gas/

2006-02-27  H.J. Lu <hongjiu.lu@intel.com>

	* gas/config/tc-i386.c (output_insn): Support Intel Merom New
	Instructions.

	* gas/config/tc-i386.h (CpuMNI): New.
	(CpuUnknownFlags): Add CpuMNI.

gas/testsuite/

2006-02-27  H.J. Lu <hongjiu.lu@intel.com>

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

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

include/opcode/

2006-02-27  H.J. Lu <hongjiu.lu@intel.com>

	* i386.h (i386_optab): Support Intel Merom New Instructions.

opcodes/

2006-02-27  H.J. Lu <hongjiu.lu@intel.com>

	* i386-dis.c (IS_3BYTE_OPCODE): New for 3-byte opcodes used by
	Intel Merom New Instructions.
	(THREE_BYTE_0): Likewise.
	(THREE_BYTE_1): Likewise.
	(three_byte_table): Likewise.
	(dis386_twobyte): Use THREE_BYTE_0 for entry 0x38. Use
	THREE_BYTE_1 for entry 0x3a.
	(twobyte_has_modrm): Updated.
	(twobyte_uses_SSE_prefix): Likewise.
	(print_insn): Handle 3-byte opcodes used by Intel Merom New
	Instructions.

9 files changed, 330 insertions, 10 deletions

diff --git a/gas/ChangeLog b/gas/ChangeLog
index 28b5575..8f51578 100644
--- a/gas/ChangeLog
+++ b/gas/ChangeLog
@@ -1,3 +1,11 @@
+2006-02-27  H.J. Lu <hongjiu.lu@intel.com>
+
+	* gas/config/tc-i386.c (output_insn): Support Intel Merom New
+	Instructions.
+
+	* gas/config/tc-i386.h (CpuMNI): New.
+	(CpuUnknownFlags): Add CpuMNI.
+
 2006-02-24  David S. Miller  <davem@sunset.davemloft.net>
 
 	* config/tc-sparc.c (priv_reg_table): Add entry for "gl".
diff --git a/gas/config/tc-i386.c b/gas/config/tc-i386.c
index 337149e..f42a9a4 100644
--- a/gas/config/tc-i386.c
+++ b/gas/config/tc-i386.c
@@ -3474,23 +3474,31 @@ output_insn ()
       /* Output normal instructions here.  */
       char *p;
       unsigned char *q;
+      unsigned int prefix;
 
-      /* All opcodes on i386 have either 1 or 2 bytes.  We may use one
-	 more higher byte to specify a prefix the instruction
-	 requires.  */
-      if ((i.tm.base_opcode & 0xff0000) != 0)
+      /* All opcodes on i386 have either 1 or 2 bytes.  Merom New
+	 Instructions have 3 bytes.  We may use one more higher byte
+	 to specify a prefix the instruction requires.  */
+      if ((i.tm.cpu_flags & CpuMNI) != 0)
 	{
+	  if (i.tm.base_opcode & 0xff000000)
+	    {
+	      prefix = (i.tm.base_opcode >> 24) & 0xff;
+	      goto check_prefix;
+	    }
+	}
+      else if ((i.tm.base_opcode & 0xff0000) != 0)
+	{
+	  prefix = (i.tm.base_opcode >> 16) & 0xff;
 	  if ((i.tm.cpu_flags & CpuPadLock) != 0)
 	    {
-	      unsigned int prefix;
-	      prefix = (i.tm.base_opcode >> 16) & 0xff;
-
+check_prefix:
 	      if (prefix != REPE_PREFIX_OPCODE
 		  || i.prefix[LOCKREP_PREFIX] != REPE_PREFIX_OPCODE)
 		add_prefix (prefix);
 	    }
 	  else
-	    add_prefix ((i.tm.base_opcode >> 16) & 0xff);
+	    add_prefix (prefix);
 	}
 
       /* The prefix bytes.  */
@@ -3512,7 +3520,13 @@ output_insn ()
 	}
       else
 	{
-	  p = frag_more (2);
+	  if ((i.tm.cpu_flags & CpuMNI) != 0)
+	    {
+	      p = frag_more (3);
+	      *p++ = (i.tm.base_opcode >> 16) & 0xff;
+	    }
+	  else
+	    p = frag_more (2);
 
 	  /* Put out high byte first: can't use md_number_to_chars!  */
 	  *p++ = (i.tm.base_opcode >> 8) & 0xff;
diff --git a/gas/config/tc-i386.h b/gas/config/tc-i386.h
index 0a9cd1c..9851704 100644
--- a/gas/config/tc-i386.h
+++ b/gas/config/tc-i386.h
@@ -184,6 +184,7 @@ typedef struct
 #define CpuPadLock    0x40000	/* VIA PadLock required */
 #define CpuSVME	      0x80000	/* AMD Secure Virtual Machine Ext-s required */
 #define CpuVMX	     0x100000	/* VMX Instructions required */
+#define CpuMNI	     0x200000	/* Merom New Instructions required */
 
   /* These flags are set by gas depending on the flag_code.  */
 #define Cpu64	     0x4000000   /* 64bit support required  */
@@ -192,7 +193,7 @@ typedef struct
   /* The default value for unknown CPUs - enable all features to avoid problems.  */
 #define CpuUnknownFlags (Cpu086|Cpu186|Cpu286|Cpu386|Cpu486|Cpu586|Cpu686 \
 	|CpuP4|CpuSledgehammer|CpuMMX|CpuMMX2|CpuSSE|CpuSSE2|CpuPNI|CpuVMX \
-	|Cpu3dnow|Cpu3dnowA|CpuK6|CpuAthlon|CpuPadLock|CpuSVME)
+	|Cpu3dnow|Cpu3dnowA|CpuK6|CpuAthlon|CpuPadLock|CpuSVME|CpuMNI)
 
   /* the bits in opcode_modifier are used to generate the final opcode from
      the base_opcode.  These bits also are used to detect alternate forms of
diff --git a/gas/testsuite/ChangeLog b/gas/testsuite/ChangeLog
index c167a21..53c6926 100644
--- a/gas/testsuite/ChangeLog
+++ b/gas/testsuite/ChangeLog
@@ -1,3 +1,12 @@
+2006-02-27  H.J. Lu <hongjiu.lu@intel.com>
+
+	* gas/i386/i386.exp: Add merom and x86-64-merom.
+
+	* gas/i386/merom.d: New file.
+	* gas/i386/merom.s: Likewise.
+	* gas/i386/x86-64-merom.d: Likewise.
+	* gas/i386/x86-64-merom.s: Likewise.
+
 2006-02-24  David S. Miller  <davem@sunset.davemloft.net>
 
 	* gas/sparc/rdhpr.s: New test.
diff --git a/gas/testsuite/gas/i386/i386.exp b/gas/testsuite/gas/i386/i386.exp
index 7bf4595..2138593 100644
--- a/gas/testsuite/gas/i386/i386.exp
+++ b/gas/testsuite/gas/i386/i386.exp
@@ -68,6 +68,7 @@ if [expr ([istarget "i*86-*-*"] ||  [istarget "x86_64-*-*"]) && [gas_32_check]]
     run_dump_test "crx"
     run_list_test "cr-err" ""
     run_dump_test "svme"
+    run_dump_test "merom"
 
     # These tests require support for 8 and 16 bit relocs,
     # so we only run them for ELF and COFF targets.
@@ -136,6 +137,7 @@ if [expr ([istarget "i*86-*-*"] || [istarget "x86_64-*-*"]) && [gas_64_check]] t
     run_dump_test "x86-64-crx-suffix"
     run_dump_test "x86-64-drx"
     run_dump_test "x86-64-drx-suffix"
+    run_dump_test "x86-64-merom"
 
     if { ![istarget "*-*-aix*"]
       && ![istarget "*-*-beos*"]
diff --git a/gas/testsuite/gas/i386/merom.d b/gas/testsuite/gas/i386/merom.d
new file mode 100644
index 0000000..a09721c
--- /dev/null
+++ b/gas/testsuite/gas/i386/merom.d
@@ -0,0 +1,73 @@
+#objdump: -dw
+#name: i386 merom
+
+.*: +file format .*
+
+Disassembly of section .text:
+
+0+000 <foo>:
+   0:	0f 38 01 01[ 	]+phaddw \(%ecx\),%mm0
+   4:	0f 38 01 c1[ 	]+phaddw %mm1,%mm0
+   8:	66 0f 38 01 01[ 	]+phaddw \(%ecx\),%xmm0
+   d:	66 0f 38 01 c1[ 	]+phaddw %xmm1,%xmm0
+  12:	0f 38 02 01[ 	]+phaddd \(%ecx\),%mm0
+  16:	0f 38 02 c1[ 	]+phaddd %mm1,%mm0
+  1a:	66 0f 38 02 01[ 	]+phaddd \(%ecx\),%xmm0
+  1f:	66 0f 38 02 c1[ 	]+phaddd %xmm1,%xmm0
+  24:	0f 38 03 01[ 	]+phaddsw \(%ecx\),%mm0
+  28:	0f 38 03 c1[ 	]+phaddsw %mm1,%mm0
+  2c:	66 0f 38 03 01[ 	]+phaddsw \(%ecx\),%xmm0
+  31:	66 0f 38 03 c1[ 	]+phaddsw %xmm1,%xmm0
+  36:	0f 38 05 01[ 	]+phsubw \(%ecx\),%mm0
+  3a:	0f 38 05 c1[ 	]+phsubw %mm1,%mm0
+  3e:	66 0f 38 05 01[ 	]+phsubw \(%ecx\),%xmm0
+  43:	66 0f 38 05 c1[ 	]+phsubw %xmm1,%xmm0
+  48:	0f 38 06 01[ 	]+phsubd \(%ecx\),%mm0
+  4c:	0f 38 06 c1[ 	]+phsubd %mm1,%mm0
+  50:	66 0f 38 06 01[ 	]+phsubd \(%ecx\),%xmm0
+  55:	66 0f 38 06 c1[ 	]+phsubd %xmm1,%xmm0
+  5a:	0f 38 07 01[ 	]+phsubsw \(%ecx\),%mm0
+  5e:	0f 38 07 c1[ 	]+phsubsw %mm1,%mm0
+  62:	66 0f 38 07 01[ 	]+phsubsw \(%ecx\),%xmm0
+  67:	66 0f 38 07 c1[ 	]+phsubsw %xmm1,%xmm0
+  6c:	0f 38 04 01[ 	]+pmaddubsw \(%ecx\),%mm0
+  70:	0f 38 04 c1[ 	]+pmaddubsw %mm1,%mm0
+  74:	66 0f 38 04 01[ 	]+pmaddubsw \(%ecx\),%xmm0
+  79:	66 0f 38 04 c1[ 	]+pmaddubsw %xmm1,%xmm0
+  7e:	0f 38 0b 01[ 	]+pmulhrsw \(%ecx\),%mm0
+  82:	0f 38 0b c1[ 	]+pmulhrsw %mm1,%mm0
+  86:	66 0f 38 0b 01[ 	]+pmulhrsw \(%ecx\),%xmm0
+  8b:	66 0f 38 0b c1[ 	]+pmulhrsw %xmm1,%xmm0
+  90:	0f 38 00 01[ 	]+pshufb \(%ecx\),%mm0
+  94:	0f 38 00 c1[ 	]+pshufb %mm1,%mm0
+  98:	66 0f 38 00 01[ 	]+pshufb \(%ecx\),%xmm0
+  9d:	66 0f 38 00 c1[ 	]+pshufb %xmm1,%xmm0
+  a2:	0f 38 08 01[ 	]+psignb \(%ecx\),%mm0
+  a6:	0f 38 08 c1[ 	]+psignb %mm1,%mm0
+  aa:	66 0f 38 08 01[ 	]+psignb \(%ecx\),%xmm0
+  af:	66 0f 38 08 c1[ 	]+psignb %xmm1,%xmm0
+  b4:	0f 38 09 01[ 	]+psignw \(%ecx\),%mm0
+  b8:	0f 38 09 c1[ 	]+psignw %mm1,%mm0
+  bc:	66 0f 38 09 01[ 	]+psignw \(%ecx\),%xmm0
+  c1:	66 0f 38 09 c1[ 	]+psignw %xmm1,%xmm0
+  c6:	0f 38 0a 01[ 	]+psignd \(%ecx\),%mm0
+  ca:	0f 38 0a c1[ 	]+psignd %mm1,%mm0
+  ce:	66 0f 38 0a 01[ 	]+psignd \(%ecx\),%xmm0
+  d3:	66 0f 38 0a c1[ 	]+psignd %xmm1,%xmm0
+  d8:	0f 3a 0f 01 02[ 	]+palignr \$0x2,\(%ecx\),%mm0
+  dd:	0f 3a 0f c1 02[ 	]+palignr \$0x2,%mm1,%mm0
+  e2:	66 0f 3a 0f 01 02[ 	]+palignr \$0x2,\(%ecx\),%xmm0
+  e8:	66 0f 3a 0f c1 02[ 	]+palignr \$0x2,%xmm1,%xmm0
+  ee:	0f 38 1c 01[ 	]+pabsb  \(%ecx\),%mm0
+  f2:	0f 38 1c c1[ 	]+pabsb  %mm1,%mm0
+  f6:	66 0f 38 1c 01[ 	]+pabsb  \(%ecx\),%xmm0
+  fb:	66 0f 38 1c c1[ 	]+pabsb  %xmm1,%xmm0
+ 100:	0f 38 1d 01[ 	]+pabsw  \(%ecx\),%mm0
+ 104:	0f 38 1d c1[ 	]+pabsw  %mm1,%mm0
+ 108:	66 0f 38 1d 01[ 	]+pabsw  \(%ecx\),%xmm0
+ 10d:	66 0f 38 1d c1[ 	]+pabsw  %xmm1,%xmm0
+ 112:	0f 38 1e 01[ 	]+pabsd  \(%ecx\),%mm0
+ 116:	0f 38 1e c1[ 	]+pabsd  %mm1,%mm0
+ 11a:	66 0f 38 1e 01[ 	]+pabsd  \(%ecx\),%xmm0
+ 11f:	66 0f 38 1e c1[ 	]+pabsd  %xmm1,%xmm0
+	...
diff --git a/gas/testsuite/gas/i386/merom.s b/gas/testsuite/gas/i386/merom.s
new file mode 100644
index 0000000..154d2f8
--- /dev/null
+++ b/gas/testsuite/gas/i386/merom.s
@@ -0,0 +1,70 @@
+#Merom New Instructions
+
+	.text
+foo:
+	phaddw		(%ecx),%mm0
+	phaddw		%mm1,%mm0
+	phaddw		(%ecx),%xmm0
+	phaddw		%xmm1,%xmm0
+	phaddd		(%ecx),%mm0
+	phaddd		%mm1,%mm0
+	phaddd		(%ecx),%xmm0
+	phaddd		%xmm1,%xmm0
+	phaddsw		(%ecx),%mm0
+	phaddsw		%mm1,%mm0
+	phaddsw		(%ecx),%xmm0
+	phaddsw		%xmm1,%xmm0
+	phsubw		(%ecx),%mm0
+	phsubw		%mm1,%mm0
+	phsubw		(%ecx),%xmm0
+	phsubw		%xmm1,%xmm0
+	phsubd		(%ecx),%mm0
+	phsubd		%mm1,%mm0
+	phsubd		(%ecx),%xmm0
+	phsubd		%xmm1,%xmm0
+	phsubsw		(%ecx),%mm0
+	phsubsw		%mm1,%mm0
+	phsubsw		(%ecx),%xmm0
+	phsubsw		%xmm1,%xmm0
+	pmaddubsw	(%ecx),%mm0
+	pmaddubsw	%mm1,%mm0
+	pmaddubsw	(%ecx),%xmm0
+	pmaddubsw	%xmm1,%xmm0
+	pmulhrsw	(%ecx),%mm0
+	pmulhrsw	%mm1,%mm0
+	pmulhrsw	(%ecx),%xmm0
+	pmulhrsw	%xmm1,%xmm0
+	pshufb		(%ecx),%mm0
+	pshufb		%mm1,%mm0
+	pshufb		(%ecx),%xmm0
+	pshufb		%xmm1,%xmm0
+	psignb		(%ecx),%mm0
+	psignb		%mm1,%mm0
+	psignb		(%ecx),%xmm0
+	psignb		%xmm1,%xmm0
+	psignw		(%ecx),%mm0
+	psignw		%mm1,%mm0
+	psignw		(%ecx),%xmm0
+	psignw		%xmm1,%xmm0
+	psignd		(%ecx),%mm0
+	psignd		%mm1,%mm0
+	psignd		(%ecx),%xmm0
+	psignd		%xmm1,%xmm0
+	palignr		$0x2,(%ecx),%mm0
+	palignr		$0x2,%mm1,%mm0
+	palignr		$0x2,(%ecx),%xmm0
+	palignr		$0x2,%xmm1,%xmm0
+	pabsb		(%ecx),%mm0
+	pabsb		%mm1,%mm0
+	pabsb		(%ecx),%xmm0
+	pabsb		%xmm1,%xmm0
+	pabsw		(%ecx),%mm0
+	pabsw		%mm1,%mm0
+	pabsw		(%ecx),%xmm0
+	pabsw		%xmm1,%xmm0
+	pabsd		(%ecx),%mm0
+	pabsd		%mm1,%mm0
+	pabsd		(%ecx),%xmm0
+	pabsd		%xmm1,%xmm0
+
+	.p2align	4,0
diff --git a/gas/testsuite/gas/i386/x86-64-merom.d b/gas/testsuite/gas/i386/x86-64-merom.d
new file mode 100644
index 0000000..f15a6e4
--- /dev/null
+++ b/gas/testsuite/gas/i386/x86-64-merom.d
@@ -0,0 +1,73 @@
+#objdump: -dw
+#name: x86-64 merom
+
+.*: +file format .*
+
+Disassembly of section .text:
+
+0+000 <foo>:
+   0:	0f 38 01 01[ 	]+phaddw \(%rcx\),%mm0
+   4:	0f 38 01 c1[ 	]+phaddw %mm1,%mm0
+   8:	66 0f 38 01 01[ 	]+phaddw \(%rcx\),%xmm0
+   d:	66 0f 38 01 c1[ 	]+phaddw %xmm1,%xmm0
+  12:	0f 38 02 01[ 	]+phaddd \(%rcx\),%mm0
+  16:	0f 38 02 c1[ 	]+phaddd %mm1,%mm0
+  1a:	66 0f 38 02 01[ 	]+phaddd \(%rcx\),%xmm0
+  1f:	66 0f 38 02 c1[ 	]+phaddd %xmm1,%xmm0
+  24:	0f 38 03 01[ 	]+phaddsw \(%rcx\),%mm0
+  28:	0f 38 03 c1[ 	]+phaddsw %mm1,%mm0
+  2c:	66 0f 38 03 01[ 	]+phaddsw \(%rcx\),%xmm0
+  31:	66 0f 38 03 c1[ 	]+phaddsw %xmm1,%xmm0
+  36:	0f 38 05 01[ 	]+phsubw \(%rcx\),%mm0
+  3a:	0f 38 05 c1[ 	]+phsubw %mm1,%mm0
+  3e:	66 0f 38 05 01[ 	]+phsubw \(%rcx\),%xmm0
+  43:	66 0f 38 05 c1[ 	]+phsubw %xmm1,%xmm0
+  48:	0f 38 06 01[ 	]+phsubd \(%rcx\),%mm0
+  4c:	0f 38 06 c1[ 	]+phsubd %mm1,%mm0
+  50:	66 0f 38 06 01[ 	]+phsubd \(%rcx\),%xmm0
+  55:	66 0f 38 06 c1[ 	]+phsubd %xmm1,%xmm0
+  5a:	0f 38 07 01[ 	]+phsubsw \(%rcx\),%mm0
+  5e:	0f 38 07 c1[ 	]+phsubsw %mm1,%mm0
+  62:	66 0f 38 07 01[ 	]+phsubsw \(%rcx\),%xmm0
+  67:	66 0f 38 07 c1[ 	]+phsubsw %xmm1,%xmm0
+  6c:	0f 38 04 01[ 	]+pmaddubsw \(%rcx\),%mm0
+  70:	0f 38 04 c1[ 	]+pmaddubsw %mm1,%mm0
+  74:	66 0f 38 04 01[ 	]+pmaddubsw \(%rcx\),%xmm0
+  79:	66 0f 38 04 c1[ 	]+pmaddubsw %xmm1,%xmm0
+  7e:	0f 38 0b 01[ 	]+pmulhrsw \(%rcx\),%mm0
+  82:	0f 38 0b c1[ 	]+pmulhrsw %mm1,%mm0
+  86:	66 0f 38 0b 01[ 	]+pmulhrsw \(%rcx\),%xmm0
+  8b:	66 0f 38 0b c1[ 	]+pmulhrsw %xmm1,%xmm0
+  90:	0f 38 00 01[ 	]+pshufb \(%rcx\),%mm0
+  94:	0f 38 00 c1[ 	]+pshufb %mm1,%mm0
+  98:	66 0f 38 00 01[ 	]+pshufb \(%rcx\),%xmm0
+  9d:	66 0f 38 00 c1[ 	]+pshufb %xmm1,%xmm0
+  a2:	0f 38 08 01[ 	]+psignb \(%rcx\),%mm0
+  a6:	0f 38 08 c1[ 	]+psignb %mm1,%mm0
+  aa:	66 0f 38 08 01[ 	]+psignb \(%rcx\),%xmm0
+  af:	66 0f 38 08 c1[ 	]+psignb %xmm1,%xmm0
+  b4:	0f 38 09 01[ 	]+psignw \(%rcx\),%mm0
+  b8:	0f 38 09 c1[ 	]+psignw %mm1,%mm0
+  bc:	66 0f 38 09 01[ 	]+psignw \(%rcx\),%xmm0
+  c1:	66 0f 38 09 c1[ 	]+psignw %xmm1,%xmm0
+  c6:	0f 38 0a 01[ 	]+psignd \(%rcx\),%mm0
+  ca:	0f 38 0a c1[ 	]+psignd %mm1,%mm0
+  ce:	66 0f 38 0a 01[ 	]+psignd \(%rcx\),%xmm0
+  d3:	66 0f 38 0a c1[ 	]+psignd %xmm1,%xmm0
+  d8:	0f 3a 0f 01 02[ 	]+palignr \$0x2,\(%rcx\),%mm0
+  dd:	0f 3a 0f c1 02[ 	]+palignr \$0x2,%mm1,%mm0
+  e2:	66 0f 3a 0f 01 02[ 	]+palignr \$0x2,\(%rcx\),%xmm0
+  e8:	66 0f 3a 0f c1 02[ 	]+palignr \$0x2,%xmm1,%xmm0
+  ee:	0f 38 1c 01[ 	]+pabsb  \(%rcx\),%mm0
+  f2:	0f 38 1c c1[ 	]+pabsb  %mm1,%mm0
+  f6:	66 0f 38 1c 01[ 	]+pabsb  \(%rcx\),%xmm0
+  fb:	66 0f 38 1c c1[ 	]+pabsb  %xmm1,%xmm0
+ 100:	0f 38 1d 01[ 	]+pabsw  \(%rcx\),%mm0
+ 104:	0f 38 1d c1[ 	]+pabsw  %mm1,%mm0
+ 108:	66 0f 38 1d 01[ 	]+pabsw  \(%rcx\),%xmm0
+ 10d:	66 0f 38 1d c1[ 	]+pabsw  %xmm1,%xmm0
+ 112:	0f 38 1e 01[ 	]+pabsd  \(%rcx\),%mm0
+ 116:	0f 38 1e c1[ 	]+pabsd  %mm1,%mm0
+ 11a:	66 0f 38 1e 01[ 	]+pabsd  \(%rcx\),%xmm0
+ 11f:	66 0f 38 1e c1[ 	]+pabsd  %xmm1,%xmm0
+	...
diff --git a/gas/testsuite/gas/i386/x86-64-merom.s b/gas/testsuite/gas/i386/x86-64-merom.s
new file mode 100644
index 0000000..a70654d
--- /dev/null
+++ b/gas/testsuite/gas/i386/x86-64-merom.s
@@ -0,0 +1,70 @@
+#Merom New Instructions
+
+	.text
+foo:
+	phaddw		(%rcx),%mm0
+	phaddw		%mm1,%mm0
+	phaddw		(%rcx),%xmm0
+	phaddw		%xmm1,%xmm0
+	phaddd		(%rcx),%mm0
+	phaddd		%mm1,%mm0
+	phaddd		(%rcx),%xmm0
+	phaddd		%xmm1,%xmm0
+	phaddsw		(%rcx),%mm0
+	phaddsw		%mm1,%mm0
+	phaddsw		(%rcx),%xmm0
+	phaddsw		%xmm1,%xmm0
+	phsubw		(%rcx),%mm0
+	phsubw		%mm1,%mm0
+	phsubw		(%rcx),%xmm0
+	phsubw		%xmm1,%xmm0
+	phsubd		(%rcx),%mm0
+	phsubd		%mm1,%mm0
+	phsubd		(%rcx),%xmm0
+	phsubd		%xmm1,%xmm0
+	phsubsw		(%rcx),%mm0
+	phsubsw		%mm1,%mm0
+	phsubsw		(%rcx),%xmm0
+	phsubsw		%xmm1,%xmm0
+	pmaddubsw	(%rcx),%mm0
+	pmaddubsw	%mm1,%mm0
+	pmaddubsw	(%rcx),%xmm0
+	pmaddubsw	%xmm1,%xmm0
+	pmulhrsw	(%rcx),%mm0
+	pmulhrsw	%mm1,%mm0
+	pmulhrsw	(%rcx),%xmm0
+	pmulhrsw	%xmm1,%xmm0
+	pshufb		(%rcx),%mm0
+	pshufb		%mm1,%mm0
+	pshufb		(%rcx),%xmm0
+	pshufb		%xmm1,%xmm0
+	psignb		(%rcx),%mm0
+	psignb		%mm1,%mm0
+	psignb		(%rcx),%xmm0
+	psignb		%xmm1,%xmm0
+	psignw		(%rcx),%mm0
+	psignw		%mm1,%mm0
+	psignw		(%rcx),%xmm0
+	psignw		%xmm1,%xmm0
+	psignd		(%rcx),%mm0
+	psignd		%mm1,%mm0
+	psignd		(%rcx),%xmm0
+	psignd		%xmm1,%xmm0
+	palignr		$0x2,(%rcx),%mm0
+	palignr		$0x2,%mm1,%mm0
+	palignr		$0x2,(%rcx),%xmm0
+	palignr		$0x2,%xmm1,%xmm0
+	pabsb		(%rcx),%mm0
+	pabsb		%mm1,%mm0
+	pabsb		(%rcx),%xmm0
+	pabsb		%xmm1,%xmm0
+	pabsw		(%rcx),%mm0
+	pabsw		%mm1,%mm0
+	pabsw		(%rcx),%xmm0
+	pabsw		%xmm1,%xmm0
+	pabsd		(%rcx),%mm0
+	pabsd		%mm1,%mm0
+	pabsd		(%rcx),%xmm0
+	pabsd		%xmm1,%xmm0
+
+	.p2align	4,0