add vpshrd(v){w,d,q}

3 files changed, 73 insertions, 0 deletions

diff --git a/gen/gen_avx512.cpp b/gen/gen_avx512.cpp
index 76d5f7d..0c11584 100644
--- a/gen/gen_avx512.cpp
+++ b/gen/gen_avx512.cpp
@@ -354,6 +354,14 @@ void putX_X_XM_IMM()
 		{ 0x70, "vpshldvw", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, false },
 		{ 0x71, "vpshldvd", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW0 | T_SAE_Z, false },
 		{ 0x71, "vpshldvq", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, false },
+
+		{ 0x72, "vpshrdw", T_66 | T_0F3A | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, true },
+		{ 0x73, "vpshrdd", T_66 | T_0F3A | T_YMM | T_MUST_EVEX | T_EW0 | T_SAE_Z, true },
+		{ 0x73, "vpshrdq", T_66 | T_0F3A | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, true },
+
+		{ 0x72, "vpshrdvw", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, false },
+		{ 0x73, "vpshrdvd", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW0 | T_SAE_Z, false },
+		{ 0x73, "vpshrdvq", T_66 | T_0F38 | T_YMM | T_MUST_EVEX | T_EW1 | T_SAE_Z, false },
 	};
 	for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
 		const Tbl *p = &tbl[i];
diff --git a/test/misc.cpp b/test/misc.cpp
index 953e787..62b23e2 100644
--- a/test/misc.cpp
+++ b/test/misc.cpp
@@ -297,4 +297,63 @@ CYBOZU_TEST_AUTO(shld)
 	CYBOZU_TEST_EQUAL(c.getSize(), n);
 	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
 }
+CYBOZU_TEST_AUTO(shrd)
+{
+	struct Code : Xbyak::CodeGenerator {
+		Code()
+		{
+			vpshrdw(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
+			vpshrdw(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
+			vpshrdw(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
+
+			vpshrdd(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
+			vpshrdd(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
+			vpshrdd(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
+
+			vpshrdq(xmm5|k3|T_z, xmm2, ptr [rax + 0x40], 5);
+			vpshrdq(ymm5|k3|T_z, ymm2, ptr [rax + 0x40], 5);
+			vpshrdq(zmm5|k3|T_z, zmm2, ptr [rax + 0x40], 5);
+
+			vpshrdvw(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
+			vpshrdvw(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
+			vpshrdvw(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
+
+			vpshrdvd(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
+			vpshrdvd(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
+			vpshrdvd(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
+
+			vpshrdvq(xmm5|k3|T_z, xmm2, ptr [rax + 0x40]);
+			vpshrdvq(ymm5|k3|T_z, ymm2, ptr [rax + 0x40]);
+			vpshrdvq(zmm5|k3|T_z, zmm2, ptr [rax + 0x40]);
+		}
+	} c;
+	const uint8_t tbl[] = {
+		0x62, 0xf3, 0xed, 0x8b, 0x72, 0x68, 0x04, 0x05,
+		0x62, 0xf3, 0xed, 0xab, 0x72, 0x68, 0x02, 0x05,
+		0x62, 0xf3, 0xed, 0xcb, 0x72, 0x68, 0x01, 0x05,
+
+		0x62, 0xf3, 0x6d, 0x8b, 0x73, 0x68, 0x04, 0x05,
+		0x62, 0xf3, 0x6d, 0xab, 0x73, 0x68, 0x02, 0x05,
+		0x62, 0xf3, 0x6d, 0xcb, 0x73, 0x68, 0x01, 0x05,
+
+		0x62, 0xf3, 0xed, 0x8b, 0x73, 0x68, 0x04, 0x05,
+		0x62, 0xf3, 0xed, 0xab, 0x73, 0x68, 0x02, 0x05,
+		0x62, 0xf3, 0xed, 0xcb, 0x73, 0x68, 0x01, 0x05,
+
+		0x62, 0xf2, 0xed, 0x8b, 0x72, 0x68, 0x04,
+		0x62, 0xf2, 0xed, 0xab, 0x72, 0x68, 0x02,
+		0x62, 0xf2, 0xed, 0xcb, 0x72, 0x68, 0x01,
+
+		0x62, 0xf2, 0x6d, 0x8b, 0x73, 0x68, 0x04,
+		0x62, 0xf2, 0x6d, 0xab, 0x73, 0x68, 0x02,
+		0x62, 0xf2, 0x6d, 0xcb, 0x73, 0x68, 0x01,
+
+		0x62, 0xf2, 0xed, 0x8b, 0x73, 0x68, 0x04,
+		0x62, 0xf2, 0xed, 0xab, 0x73, 0x68, 0x02,
+		0x62, 0xf2, 0xed, 0xcb, 0x73, 0x68, 0x01,
+	};
+	const size_t n = sizeof(tbl) / sizeof(tbl[0]);
+	CYBOZU_TEST_EQUAL(c.getSize(), n);
+	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
+}
 #endif
diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
index 00ca540..06da38e 100644
--- a/xbyak/xbyak_mnemonic.h
+++ b/xbyak/xbyak_mnemonic.h
@@ -1863,6 +1863,12 @@ void vpshldvd(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1
 void vpshldvq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x71); }
 void vpshldvw(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x70); }
 void vpshldw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8 imm) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F3A | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x70, imm); }
+void vpshrdd(const Xmm& x1, const Xmm& x2, const Operand& op, uint8 imm) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F3A | T_EW0 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x73, imm); }
+void vpshrdq(const Xmm& x1, const Xmm& x2, const Operand& op, uint8 imm) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F3A | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x73, imm); }
+void vpshrdvd(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW0 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x73); }
+void vpshrdvq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x73); }
+void vpshrdvw(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x72); }
+void vpshrdw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8 imm) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F3A | T_EW1 | T_YMM | T_SAE_Z | T_MUST_EVEX, 0x72, imm); }
 void vpsllvw(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66 | T_0F38 | T_EW1 | T_YMM | T_MUST_EVEX, 0x12); }
 void vpsraq(const Xmm& x, const Operand& op, uint8 imm) { opAVX_X_X_XM(Xmm(x.getKind(), 4), x, op, T_66 | T_0F | T_EW1 | T_YMM | T_MUST_EVEX | T_B64, 0x72, imm); }
 void vpsraq(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_N16 | T_66 | T_0F | T_EW1 | T_YMM | T_MUST_EVEX, 0xE2); }