add 6(a|l) sse support to 8(a|l)

2 files changed, 233 insertions, 3 deletions

diff --git a/sys/src/cmd/8a/a.y b/sys/src/cmd/8a/a.y
index 7c9487f64..0ec5dd263 100644
--- a/sys/src/cmd/8a/a.y
+++ b/sys/src/cmd/8a/a.y
@@ -20,9 +20,9 @@
 %left	'+' '-'
 %left	'*' '/' '%'
 %token	<lval>	LTYPE0 LTYPE1 LTYPE2 LTYPE3 LTYPE4
-%token	<lval>	LTYPEC LTYPED LTYPEN LTYPER LTYPET LTYPES LTYPEM LTYPEI LTYPEG
+%token	<lval>	LTYPEC LTYPED LTYPEN LTYPER LTYPET LTYPES LTYPEM LTYPEI LTYPEG LTYPEXC LTYPEX
 %token	<lval>	LCONST LFP LPC LSB
-%token	<lval>	LBREG LLREG LSREG LFREG
+%token	<lval>	LBREG LLREG LSREG LFREG LMREG LXREG
 %token	<dval>	LFCONST
 %token	<sval>	LSCONST LSP
 %token	<sym>	LNAME LLAB LVAR
@@ -30,7 +30,7 @@
 %type	<con2>	con2
 %type	<gen>	mem imm imm2 reg nam rel rem rim rom omem nmem
 %type	<gen2>	nonnon nonrel nonrem rimnon rimrem remrim
-%type	<gen2>	spec1 spec2 spec3 spec4 spec5 spec6 spec7 spec8
+%type	<gen2>	spec1 spec2 spec3 spec4 spec5 spec6 spec7 spec8 spec9 spec10
 %%
 prog:
 |	prog line
@@ -79,6 +79,8 @@ inst:
 |	LTYPEM spec6	{ outcode($1, &$2); }
 |	LTYPEI spec7	{ outcode($1, &$2); }
 |	LTYPEG spec8	{ outcode($1, &$2); }
+|	LTYPEXC spec9	{ outcode($1, &$2); }
+|	LTYPEX spec10	{ outcode($1, &$2); }
 
 nonnon:
 	{
@@ -237,6 +239,23 @@ spec8:	/* GLOBL */
 		$$.from.scale = $3;
 		$$.to = $5;
 	}
+spec9:	/* CMPPS/CMPPD */
+	reg ',' rem ',' con
+	{
+		$$.from = $1;
+		$$.to = $3;
+		$$.from.offset = $5;
+	}
+
+spec10:	/* shufl */
+	imm ',' rem ',' reg
+	{
+		$$.from = $3;
+		$$.to = $5;
+		if($1.type != D_CONST)
+			yyerror("illegal constant");
+		$$.to.offset = $1.offset;
+	}
 
 rem:
 	reg
@@ -301,6 +320,11 @@ reg:
 		$$ = nullgen;
 		$$.type = $1;
 	}
+|	LMREG
+	{
+		$$ = nullgen;
+		$$.type = $1;
+	}
 |	LSP
 	{
 		$$ = nullgen;
@@ -311,6 +335,11 @@ reg:
 		$$ = nullgen;
 		$$.type = $1;
 	}
+|	LXREG
+	{
+		$$ = nullgen;
+		$$.type = $1;
+	}
 
 imm:
 	'$' con
diff --git a/sys/src/cmd/8a/lex.c b/sys/src/cmd/8a/lex.c
index 7aaf25737..b5c736c1d 100644
--- a/sys/src/cmd/8a/lex.c
+++ b/sys/src/cmd/8a/lex.c
@@ -192,6 +192,24 @@ struct
 	"F6",		LFREG,	D_F0+6,
 	"F7",		LFREG,	D_F0+7,
 
+	"M0",		LMREG,	D_M0+0,
+	"M1",		LMREG,	D_M0+1,
+	"M2",		LMREG,	D_M0+2,
+	"M3",		LMREG,	D_M0+3,
+	"M4",		LMREG,	D_M0+4,
+	"M5",		LMREG,	D_M0+5,
+	"M6",		LMREG,	D_M0+6,
+	"M7",		LMREG,	D_M0+7,
+
+	"X0",		LXREG,	D_X0+0,
+	"X1",		LXREG,	D_X0+1,
+	"X2",		LXREG,	D_X0+2,
+	"X3",		LXREG,	D_X0+3,
+	"X4",		LXREG,	D_X0+4,
+	"X5",		LXREG,	D_X0+5,
+	"X6",		LXREG,	D_X0+6,
+	"X7",		LXREG,	D_X0+7,
+
 	"CS",		LSREG,	D_CS,
 	"SS",		LSREG,	D_SS,
 	"DS",		LSREG,	D_DS,
@@ -277,6 +295,7 @@ struct
 	"CMPXCHGB",	LTYPE3,	ACMPXCHGB,
 	"CMPXCHGL",	LTYPE3,	ACMPXCHGL,
 	"CMPXCHGW",	LTYPE3,	ACMPXCHGW,
+	"CPUID",	LTYPE0,	ACPUID,
 	"DAA",		LTYPE0,	ADAA,
 	"DAS",		LTYPE0,	ADAS,
 	"DATA",		LTYPED,	ADATA,
@@ -639,6 +658,188 @@ struct
 	"FYL2X",	LTYPE0, AFYL2X,
 	"FYL2XP1",	LTYPE0, AFYL2XP1,
 
+	"ADDPD",	LTYPE3,	AADDPD,
+	"ADDPS",	LTYPE3,	AADDPS,
+	"ADDSD",	LTYPE3,	AADDSD,
+	"ADDSS",	LTYPE3,	AADDSS,
+	"ANDNPD",	LTYPE3,	AANDNPD,
+	"ANDNPS",	LTYPE3,	AANDNPS,
+	"ANDPD",	LTYPE3,	AANDPD,
+	"ANDPS",	LTYPE3,	AANDPS,
+	"CMPPD",	LTYPEXC,ACMPPD,
+	"CMPPS",	LTYPEXC,ACMPPS,
+	"CMPSD",	LTYPEXC,ACMPSD,
+	"CMPSS",	LTYPEXC,ACMPSS,
+	"COMISD",	LTYPE3,	ACOMISD,
+	"COMISS",	LTYPE3,	ACOMISS,
+	"CVTPL2PD",	LTYPE3,	ACVTPL2PD,
+	"CVTPL2PS",	LTYPE3,	ACVTPL2PS,
+	"CVTPD2PL",	LTYPE3,	ACVTPD2PL,
+	"CVTPD2PS",	LTYPE3,	ACVTPD2PS,
+	"CVTPS2PL",	LTYPE3,	ACVTPS2PL,
+	"PF2IW",	LTYPE3,	APF2IW,
+	"PF2IL",	LTYPE3,	APF2IL,
+	"PF2ID",	LTYPE3,	APF2IL,	/* syn */
+	"PI2FL",	LTYPE3,	API2FL,
+	"PI2FD",	LTYPE3,	API2FL,	/* syn */
+	"PI2FW",	LTYPE3,	API2FW,
+	"CVTPS2PD",	LTYPE3,	ACVTPS2PD,
+	"CVTSD2SL",	LTYPE3,	ACVTSD2SL,
+	"CVTSD2SS",	LTYPE3,	ACVTSD2SS,
+	"CVTSL2SD",	LTYPE3,	ACVTSL2SD,
+	"CVTSL2SS",	LTYPE3,	ACVTSL2SS,
+	"CVTSS2SD",	LTYPE3,	ACVTSS2SD,
+	"CVTSS2SL",	LTYPE3,	ACVTSS2SL,
+	"CVTTPD2PL",	LTYPE3,	ACVTTPD2PL,
+	"CVTTPS2PL",	LTYPE3,	ACVTTPS2PL,
+	"CVTTSD2SL",	LTYPE3,	ACVTTSD2SL,
+	"CVTTSS2SL",	LTYPE3,	ACVTTSS2SL,
+	"DIVPD",	LTYPE3,	ADIVPD,
+	"DIVPS",	LTYPE3,	ADIVPS,
+	"DIVSD",	LTYPE3,	ADIVSD,
+	"DIVSS",	LTYPE3,	ADIVSS,
+	"FXRSTOR",	LTYPE2,	AFXRSTOR,
+	"FXSAVE",	LTYPE1,	AFXSAVE,
+	"LDMXCSR",	LTYPE2,	ALDMXCSR,
+	"MASKMOVOU",	LTYPE3,	AMASKMOVOU,
+	"MASKMOVDQU",	LTYPE3,	AMASKMOVOU,	/* syn */
+	"MAXPD",	LTYPE3,	AMAXPD,
+	"MAXPS",	LTYPE3,	AMAXPS,
+	"MAXSD",	LTYPE3,	AMAXSD,
+	"MAXSS",	LTYPE3,	AMAXSS,
+	"MINPD",	LTYPE3,	AMINPD,
+	"MINPS",	LTYPE3,	AMINPS,
+	"MINSD",	LTYPE3,	AMINSD,
+	"MINSS",	LTYPE3,	AMINSS,
+	"MOVAPD",	LTYPE3,	AMOVAPD,
+	"MOVAPS",	LTYPE3,	AMOVAPS,
+	"MOVO",		LTYPE3,	AMOVO,
+	"MOVOA",	LTYPE3,	AMOVO,	/* syn */
+	"MOVOU",	LTYPE3,	AMOVOU,
+	"MOVHLPS",	LTYPE3,	AMOVHLPS,
+	"MOVHPD",	LTYPE3,	AMOVHPD,
+	"MOVHPS",	LTYPE3,	AMOVHPS,
+	"MOVLHPS",	LTYPE3,	AMOVLHPS,
+	"MOVLPD",	LTYPE3,	AMOVLPD,
+	"MOVLPS",	LTYPE3,	AMOVLPS,
+	"MOVMSKPD",	LTYPE3,	AMOVMSKPD,
+	"MOVMSKPS",	LTYPE3,	AMOVMSKPS,
+	"MOVNTO",	LTYPE3,	AMOVNTO,
+	"MOVNTDQ",	LTYPE3,	AMOVNTO,	/* syn */
+	"MOVNTPD",	LTYPE3,	AMOVNTPD,
+	"MOVNTPS",	LTYPE3,	AMOVNTPS,
+	"MOVNTQ",	LTYPE3,	AMOVNTQ,
+	"MOVQOZX",	LTYPE3,	AMOVQOZX,
+	"MOVSD",	LTYPE3,	AMOVSD,
+	"MOVSS",	LTYPE3,	AMOVSS,
+	"MOVUPD",	LTYPE3,	AMOVUPD,
+	"MOVUPS",	LTYPE3,	AMOVUPS,
+	"MULPD",	LTYPE3,	AMULPD,
+	"MULPS",	LTYPE3,	AMULPS,
+	"MULSD",	LTYPE3,	AMULSD,
+	"MULSS",	LTYPE3,	AMULSS,
+	"ORPD",		LTYPE3,	AORPD,
+	"ORPS",		LTYPE3,	AORPS,
+	"PACKSSLW",	LTYPE3,	APACKSSLW,
+	"PACKSSWB",	LTYPE3,	APACKSSWB,
+	"PACKUSWB",	LTYPE3,	APACKUSWB,
+	"PADDB",	LTYPE3,	APADDB,
+	"PADDL",	LTYPE3,	APADDL,
+	"PADDQ",	LTYPE3,	APADDQ,
+	"PADDSB",	LTYPE3,	APADDSB,
+	"PADDSW",	LTYPE3,	APADDSW,
+	"PADDUSB",	LTYPE3,	APADDUSB,
+	"PADDUSW",	LTYPE3,	APADDUSW,
+	"PADDW",	LTYPE3,	APADDW,
+	"PAND",		LTYPE3, APAND,
+	"PANDB",	LTYPE3,	APANDB,
+	"PANDL",	LTYPE3,	APANDL,
+	"PANDSB",	LTYPE3,	APANDSB,
+	"PANDSW",	LTYPE3,	APANDSW,
+	"PANDUSB",	LTYPE3,	APANDUSB,
+	"PANDUSW",	LTYPE3,	APANDUSW,
+	"PANDW",	LTYPE3,	APANDW,
+	"PANDN",	LTYPE3, APANDN,
+	"PAVGB",	LTYPE3,	APAVGB,
+	"PAVGW",	LTYPE3,	APAVGW,
+	"PCMPEQB",	LTYPE3,	APCMPEQB,
+	"PCMPEQL",	LTYPE3,	APCMPEQL,
+	"PCMPEQW",	LTYPE3,	APCMPEQW,
+	"PCMPGTB",	LTYPE3,	APCMPGTB,
+	"PCMPGTL",	LTYPE3,	APCMPGTL,	
+	"PCMPGTW",	LTYPE3,	APCMPGTW,
+	"PEXTRW",	LTYPEX,	APEXTRW,
+	"PINSRW",	LTYPEX,	APINSRW,
+	"PMADDWL",	LTYPE3,	APMADDWL,
+	"PMAXSW",	LTYPE3,	APMAXSW,
+	"PMAXUB",	LTYPE3,	APMAXUB,
+	"PMINSW",	LTYPE3,	APMINSW,
+	"PMINUB",	LTYPE3,	APMINUB,
+	"PMOVMSKB",	LTYPE3,	APMOVMSKB,
+	"PMULHRW",	LTYPE3,	APMULHRW,
+	"PMULHUW",	LTYPE3,	APMULHUW,
+	"PMULHW",	LTYPE3,	APMULHW,
+	"PMULLW",	LTYPE3,	APMULLW,
+	"PMULULQ",	LTYPE3,	APMULULQ,
+	"POR",		LTYPE3,	APOR,
+	"PSADBW",	LTYPE3,	APSADBW,
+	"PSHUFHW",	LTYPEX,	APSHUFHW,
+	"PSHUFL",	LTYPEX,	APSHUFL,
+	"PSHUFLW",	LTYPEX,	APSHUFLW,
+	"PSHUFW",	LTYPEX, APSHUFW,
+	"PSLLO",	LTYPE3,	APSLLO,
+	"PSLLDQ",	LTYPE3,	APSLLO,	/* syn */
+	"PSLLL",	LTYPE3,	APSLLL,
+	"PSLLQ",	LTYPE3,	APSLLQ,
+	"PSLLW",	LTYPE3,	APSLLW,
+	"PSRAL",	LTYPE3,	APSRAL,
+	"PSRAW",	LTYPE3,	APSRAW,
+	"PSRLO",	LTYPE3,	APSRLO,
+	"PSRLDQ",	LTYPE3,	APSRLO,	/* syn */
+	"PSRLL",	LTYPE3,	APSRLL,
+	"PSRLQ",	LTYPE3,	APSRLQ,
+	"PSRLW",	LTYPE3,	APSRLW,
+	"PSUBB",	LTYPE3,	APSUBB,
+	"PSUBL",	LTYPE3,	APSUBL,
+	"PSUBQ",	LTYPE3,	APSUBQ,
+	"PSUBSB",	LTYPE3,	APSUBSB,
+	"PSUBSW",	LTYPE3,	APSUBSW,
+	"PSUBUSB",	LTYPE3,	APSUBUSB,
+	"PSUBUSW",	LTYPE3,	APSUBUSW,
+	"PSUBW",	LTYPE3,	APSUBW,
+	"PUNPCKHBW",	LTYPE3,	APUNPCKHBW,
+	"PUNPCKHLQ",	LTYPE3,	APUNPCKHLQ,
+	"PUNPCKHQDQ",	LTYPE3,	APUNPCKHQDQ,
+	"PUNPCKHWL",	LTYPE3,	APUNPCKHWL,
+	"PUNPCKLBW",	LTYPE3,	APUNPCKLBW,
+	"PUNPCKLLQ",	LTYPE3,	APUNPCKLLQ,
+	"PUNPCKLQDQ",	LTYPE3,	APUNPCKLQDQ,
+	"PUNPCKLWL",	LTYPE3,	APUNPCKLWL,
+	"PXOR",		LTYPE3,	APXOR,
+	"RCPPS",	LTYPE3,	ARCPPS,
+	"RCPSS",	LTYPE3,	ARCPSS,
+	"RSQRTPS",	LTYPE3,	ARSQRTPS,
+	"RSQRTSS",	LTYPE3,	ARSQRTSS,
+	"SHUFPD",	LTYPEX,	ASHUFPD,
+	"SHUFPS",	LTYPEX,	ASHUFPS,
+	"SQRTPD",	LTYPE3,	ASQRTPD,
+	"SQRTPS",	LTYPE3,	ASQRTPS,
+	"SQRTSD",	LTYPE3,	ASQRTSD,
+	"SQRTSS",	LTYPE3,	ASQRTSS,
+	"STMXCSR",	LTYPE1,	ASTMXCSR,
+	"SUBPD",	LTYPE3,	ASUBPD,
+	"SUBPS",	LTYPE3,	ASUBPS,
+	"SUBSD",	LTYPE3,	ASUBSD,
+	"SUBSS",	LTYPE3,	ASUBSS,
+	"UCOMISD",	LTYPE3,	AUCOMISD,
+	"UCOMISS",	LTYPE3,	AUCOMISS,
+	"UNPCKHPD",	LTYPE3,	AUNPCKHPD,
+	"UNPCKHPS",	LTYPE3,	AUNPCKHPS,
+	"UNPCKLPD",	LTYPE3,	AUNPCKLPD,
+	"UNPCKLPS",	LTYPE3,	AUNPCKLPS,
+	"XORPD",	LTYPE3,	AXORPD,
+	"XORPS",	LTYPE3,	AXORPS,
+
 	0
 };