cmd/internal/obj/x86: add some missing AMD64 instructions

Add VBROADCASTSD, BROADCASTSS, MOVDDUP, MOVSHDUP, MOVSLDUP,
VMOVDDUP, VMOVSHDUP, VMOVSLDUP.

Fixes #16007

Change-Id: I9614e58eed6c1b6f299d9b4f0b1a7750aa7c1725
Reviewed-on: https://go-review.googlesource.com/31491
Run-TryBot: Ilya Tocar <ilya.tocar@intel.com>
TryBot-Result: Gobot Gobot <gobot@golang.org>
Reviewed-by: David Chase <drchase@google.com>

src/cmd/internal/obj/x86/a.out.go

@@ -771,6 +771,9 @@ const (
 	AROUNDSS
 	AROUNDPD
 	AROUNDSD
+	AMOVDDUP
+	AMOVSHDUP
+	AMOVSLDUP
 
 	APSHUFD
 	APCLMULQDQ
@@ -803,6 +806,11 @@ const (
 	AVPERM2I128
 	ARORXL
 	ARORXQ
+	AVBROADCASTSS
+	AVBROADCASTSD
+	AVMOVDDUP
+	AVMOVSHDUP
+	AVMOVSLDUP
 
 	// from 386
 	AJCXZW

src/cmd/internal/obj/x86/anames.go

@@ -708,6 +708,9 @@ var Anames = []string{
 	"ROUNDSS",
 	"ROUNDPD",
 	"ROUNDSD",
+	"MOVDDUP",
+	"MOVSHDUP",
+	"MOVSLDUP",
 	"PSHUFD",
 	"PCLMULQDQ",
 	"VZEROUPPER",
@@ -738,6 +741,11 @@ var Anames = []string{
 	"VPERM2I128",
 	"RORXL",
 	"RORXQ",
+	"VBROADCASTSS",
+	"VBROADCASTSD",
+	"VMOVDDUP",
+	"VMOVSHDUP",
+	"VMOVSLDUP",
 	"JCXZW",
 	"FCMOVCC",
 	"FCMOVCS",

src/cmd/internal/obj/x86/asm6.go

@@ -866,6 +866,10 @@ var yvex_vpbroadcast = []ytab{
 	{Yxm, Ynone, Yyr, Zvex_rm_v_r, 2},
 }
 
+var yvex_vpbroadcast_sd = []ytab{
+	{Yxm, Ynone, Yyr, Zvex_rm_v_r, 2},
+}
+
 var ymmxmm0f38 = []ytab{
 	{Ymm, Ynone, Ymr, Zlitm_r, 3},
 	{Yxm, Ynone, Yxr, Zlitm_r, 5},
@@ -1630,6 +1634,9 @@ var optab =
 	{APSHUFD, yxshuf, Pq, [23]uint8{0x70, 0}},
 	{APCLMULQDQ, yxshuf, Pq, [23]uint8{0x3a, 0x44, 0}},
 	{APCMPESTRI, yxshuf, Pq, [23]uint8{0x3a, 0x61, 0}},
+	{AMOVDDUP, yxm, Pf2, [23]uint8{0x12}},
+	{AMOVSHDUP, yxm, Pf3, [23]uint8{0x16}},
+	{AMOVSLDUP, yxm, Pf3, [23]uint8{0x12}},
 
 	{AANDNL, yvex_r3, Pvex, [23]uint8{VEX_LZ_0F38_W0, 0xF2}},
 	{AANDNQ, yvex_r3, Pvex, [23]uint8{VEX_LZ_0F38_W1, 0xF2}},
@@ -1678,6 +1685,11 @@ var optab =
 	{AVPERM2I128, yvex_yyi4, Pvex, [23]uint8{VEX_256_66_0F3A_WIG, 0x46}},
 	{ARORXL, yvex_ri3, Pvex, [23]uint8{VEX_LZ_F2_0F3A_W0, 0xf0}},
 	{ARORXQ, yvex_ri3, Pvex, [23]uint8{VEX_LZ_F2_0F3A_W1, 0xf0}},
+	{AVBROADCASTSD, yvex_vpbroadcast_sd, Pvex, [23]uint8{VEX_256_66_0F38_W0, 0x19}},
+	{AVBROADCASTSS, yvex_vpbroadcast, Pvex, [23]uint8{VEX_128_66_0F38_W0, 0x18, VEX_256_66_0F38_W0, 0x18}},
+	{AVMOVDDUP, yvex_xy2, Pvex, [23]uint8{VEX_128_F2_0F_WIG, 0x12, VEX_256_F2_0F_WIG, 0x12}},
+	{AVMOVSHDUP, yvex_xy2, Pvex, [23]uint8{VEX_128_F3_0F_WIG, 0x16, VEX_256_F3_0F_WIG, 0x16}},
+	{AVMOVSLDUP, yvex_xy2, Pvex, [23]uint8{VEX_128_F3_0F_WIG, 0x12, VEX_256_F3_0F_WIG, 0x12}},
 
 	{AXACQUIRE, ynone, Px, [23]uint8{0xf2}},
 	{AXRELEASE, ynone, Px, [23]uint8{0xf3}},
@@ -3379,7 +3391,8 @@ func doasm(ctxt *obj.Link, p *obj.Prog) {
 				log.Fatalf("asmins bad table %v", p)
 			}
 			op = int(o.op[z])
-			if op == 0x0f {
+			// In vex case 0x0f is actually VEX_256_F2_0F_WIG
+			if op == 0x0f && o.prefix != Pvex {
 				ctxt.AsmBuf.Put1(byte(op))
 				z++
 				op = int(o.op[z])