cmd/6a, cmd/6l: add support for AES-NI instrutions and PSHUFD

This CL adds support for the these 7 new instructions to 6a/6l in
preparation of the upcoming CL for AES-NI accelerated crypto/aes:
AESENC, AESENCLAST, AESDEC, AESDECLAST, AESIMC, AESKEYGENASSIST,
and PSHUFD.

R=golang-dev, rsc
CC=golang-dev
https://golang.org/cl/5970055

diff --git a/src/cmd/6a/lex.c b/src/cmd/6a/lex.c
index 6a1c652654cbfcf28fbb27a76b835126bd9b34a8..376a4b67e534e7f4b189d4fb38fa9d89d04f940e 100644 (file)
--- a/src/cmd/6a/lex.c
+++ b/src/cmd/6a/lex.c
@@ -1008,6 +1008,13 @@ struct
        "PREFETCHT2",           LTYPE2, APREFETCHT2,
        "PREFETCHNTA",          LTYPE2, APREFETCHNTA,
        "UNDEF",        LTYPE0, AUNDEF,
+       "AESENC",       LTYPE3, AAESENC,
+       "AESENCLAST",   LTYPE3, AAESENCLAST,
+       "AESDEC",       LTYPE3, AAESDEC,
+       "AESDECLAST",   LTYPE3, AAESDECLAST,
+       "AESIMC",       LTYPE3, AAESIMC,
+       "AESKEYGENASSIST", LTYPEX, AAESKEYGENASSIST,
+       "PSHUFD", LTYPEX, APSHUFD,
 
        0
 };

diff --git a/src/cmd/6l/6.out.h b/src/cmd/6l/6.out.h
index 769df3b5ac01dc1caddc8c0ae1961c4bffaac9d5..6acc1f459edb6e6b1a40939225d50afd72e4ca6d 100644 (file)
--- a/src/cmd/6l/6.out.h
+++ b/src/cmd/6l/6.out.h
@@ -748,6 +748,15 @@ enum       as
        
        AUNDEF,
 
+       AAESENC,
+       AAESENCLAST,
+       AAESDEC,
+       AAESDECLAST,
+       AAESIMC,
+       AAESKEYGENASSIST,
+
+       APSHUFD,
+
        ALAST
 };
 

diff --git a/src/cmd/6l/optab.c b/src/cmd/6l/optab.c
index 0716fa4453fec5e72eb7944535df502ab53af242..e81c1eda0130a20ff809990e78f154b4ac2c08a6 100644 (file)
--- a/src/cmd/6l/optab.c
+++ b/src/cmd/6l/optab.c
@@ -269,7 +269,7 @@ uchar       yimul[] =
 };
 uchar  yimul3[] =
 {
-       Yml,    Yrl,    Zibm_r, 1,
+       Yml,    Yrl,    Zibm_r, 2,
        0
 };
 uchar  ybyte[] =
@@ -518,17 +518,17 @@ uchar     ymrxr[] =
 };
 uchar  ymshuf[] =
 {
-       Ymm,    Ymr,    Zibm_r, 1,
+       Ymm,    Ymr,    Zibm_r, 2,
        0
 };
 uchar  yxshuf[] =
 {
-       Yxm,    Yxr,    Zibm_r, 1,
+       Yxm,    Yxr,    Zibm_r, 2,
        0
 };
 uchar  yextrw[] =
 {
-       Yxr,    Yrl,    Zibm_r, 1,
+       Yxr,    Yrl,    Zibm_r, 2,
        0
 };
 uchar  ypsdq[] =
@@ -551,6 +551,16 @@ uchar      yprefetch[] =
        Ym,     Ynone,  Zm_o,   2,
        0,
 };
+uchar  yaes[] =
+{
+       Yxm,    Yxr,    Zlitm_r,        2,
+       0
+};
+uchar  yaes2[] =
+{
+       Yxm,    Yxr,    Zibm_r, 2,
+       0
+};
 
 /*
  * You are doasm, holding in your hand a Prog* with p->as set to, say, ACRC32,
@@ -791,7 +801,7 @@ Optab optab[] =
        { AIMULL,       yimul,  Px, 0xf7,(05),0x6b,0x69,Pm,0xaf },
        { AIMULQ,       yimul,  Pw, 0xf7,(05),0x6b,0x69,Pm,0xaf },
        { AIMULW,       yimul,  Pe, 0xf7,(05),0x6b,0x69,Pm,0xaf },
-       { AIMUL3Q,      yimul3, Pw, 0x6b },
+       { AIMUL3Q,      yimul3, Pw, 0x6b,(00) },
        { AINB,         yin,    Pb, 0xe4,0xec },
        { AINCB,        yincb,  Pb, 0xfe,(00) },
        { AINCL,        yincl,  Px, 0xff,(00) },
@@ -950,7 +960,7 @@ Optab optab[] =
        { APCMPGTB,     ymm,    Py, 0x64,Pe,0x64 },
        { APCMPGTL,     ymm,    Py, 0x66,Pe,0x66 },
        { APCMPGTW,     ymm,    Py, 0x65,Pe,0x65 },
-       { APEXTRW,      yextrw, Pq, 0xc5 },
+       { APEXTRW,      yextrw, Pq, 0xc5,(00) },
        { APF2IL,       ymfp,   Px, 0x1d },
        { APF2IW,       ymfp,   Px, 0x1c },
        { API2FL,       ymfp,   Px, 0x0d },
@@ -971,7 +981,7 @@ Optab optab[] =
        { APFRSQRT,     ymfp,   Px, 0x97 },
        { APFSUB,       ymfp,   Px, 0x9a },
        { APFSUBR,      ymfp,   Px, 0xaa },
-       { APINSRW,      yextrw, Pq, 0xc4 },
+       { APINSRW,      yextrw, Pq, 0xc4,(00) },
        { APMADDWL,     ymm,    Py, 0xf5,Pe,0xf5 },
        { APMAXSW,      yxm,    Pe, 0xee },
        { APMAXUB,      yxm,    Pe, 0xde },
@@ -993,10 +1003,10 @@ Optab optab[] =
        { APOPW,        ypopl,  Pe, 0x58,0x8f,(00) },
        { APOR,         ymm,    Py, 0xeb,Pe,0xeb },
        { APSADBW,      yxm,    Pq, 0xf6 },
-       { APSHUFHW,     yxshuf, Pf3, 0x70 },
-       { APSHUFL,      yxshuf, Pq, 0x70 },
-       { APSHUFLW,     yxshuf, Pf2, 0x70 },
-       { APSHUFW,      ymshuf, Pm, 0x70 },
+       { APSHUFHW,     yxshuf, Pf3, 0x70,(00) },
+       { APSHUFL,      yxshuf, Pq, 0x70,(00) },
+       { APSHUFLW,     yxshuf, Pf2, 0x70,(00) },
+       { APSHUFW,      ymshuf, Pm, 0x70,(00) },
        { APSLLO,       ypsdq,  Pq, 0x73,(07) },
        { APSLLL,       yps,    Py, 0xf2, 0x72,(06), Pe,0xf2, Pe,0x72,(06) },
        { APSLLQ,       yps,    Py, 0xf3, 0x73,(06), Pe,0xf3, Pe,0x73,(06) },
@@ -1101,8 +1111,8 @@ Optab optab[] =
        { ASHRL,        yshl,   Px, 0xd1,(05),0xc1,(05),0xd3,(05),0xd3,(05) },
        { ASHRQ,        yshl,   Pw, 0xd1,(05),0xc1,(05),0xd3,(05),0xd3,(05) },
        { ASHRW,        yshl,   Pe, 0xd1,(05),0xc1,(05),0xd3,(05),0xd3,(05) },
-       { ASHUFPD,      yxshuf, Pq, 0xc6 },
-       { ASHUFPS,      yxshuf, Pm, 0xc6 },
+       { ASHUFPD,      yxshuf, Pq, 0xc6,(00) },
+       { ASHUFPS,      yxshuf, Pm, 0xc6,(00) },
        { ASQRTPD,      yxm,    Pe, 0x51 },
        { ASQRTPS,      yxm,    Pm, 0x51 },
        { ASQRTSD,      yxm,    Pf2, 0x51 },
@@ -1296,6 +1306,15 @@ Optab optab[] =
 
        { AUNDEF,               ynone,  Px, 0x0f, 0x0b },
 
+       { AAESENC,      yaes,   Pq, 0x38,0xdc,(0) },
+       { AAESENCLAST,  yaes,   Pq, 0x38,0xdd,(0) },
+       { AAESDEC,      yaes,   Pq, 0x38,0xde,(0) },
+       { AAESDECLAST,  yaes,   Pq, 0x38,0xdf,(0) },
+       { AAESIMC,      yaes,   Pq, 0x38,0xdb,(0) },
+       { AAESKEYGENASSIST,     yaes2,  Pq, 0x3a,0xdf,(0) },
+
+       { APSHUFD,      yaes2,  Pq,     0x70,(0) },
+
        { AEND },
        0
 };

diff --git a/src/cmd/6l/span.c b/src/cmd/6l/span.c
index 60916c0412f29ced4ec4d847476fd4199a7628f9..abffc2935cfeae73afd36834057c3f3e4ab69a9a 100644 (file)
--- a/src/cmd/6l/span.c
+++ b/src/cmd/6l/span.c
@@ -1247,7 +1247,8 @@ found:
                break;
 
        case Zibm_r:
-               *andptr++ = op;
+               while ((op = o->op[z++]) != 0)
+                       *andptr++ = op;
                asmand(&p->from, &p->to);
                *andptr++ = p->to.offset;
                break;