_PE
--- a/gas/testsuite/gas/i386/insn-32.d
+++ b/gas/testsuite/gas/i386/insn-32.d
@@ -11,4 +11,6 @@  Disassembly of section .text:
 [ 	]*[a-f0-9]+:	f3 90[ 	]+pause
 [ 	]*[a-f0-9]+:	d9 ee[ 	]+fldz
 [ 	]*[a-f0-9]+:	f3 0f 01 e8[ 	]+setssbsy
+[ 	]*[a-f0-9]+:	c5 fc 77[ 	]+vzeroall
+[ 	]*[a-f0-9]+:	c4 e1 7c 77[ 	]+vzeroall
 #pass
--- a/gas/testsuite/gas/i386/insn-32.s
+++ b/gas/testsuite/gas/i386/insn-32.s
@@ -12,3 +12,7 @@  insn:
 
 	# setssbsy
 	.insn 0xf30f01e8
+
+	# vzeroall
+	.insn VEX.256.0F.WIG 0x77
+	.insn {vex3} VEX.L1 0x0f77
--- a/gas/testsuite/gas/i386/insn-64.d
+++ b/gas/testsuite/gas/i386/insn-64.d
@@ -11,4 +11,6 @@  Disassembly of section .text:
 [ 	]*[a-f0-9]+:	f3 90[ 	]+pause
 [ 	]*[a-f0-9]+:	d9 ee[ 	]+fldz
 [ 	]*[a-f0-9]+:	f3 0f 01 e8[ 	]+setssbsy
+[ 	]*[a-f0-9]+:	c5 fc 77[ 	]+vzeroall
+[ 	]*[a-f0-9]+:	c4 e1 7c 77[ 	]+vzeroall
 #pass
--- a/gas/testsuite/gas/i386/insn-64.s
+++ b/gas/testsuite/gas/i386/insn-64.s
@@ -12,3 +12,7 @@  insn:
 
 	# setssbsy
 	.insn 0xf30f01e8
+
+	# vzeroall
+	.insn VEX.256.0F.WIG 0x77
+	.insn {vex3} VEX.L1 0x0f77
--- a/opcodes/i386-opc.h
+++ b/opcodes/i386-opc.h
@@ -643,12 +643,14 @@  enum
 	3: 256bit EVEX prefix.
 	4: Length-ignored (LIG) EVEX prefix.
 	5: Length determined from actual operands.
+	6: L'L = 3 (reserved, .insn only)
    */
 #define EVEX512                1
 #define EVEX128                2
 #define EVEX256                3
 #define EVEXLIG                4
 #define EVEXDYN                5
+#define EVEX_L3                6
   EVex,
 
   /* AVX512 masking support: