From 50875294c25b38168c5ff60177295541df1ad30d Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Fri, 17 Nov 2023 14:39:03 +0900
Subject: [PATCH] add tests of 2-op imul

---
 test/apx.cpp  | 69 ++++++++++++++++++++++++++++++---------------------
 xbyak/xbyak.h |  5 ++--
 2 files changed, 44 insertions(+), 30 deletions(-)

diff --git a/test/apx.cpp b/test/apx.cpp
index a5ada010..08cbbb91 100644
--- a/test/apx.cpp
+++ b/test/apx.cpp
@@ -735,7 +735,7 @@ CYBOZU_TEST_AUTO(inc_dec)
 	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
 }
 
-CYBOZU_TEST_AUTO(div)
+CYBOZU_TEST_AUTO(div_op1)
 {
 	struct Code : Xbyak::CodeGenerator {
 		Code()
@@ -804,29 +804,6 @@ CYBOZU_TEST_AUTO(div)
 			not_(word [r20+r30*1]);
 			not_(dword [r20+r30*1]);
 			not_(qword [r20+r30*1]);
-
-			// 2op
-			// imul(r30b, al); // QQQ : not supported?
-			imul(r30w, ax);
-			imul(r30d, eax);
-			imul(r30, rax);
-			imul(r30|T_nf, rax);
-			imul(rcx|T_nf, rax);
-			imul(rcx, ptr [r30]);
-
-			neg(r30b, al);
-			neg(r30w, ax);
-			neg(r30d, eax);
-			neg(r30, rax);
-			neg(r30|T_nf, rax);
-			neg(rcx|T_nf, rax);
-			neg(rcx, ptr [r30]);
-
-			not_(r30b, al);
-			not_(r30w, ax);
-			not_(r30d, eax);
-			not_(r30, rax);
-			not_(rcx, ptr [r30]);
 		}
 	} c;
 	const uint8_t tbl[] = {
@@ -862,12 +839,48 @@ CYBOZU_TEST_AUTO(div)
 		0xf7, 0xd4, 0x62, 0xfc, 0xfc, 0x08, 0xf7, 0xd4, 0x62, 0xbc, 0x78, 0x08, 0xf6, 0x14, 0x34, 0x62,
 		0xbc, 0x79, 0x08, 0xf7, 0x14, 0x34, 0x62, 0xbc, 0x78, 0x08, 0xf7, 0x14, 0x34, 0x62, 0xbc, 0xf8,
 		0x08, 0xf7, 0x14, 0x34,
+	};
+	const size_t n = sizeof(tbl);
+	CYBOZU_TEST_EQUAL(c.getSize(), n);
+	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
+}
 
-		// 2op
+CYBOZU_TEST_AUTO(imul_2op)
+{
+	struct Code : Xbyak::CodeGenerator {
+		Code()
+		{
+			// imul(r30b, al); // QQQ : not supported?
+			imul(r30w, ax);
+			imul(r30d, eax);
+			imul(r30, rax);
+			imul(r30|T_nf, rax);
+			imul(rcx|T_nf, rax);
+			imul(rcx, ptr [r30]);
+
+			neg(r30b, al);
+			neg(r30w, ax);
+			neg(r30d, eax);
+			neg(r30, rax);
+			neg(r30|T_nf, rax);
+			neg(rcx|T_nf, rax);
+			neg(rcx, ptr [r30]);
+
+			not_(r30b, al);
+			not_(r30w, ax);
+			not_(r30d, eax);
+			not_(r30, rax);
+			not_(rcx, ptr [r30]);
+		}
+	} c;
+	const uint8_t tbl[] = {
 		// imul
-		0x62, 0x64, 0x7d, 0x08, 0xaf, 0xf0, 0x62, 0x64, 0x7c, 0x08, 0xaf, 0xf0, 0x62, 0x64, 0xfc, 0x08,
-		0xaf, 0xf0, 0x62, 0x64, 0xfc, 0x0c, 0xaf, 0xf0, 0x62, 0xf4, 0xfc, 0x0c, 0xaf, 0xc8, 0x62, 0xdc,
-		0xfc, 0x08, 0xaf, 0x0e,
+		0x62, 0x64, 0x7d, 0x08, 0xaf, 0xf0,
+		0x62, 0x64, 0x7c, 0x08, 0xaf, 0xf0,
+		0x62, 0x64, 0xfc, 0x08, 0xaf, 0xf0,
+		0x62, 0x64, 0xfc, 0x0c, 0xaf, 0xf0,
+		0x62, 0xf4, 0xfc, 0x0c, 0xaf, 0xc8,
+		0x62, 0xdc, 0xfc, 0x08, 0xaf, 0x0e,
 
 		// neg
 		0x62, 0xf4, 0x0c, 0x10, 0xf6, 0xd8, 0x62, 0xf4, 0x0d, 0x10, 0xf7, 0xd8, 0x62, 0xf4, 0x0c, 0x10,
diff --git a/xbyak/xbyak.h b/xbyak/xbyak.h
index b4243349..afa6c6ae 100644
--- a/xbyak/xbyak.h
+++ b/xbyak/xbyak.h
@@ -606,6 +606,7 @@ class Operand {
 	XBYAK_CONSTEXPR bool hasEvex() const { return isZMM() || isExtIdx2() || getOpmaskIdx() || getRounding(); }
 	XBYAK_CONSTEXPR bool hasRex() const { return isExt8bit() || isREG(64) || isExtIdx(); }
 	XBYAK_CONSTEXPR bool hasRex2() const;
+	XBYAK_CONSTEXPR bool hasRex2NF() const { return hasRex2() || NF_; }
 	XBYAK_CONSTEXPR bool hasZero() const { return zero_; }
 	XBYAK_CONSTEXPR int getOpmaskIdx() const { return mask_; }
 	XBYAK_CONSTEXPR int getRounding() const { return rounding_; }
@@ -2176,7 +2177,7 @@ class CodeGenerator : public CodeArray {
 		int opBit = op.getBit();
 		if (disableRex && opBit == 64) opBit = 32;
 		const Reg r(ext, Operand::REG, opBit);
-		if ((type & T_VEX) && (op.hasRex2() || op.getNF()) && opROO(Reg(0, Operand::REG, opBit), op, r, type, code)) return;
+		if ((type & T_VEX) && op.hasRex2NF() && opROO(Reg(0, Operand::REG, opBit), op, r, type, code)) return;
 		if (op.isMEM()) {
 			opMR(op.getAddress(), r, type, code, immSize);
 		} else if (op.isREG(bit)) {
@@ -2899,7 +2900,7 @@ class CodeGenerator : public CodeArray {
 	// (r, r, m) or (r, m, r)
 	bool opROO(const Reg& d, const Operand& op1, const Operand& op2, uint64_t type, int code, int immSize = 0)
 	{
-		if (!(type & T_VEX) && !d.isREG() && !(d.hasRex2() || op1.hasRex2() || op2.hasRex2())) return false;
+		if (!d.isREG() && !(d.hasRex2NF() || op1.hasRex2NF() || op2.hasRex2NF())) return false;
 		const Operand *p1 = &op1, *p2 = &op2;
 		if (p1->isMEM()) { std::swap(p1, p2); } else { if (p2->isMEM()) code |= 2; }
 		if (p1->isMEM()) XBYAK_THROW_RET(ERR_BAD_COMBINATION, false)