gas/

2007-04-18  H.J. Lu <hongjiu.lu@intel.com>

	* config/tc-i386.c (cpu_arch): Add .sse4.2 and .sse4.
	(match_template): Handle operand size for crc32 in SSE4.2.
	(process_suffix): Handle operand type for crc32 in SSE4.2.
	(output_insn): Support SSE4.2.

gas/testsuite/

2007-04-18  H.J. Lu <hongjiu.lu@intel.com>

	* gas/i386/i386.exp: Add sse4.2 and x86-64-sse4.2.

	* gas/i386/sse4_2.d: New file.
	* gas/i386/sse4_2.s: Likewise.
	* gas/i386/x86-64-sse4_2.d: Likewise.
	* gas/i386/x86-64-sse4_2.s: Likewise.

opcodes/

2007-04-18  H.J. Lu <hongjiu.lu@intel.com>

	* i386-dis.c (CRC32_Fixup): New.
	(PREGRP85, PREGRP86, PREGRP87, PREGRP88, PREGRP89, PREGRP90,
	 PREGRP91): New.
	(threebyte_0x38_uses_DATA_prefix): Updated for SSE4.2.
	(threebyte_0x3a_uses_DATA_prefix): Likewise.
	(prefix_user_table): Add PREGRP85, PREGRP86, PREGRP87,
	PREGRP88, PREGRP89, PREGRP90 and PREGRP91.
	(three_byte_table): Likewise.

	* i386-opc.c (i386_optab): Add SSE4.2 opcodes.

	* gas/config/tc-i386.h (CpuSSE4_2): New.
	(CpuSSE4): Likewise.
	(CpuUnknownFlags): Add CpuSSE4_2.

8 files changed, 216 insertions, 17 deletions

diff --git a/gas/ChangeLog b/gas/ChangeLog
index 05ccfc0..8d5a08f 100644
--- a/gas/ChangeLog
+++ b/gas/ChangeLog
@@ -1,5 +1,12 @@
 2007-04-18  H.J. Lu <hongjiu.lu@intel.com>
 
+	* config/tc-i386.c (cpu_arch): Add .sse4.2 and .sse4.
+	(match_template): Handle operand size for crc32 in SSE4.2.
+	(process_suffix): Handle operand type for crc32 in SSE4.2.
+	(output_insn): Support SSE4.2.
+
+2007-04-18  H.J. Lu <hongjiu.lu@intel.com>
+
 	* config/tc-i386.c (cpu_arch): Add .sse4.1.
 	(process_operands): Adjust implicit operand for blendvpd,
 	blendvps and pblendvb in SSE4.1.
diff --git a/gas/config/tc-i386.c b/gas/config/tc-i386.c
index 6ee4010..0b60387 100644
--- a/gas/config/tc-i386.c
+++ b/gas/config/tc-i386.c
@@ -500,6 +500,10 @@ static const arch_entry cpu_arch[] =
    CpuMMX|CpuMMX2|CpuSSE|CpuSSE2|CpuSSE3|CpuSSSE3},
   {".sse4.1", PROCESSOR_UNKNOWN,
    CpuMMX|CpuMMX2|CpuSSE|CpuSSE2|CpuSSE3|CpuSSSE3|CpuSSE4_1},
+  {".sse4.2", PROCESSOR_UNKNOWN,
+   CpuMMX|CpuMMX2|CpuSSE|CpuSSE2|CpuSSE3|CpuSSSE3|CpuSSE4},
+  {".sse4", PROCESSOR_UNKNOWN,
+   CpuMMX|CpuMMX2|CpuSSE|CpuSSE2|CpuSSE3|CpuSSSE3|CpuSSE4},
   {".3dnow", PROCESSOR_UNKNOWN,
    CpuMMX|Cpu3dnow},
   {".3dnowa", PROCESSOR_UNKNOWN,
@@ -2640,9 +2644,10 @@ match_template (void)
 	      || !MATCH (overlap1, i.types[1], operand_types[1])
 	      /* monitor in SSE3 is a very special case.  The first
 		 register and the second register may have different
-		 sizes.  */
+		 sizes.  The same applies to crc32 in SSE4.2.  */
 	      || !((t->base_opcode == 0x0f01
 		    && t->extension_opcode == 0xc8)
+		   || t->base_opcode == 0xf20f38f1
 		   || CONSISTENT_REGISTER_MATCH (overlap0, i.types[0],
 						 operand_types[0],
 						 overlap1, i.types[1],
@@ -2829,19 +2834,30 @@ process_suffix (void)
 	{
 	  /* We take i.suffix from the last register operand specified,
 	     Destination register type is more significant than source
-	     register type.  */
-	  int op;
-
-	  for (op = i.operands; --op >= 0;)
-	    if ((i.types[op] & Reg)
-		&& !(i.tm.operand_types[op] & InOutPortReg))
-	      {
-		i.suffix = ((i.types[op] & Reg8) ? BYTE_MNEM_SUFFIX :
-			    (i.types[op] & Reg16) ? WORD_MNEM_SUFFIX :
-			    (i.types[op] & Reg64) ? QWORD_MNEM_SUFFIX :
+	     register type.  crc32 in SSE4.2 prefers source register
+	     type. */
+	  if (i.tm.base_opcode == 0xf20f38f1)
+	    {
+	      if ((i.types[0] & Reg))
+		i.suffix = ((i.types[0] & Reg16) ? WORD_MNEM_SUFFIX :
 			    LONG_MNEM_SUFFIX);
-		break;
-	      }
+	    }
+
+	  if (!i.suffix)
+	    {
+	      int op;
+
+	      for (op = i.operands; --op >= 0;)
+		if ((i.types[op] & Reg)
+		    && !(i.tm.operand_types[op] & InOutPortReg))
+		  {
+		    i.suffix = ((i.types[op] & Reg8) ? BYTE_MNEM_SUFFIX :
+				(i.types[op] & Reg16) ? WORD_MNEM_SUFFIX :
+				(i.types[op] & Reg64) ? QWORD_MNEM_SUFFIX :
+				LONG_MNEM_SUFFIX);
+		    break;
+		  }
+	    }
 	}
       else if (i.suffix == BYTE_MNEM_SUFFIX)
 	{
@@ -3929,9 +3945,11 @@ output_insn (void)
       unsigned int prefix;
 
       /* All opcodes on i386 have either 1 or 2 bytes.  SSSE3 and
-	 SSE4.1 instructions have 3 bytes.  We may use one more higher
-	 byte to specify a prefix the instruction requires.  */
-      if ((i.tm.cpu_flags & (CpuSSSE3 | CpuSSE4_1)) != 0)
+	 SSE4 instructions have 3 bytes.  We may use one more higher
+	 byte to specify a prefix the instruction requires.  Exclude
+	 instructions which are in both SSE4 and ABM.  */
+      if ((i.tm.cpu_flags & (CpuSSSE3 | CpuSSE4)) != 0
+	  && (i.tm.cpu_flags & CpuABM) == 0)
 	{
 	  if (i.tm.base_opcode & 0xff000000)
 	    {
@@ -3972,7 +3990,8 @@ output_insn (void)
 	}
       else
 	{
-	  if ((i.tm.cpu_flags & (CpuSSSE3 | CpuSSE4_1)) != 0)
+	  if ((i.tm.cpu_flags & (CpuSSSE3 | CpuSSE4)) != 0
+	      && (i.tm.cpu_flags & CpuABM) == 0)
 	    {
 	      p = frag_more (3);
 	      *p++ = (i.tm.base_opcode >> 16) & 0xff;
diff --git a/gas/testsuite/ChangeLog b/gas/testsuite/ChangeLog
index 0679212..263fde4 100644
--- a/gas/testsuite/ChangeLog
+++ b/gas/testsuite/ChangeLog
@@ -1,5 +1,14 @@
 2007-04-18  H.J. Lu <hongjiu.lu@intel.com>
 
+	* gas/i386/i386.exp: Add sse4.2 and x86-64-sse4.2.
+
+	* gas/i386/sse4_2.d: New file.
+	* gas/i386/sse4_2.s: Likewise.
+	* gas/i386/x86-64-sse4_2.d: Likewise.
+	* gas/i386/x86-64-sse4_2.s: Likewise.
+
+2007-04-18  H.J. Lu <hongjiu.lu@intel.com>
+
 	* gas/i386/i386.exp: Add sse4.1 and x86-64-sse4.1.
 
 	* gas/i386/sse4_1.d: New file.
diff --git a/gas/testsuite/gas/i386/i386.exp b/gas/testsuite/gas/i386/i386.exp
index 30053e5..e701a8d 100644
--- a/gas/testsuite/gas/i386/i386.exp
+++ b/gas/testsuite/gas/i386/i386.exp
@@ -87,6 +87,7 @@ if [expr ([istarget "i*86-*-*"] ||  [istarget "x86_64-*-*"]) && [gas_32_check]]
     run_dump_test "addr16"
     run_dump_test "addr32"
     run_dump_test "sse4_1"
+    run_dump_test "sse4_2"
 
     # These tests require support for 8 and 16 bit relocs,
     # so we only run them for ELF and COFF targets.
@@ -173,6 +174,7 @@ if [expr ([istarget "i*86-*-*"] || [istarget "x86_64-*-*"]) && [gas_64_check]] t
     run_dump_test "x86-64-nops-1-nocona"
     run_dump_test "x86-64-nops-1-merom"
     run_dump_test "x86-64-sse4_1"
+    run_dump_test "x86-64-sse4_2"
 
     if { ![istarget "*-*-aix*"]
       && ![istarget "*-*-beos*"]
diff --git a/gas/testsuite/gas/i386/sse4_2.d b/gas/testsuite/gas/i386/sse4_2.d
new file mode 100644
index 0000000..e6468b3
--- /dev/null
+++ b/gas/testsuite/gas/i386/sse4_2.d
@@ -0,0 +1,37 @@
+#objdump: -dw
+#name: i386 SSE4.2
+
+.*:     file format .*
+
+Disassembly of section .text:
+
+0+000 <foo>:
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 19       	crc32l \(%ecx\),%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f0 d9       	crc32b %cl,%ebx
+[ 	]*[0-9a-f]+:	66 f2 0f 38 f1 d9    	crc32w %cx,%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 d9       	crc32l %ecx,%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f0 19       	crc32b \(%ecx\),%ebx
+[ 	]*[0-9a-f]+:	66 f2 0f 38 f1 19    	crc32w \(%ecx\),%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 19       	crc32l \(%ecx\),%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f0 d9       	crc32b %cl,%ebx
+[ 	]*[0-9a-f]+:	66 f2 0f 38 f1 d9    	crc32w %cx,%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 d9       	crc32l %ecx,%ebx
+[ 	]*[0-9a-f]+:	66 0f 38 37 01       	pcmpgtq \(%ecx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 38 37 c1       	pcmpgtq %xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 61 01 00    	pcmpestri \$0x0,\(%ecx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 61 c1 00    	pcmpestri \$0x0,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 60 01 01    	pcmpestrm \$0x1,\(%ecx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 60 c1 01    	pcmpestrm \$0x1,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 63 01 02    	pcmpistri \$0x2,\(%ecx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 63 c1 02    	pcmpistri \$0x2,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 62 01 03    	pcmpistrm \$0x3,\(%ecx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 62 c1 03    	pcmpistrm \$0x3,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 f3 0f b8 19       	popcnt \(%ecx\),%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 19          	popcnt \(%ecx\),%ebx
+[ 	]*[0-9a-f]+:	66 f3 0f b8 19       	popcnt \(%ecx\),%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 19          	popcnt \(%ecx\),%ebx
+[ 	]*[0-9a-f]+:	66 f3 0f b8 d9       	popcnt %cx,%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 d9          	popcnt %ecx,%ebx
+[ 	]*[0-9a-f]+:	66 f3 0f b8 d9       	popcnt %cx,%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 d9          	popcnt %ecx,%ebx
+#pass
diff --git a/gas/testsuite/gas/i386/sse4_2.s b/gas/testsuite/gas/i386/sse4_2.s
new file mode 100644
index 0000000..584757c
--- /dev/null
+++ b/gas/testsuite/gas/i386/sse4_2.s
@@ -0,0 +1,34 @@
+# Streaming SIMD extensions 4.2 Instructions
+
+	.text
+foo:
+	crc32		(%ecx),%ebx
+	crc32		%cl,%ebx
+	crc32		%cx,%ebx
+	crc32		%ecx,%ebx
+	crc32b		(%ecx),%ebx
+	crc32w		(%ecx),%ebx
+	crc32l		(%ecx),%ebx
+	crc32b		%cl,%ebx
+	crc32w		%cx,%ebx
+	crc32l		%ecx,%ebx
+	pcmpgtq		(%ecx),%xmm0
+	pcmpgtq		%xmm1,%xmm0
+	pcmpestri	$0x0,(%ecx),%xmm0
+	pcmpestri	$0x0,%xmm1,%xmm0
+	pcmpestrm	$0x1,(%ecx),%xmm0
+	pcmpestrm	$0x1,%xmm1,%xmm0
+	pcmpistri	$0x2,(%ecx),%xmm0
+	pcmpistri	$0x2,%xmm1,%xmm0
+	pcmpistrm	$0x3,(%ecx),%xmm0
+	pcmpistrm	$0x3,%xmm1,%xmm0
+	popcnt		(%ecx),%bx
+	popcnt		(%ecx),%ebx
+	popcntw		(%ecx),%bx
+	popcntl		(%ecx),%ebx
+	popcnt		%cx,%bx
+	popcnt		%ecx,%ebx
+	popcntw		%cx,%bx
+	popcntl		%ecx,%ebx
+
+	.p2align	4,0
diff --git a/gas/testsuite/gas/i386/x86-64-sse4_2.d b/gas/testsuite/gas/i386/x86-64-sse4_2.d
new file mode 100644
index 0000000..b541f74
--- /dev/null
+++ b/gas/testsuite/gas/i386/x86-64-sse4_2.d
@@ -0,0 +1,47 @@
+#objdump: -dw
+#name: x86-64 SSE4.2
+
+.*:     file format .*
+
+Disassembly of section .text:
+
+0+000 <foo>:
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 19       	crc32l \(%rcx\),%ebx
+[ 	]*[0-9a-f]+:	f2 48 0f 38 f1 19    	crc32q \(%rcx\),%rbx
+[ 	]*[0-9a-f]+:	f2 0f 38 f0 d9       	crc32b %cl,%ebx
+[ 	]*[0-9a-f]+:	f2 48 0f 38 f0 d9    	crc32b %cl,%rbx
+[ 	]*[0-9a-f]+:	66 f2 0f 38 f1 d9    	crc32w %cx,%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 d9       	crc32l %ecx,%ebx
+[ 	]*[0-9a-f]+:	f2 48 0f 38 f1 d9    	crc32q %rcx,%rbx
+[ 	]*[0-9a-f]+:	f2 0f 38 f0 19       	crc32b \(%rcx\),%ebx
+[ 	]*[0-9a-f]+:	66 f2 0f 38 f1 19    	crc32w \(%rcx\),%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 19       	crc32l \(%rcx\),%ebx
+[ 	]*[0-9a-f]+:	f2 48 0f 38 f1 19    	crc32q \(%rcx\),%rbx
+[ 	]*[0-9a-f]+:	f2 0f 38 f0 d9       	crc32b %cl,%ebx
+[ 	]*[0-9a-f]+:	f2 48 0f 38 f0 d9    	crc32b %cl,%rbx
+[ 	]*[0-9a-f]+:	66 f2 0f 38 f1 d9    	crc32w %cx,%ebx
+[ 	]*[0-9a-f]+:	f2 0f 38 f1 d9       	crc32l %ecx,%ebx
+[ 	]*[0-9a-f]+:	f2 48 0f 38 f1 d9    	crc32q %rcx,%rbx
+[ 	]*[0-9a-f]+:	66 0f 38 37 01       	pcmpgtq \(%rcx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 38 37 c1       	pcmpgtq %xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 61 01 00    	pcmpestri \$0x0,\(%rcx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 61 c1 00    	pcmpestri \$0x0,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 60 01 01    	pcmpestrm \$0x1,\(%rcx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 60 c1 01    	pcmpestrm \$0x1,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 63 01 02    	pcmpistri \$0x2,\(%rcx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 63 c1 02    	pcmpistri \$0x2,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 62 01 03    	pcmpistrm \$0x3,\(%rcx\),%xmm0
+[ 	]*[0-9a-f]+:	66 0f 3a 62 c1 03    	pcmpistrm \$0x3,%xmm1,%xmm0
+[ 	]*[0-9a-f]+:	66 f3 0f b8 19       	popcnt \(%rcx\),%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 19          	popcnt \(%rcx\),%ebx
+[ 	]*[0-9a-f]+:	f3 48 0f b8 19       	popcnt \(%rcx\),%rbx
+[ 	]*[0-9a-f]+:	66 f3 0f b8 19       	popcnt \(%rcx\),%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 19          	popcnt \(%rcx\),%ebx
+[ 	]*[0-9a-f]+:	f3 48 0f b8 19       	popcnt \(%rcx\),%rbx
+[ 	]*[0-9a-f]+:	66 f3 0f b8 d9       	popcnt %cx,%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 d9          	popcnt %ecx,%ebx
+[ 	]*[0-9a-f]+:	f3 48 0f b8 d9       	popcnt %rcx,%rbx
+[ 	]*[0-9a-f]+:	66 f3 0f b8 d9       	popcnt %cx,%bx
+[ 	]*[0-9a-f]+:	f3 0f b8 d9          	popcnt %ecx,%ebx
+[ 	]*[0-9a-f]+:	f3 48 0f b8 d9       	popcnt %rcx,%rbx
+#pass
diff --git a/gas/testsuite/gas/i386/x86-64-sse4_2.s b/gas/testsuite/gas/i386/x86-64-sse4_2.s
new file mode 100644
index 0000000..fdc7491
--- /dev/null
+++ b/gas/testsuite/gas/i386/x86-64-sse4_2.s
@@ -0,0 +1,44 @@
+# Streaming SIMD extensions 4.2 Instructions
+
+	.text
+foo:
+	crc32		(%rcx),%ebx
+	crc32		(%rcx),%rbx
+	crc32		%cl,%ebx
+	crc32		%cl,%rbx
+	crc32		%cx,%ebx
+	crc32		%ecx,%ebx
+	crc32		%rcx,%rbx
+	crc32b		(%rcx),%ebx
+	crc32w		(%rcx),%ebx
+	crc32l		(%rcx),%ebx
+	crc32q		(%rcx),%rbx
+	crc32b		%cl,%ebx
+	crc32b		%cl,%rbx
+	crc32w		%cx,%ebx
+	crc32l		%ecx,%ebx
+	crc32q		%rcx,%rbx
+	pcmpgtq		(%rcx),%xmm0
+	pcmpgtq		%xmm1,%xmm0
+	pcmpestri	$0x0,(%rcx),%xmm0
+	pcmpestri	$0x0,%xmm1,%xmm0
+	pcmpestrm	$0x1,(%rcx),%xmm0
+	pcmpestrm	$0x1,%xmm1,%xmm0
+	pcmpistri	$0x2,(%rcx),%xmm0
+	pcmpistri	$0x2,%xmm1,%xmm0
+	pcmpistrm	$0x3,(%rcx),%xmm0
+	pcmpistrm	$0x3,%xmm1,%xmm0
+	popcnt		(%rcx),%bx
+	popcnt		(%rcx),%ebx
+	popcnt		(%rcx),%rbx
+	popcntw		(%rcx),%bx
+	popcntl		(%rcx),%ebx
+	popcntq		(%rcx),%rbx
+	popcnt		%cx,%bx
+	popcnt		%ecx,%ebx
+	popcnt		%rcx,%rbx
+	popcntw		%cx,%bx
+	popcntl		%ecx,%ebx
+	popcntq		%rcx,%rbx
+
+	.p2align	4,0