add vmpsadbw for avx10.2

herumi · Oct 13, 2024 · 8552268 · 8552268
1 parent 2d3a9ce
commit 8552268
Show file tree

Hide file tree

Showing 8 changed files with 51 additions and 13 deletions.
diff --git a/gen/gen_avx512.cpp b/gen/gen_avx512.cpp
@@ -447,15 +447,14 @@ void putX_X_XM_IMM()
 		{ 0x1B, "vcvtne2ph2hf8s", T_MUST_EVEX | T_F2 | T_MAP5 | T_EW0 | T_YMM | T_B16 | T_N1, false },
 
 		{ 0x52, "vdpphps", T_MUST_EVEX | T_0F38 | T_EW0 | T_YMM | T_B32, false },
-//		{ 0x42, "vmpsadbw", T_MUST_EVEX | T_F3 | T_0F3A | T_EW0 | T_YMM | T_B32, true },
 	};
 	for (size_t i = 0; i < NUM_OF_ARRAY(tbl); i++) {
 		const Tbl *p = &tbl[i];
 		std::string s = type2String(p->type);
 		printf("void %s(const Xmm& x1, const Xmm& x2, const Operand& op%s) { opAVX_X_X_XM(x1, x2, op, %s, 0x%02X%s); }\n"
 			, p->name, p->hasIMM ? ", uint8_t imm" : "", s.c_str(), p->code, p->hasIMM ? ", imm" : "");
 	}
-//	puts("void vmpsadbw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_0F3A | T_YMM, 0x42, encoding, imm, T_66 | T_W0 | T_YMM, T_F3 | T_EW0 | T_B32); }");
+	puts("void vmpsadbw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_0F3A | T_YMM, 0x42, encoding, imm, T_66 | T_W0 | T_YMM, T_F3 | T_0F3A | T_EW0 | T_B32, 1); }");
 }
 
 void putShift()

diff --git a/gen/gen_code.cpp b/gen/gen_code.cpp
@@ -57,7 +57,7 @@ void putX_X_XM(bool omitOnly)
 			{ 0x0C, "blendps", T_0F3A | T_66 | T_W0 | T_YMM, true, true, 3 },
 			{ 0x41, "dppd", T_0F3A | T_66 | T_W0, true, true, 3 },
 			{ 0x40, "dpps", T_0F3A | T_66 | T_W0 | T_YMM, true, true, 3 },
-			{ 0x42, "mpsadbw", T_0F3A | T_66 | T_W0 | T_YMM, true, true, 3 },
+			{ 0x42, "mpsadbw", T_0F3A | T_66 | T_W0 | T_YMM, true, true, 1 },
 			{ 0x0E, "pblendw", T_0F3A | T_66 | T_W0 | T_YMM, true, true, 3 },
 			{ 0x02, "pblendd", T_0F3A | T_66 | T_W0 | T_YMM, true, true, 2 },
 			{ 0x0B, "roundsd", T_0F3A | T_66 | T_W0, true, true, 3 },

diff --git a/test/avx10/misc.txt b/test/avx10/misc.txt
@@ -10,4 +10,11 @@ vdpphps(zm1, zm2, zm3);
 vdpphps(zm1, zm2, ptr[rax+128]);
 vdpphps(zm1, zm2, ptr_b[rax+128]);
 
-// skip vmpsadbw
+vmpsadbw(xm1, xm3, xm15, 3);
+vmpsadbw(xm1|T_z, xm4, ptr[rax+128], 5);
+
+vmpsadbw(ym1|k4, ym3, ym15, 3);
+vmpsadbw(ym1, ym4, ptr[rax+128], 5);
+
+vmpsadbw(zm1|k4, zm3, zm15, 3);
+vmpsadbw(zm1, zm4, ptr[rax+128], 5);
diff --git a/test/avx10_test.cpp b/test/avx10_test.cpp
@@ -228,3 +228,27 @@ CYBOZU_TEST_AUTO(ymm_with_sae)
 	CYBOZU_TEST_EQUAL(c.getSize(), n);
 	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
 }
+
+CYBOZU_TEST_AUTO(vmpsadbw)
+{
+	struct Code : Xbyak::CodeGenerator {
+		Code()
+		{
+			setDefaultEncoding();
+			vmpsadbw(xm1, xm3, xm15, 3); // vex(avx)
+			vmpsadbw(ym1, ym3, ptr[rax+128], 3); // vex(avx2)
+			setDefaultEncoding(VexEncoding, EvexEncoding);
+			vmpsadbw(ym1, ym3, ym15, 3); // evex(avx10.2)
+			vmpsadbw(ym1, ym3, ptr[rax+128], 3); // evex(avx10.2)
+		}
+	} c;
+	const uint8_t tbl[] = {
+		0xc4, 0xc3, 0x61, 0x42, 0xcf, 0x03,
+		0xc4, 0xe3, 0x65, 0x42, 0x88, 0x80, 0x00, 0x00, 0x00, 0x03,
+		0x62, 0xd3, 0x66, 0x28, 0x42, 0xcf, 0x03,
+		0x62, 0xf3, 0x66, 0x28, 0x42, 0x48, 0x04, 0x03,
+	};
+	const size_t n = sizeof(tbl) / sizeof(tbl[0]);
+	CYBOZU_TEST_EQUAL(c.getSize(), n);
+	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
+}
diff --git a/test/test_by_xed.cpp b/test/test_by_xed.cpp
@@ -1,10 +1,13 @@
 #include <stdio.h>
 #include <xbyak/xbyak.h>
 
+using namespace Xbyak;
+
 struct Code : Xbyak::CodeGenerator {
 	Code()
 		 : Xbyak::CodeGenerator(4096*8)
 	{
+		setDefaultEncoding(VexEncoding, EvexEncoding);
 #include "tmp.cpp"
 	}
 };

diff --git a/test/test_by_xed.py b/test/test_by_xed.py
@@ -210,6 +210,11 @@ def parseNmemonic(s):
   args = []
   attrs = []
 
+  # remove Xbyak::{Evex,Vex}Encoding
+  r = re.search(r'(,[^,]*Encoding)', s)
+  if r:
+    s = s.replace(r.group(1), '')
+
   (s, broadcast) = parseBroadcast(s)
 
   # replace xm0 with xmm0

diff --git a/xbyak/xbyak.h b/xbyak/xbyak.h
@@ -2661,21 +2661,21 @@ class CodeGenerator : public CodeArray {
 		if (addr.getRegExp().getIndex().getKind() != kind) XBYAK_THROW(ERR_BAD_VSIB_ADDRESSING)
 		opVex(x, 0, addr, type, code);
 	}
-	void opEncoding(const Xmm& x1, const Xmm& x2, const Operand& op, uint64_t type, int code, PreferredEncoding encoding, int sel = 0)
+	void opEncoding(const Xmm& x1, const Xmm& x2, const Operand& op, uint64_t type, int code, PreferredEncoding encoding, int imm = NONE, uint64_t typeVex = 0, uint64_t typeEvex = 0, int sel = 0)
 	{
-		opAVX_X_X_XM(x1, x2, op, type | orEvexIf(encoding, sel), code);
+		opAVX_X_X_XM(x1, x2, op, type | orEvexIf(encoding, typeVex, typeEvex, sel), code, imm);
 	}
-	int orEvexIf(PreferredEncoding encoding, int sel = 0) {
+	int orEvexIf(PreferredEncoding encoding, uint64_t typeVex, uint64_t typeEvex, int sel) {
 		if (encoding == DefaultEncoding) {
 			encoding = defaultEncoding_[sel];
 		}
 		if (encoding == EvexEncoding) {
 #ifdef XBYAK_DISABLE_AVX512
 			XBYAK_THROW(ERR_EVEX_IS_INVALID)
 #endif
-			return T_MUST_EVEX;
+			return T_MUST_EVEX | typeEvex;
 		}
-		return 0;
+		return typeVex;
 	}
 	void opInOut(const Reg& a, const Reg& d, uint8_t code)
 	{
@@ -3132,8 +3132,8 @@ class CodeGenerator : public CodeArray {
 #endif
 		, isDefaultJmpNEAR_(false)
 	{
-		defaultEncoding_[0] = EvexEncoding; // use avx512-vnni not avx-vnni
-		defaultEncoding_[1] = VexEncoding; // use vmpsadbw(avx) not avx10.2
+		// select avx512-vnni, vmpsadbw(avx)
+		setDefaultEncoding();
 		labelMgr_.set(this);
 	}
 	void reset()
@@ -3171,7 +3171,7 @@ class CodeGenerator : public CodeArray {
 #endif
 
 	// set default encoding to select Vex or Evex
-	void setDefaultEncoding(PreferredEncoding vnniEnc, PreferredEncoding mpsadbwEnc = VexEncoding)
+	void setDefaultEncoding(PreferredEncoding vnniEnc = EvexEncoding, PreferredEncoding mpsadbwEnc = VexEncoding)
 	{ defaultEncoding_[0] = vnniEnc; defaultEncoding_[1] = mpsadbwEnc; }
 
 	void sha1msg12(const Xmm& x, const Operand& op)

diff --git a/xbyak/xbyak_mnemonic.h b/xbyak/xbyak_mnemonic.h
@@ -1369,7 +1369,6 @@ void vmovupd(const Address& addr, const Xmm& xmm) { opAVX_X_XM_IMM(xmm, addr, T_
 void vmovupd(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, T_66|T_0F|T_EW1|T_YMM|T_EVEX, 0x10); }
 void vmovups(const Address& addr, const Xmm& xmm) { opAVX_X_XM_IMM(xmm, addr, T_0F|T_EW0|T_YMM|T_EVEX|T_M_K, 0x11); }
 void vmovups(const Xmm& xm, const Operand& op) { opAVX_X_XM_IMM(xm, op, T_0F|T_EW0|T_YMM|T_EVEX, 0x10); }
-void vmpsadbw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm) { opAVX_X_X_XM(x1, x2, op, T_66|T_0F3A|T_W0|T_YMM, 0x42, imm); }
 void vmulpd(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_0F | T_66 | T_EW1 | T_YMM | T_EVEX | T_ER_Z | T_B64, 0x59); }
 void vmulps(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_0F | T_EW0 | T_YMM | T_EVEX | T_ER_Z | T_B32, 0x59); }
 void vmulsd(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_0F | T_F2 | T_EW1 | T_EVEX | T_ER_X | T_N8, 0x59); }
@@ -2408,6 +2407,7 @@ void vmovsh(const Xmm& x1, const Xmm& x2, const Xmm& x3) { opAVX_X_X_XM(x1, x2,
 void vmovw(const Address& addr, const Xmm& x) { opAVX_X_XM_IMM(x, addr, T_N2|T_66|T_MAP5|T_MUST_EVEX, 0x7E); }
 void vmovw(const Reg32e& r, const Xmm& x) { opAVX_X_X_XM(x, xm0, r, T_N2|T_66|T_MAP5|T_MUST_EVEX, 0x7E); }
 void vmovw(const Xmm& x, const Operand& op) { if (!op.isREG(32|64) && !op.isMEM()) XBYAK_THROW(ERR_BAD_COMBINATION) opAVX_X_X_XM(x, xm0, op, T_N2|T_66|T_MAP5|T_MUST_EVEX, 0x6E); }
+void vmpsadbw(const Xmm& x1, const Xmm& x2, const Operand& op, uint8_t imm, PreferredEncoding encoding = DefaultEncoding) { opEncoding(x1, x2, op, T_0F3A | T_YMM, 0x42, encoding, imm, T_66 | T_W0 | T_YMM, T_F3 | T_0F3A | T_EW0 | T_B32, 1); }
 void vmulnepbf16(const Xmm& x1, const Xmm& x2, const Operand& op) { opAVX_X_X_XM(x1, x2, op, T_66|T_MAP5|T_EW0|T_YMM|T_MUST_EVEX|T_B16, 0x59); }
 void vmulph(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_MAP5 | T_EW0 | T_YMM | T_MUST_EVEX | T_ER_Z | T_B16, 0x59); }
 void vmulsh(const Xmm& xmm, const Operand& op1, const Operand& op2 = Operand()) { opAVX_X_X_XM(xmm, op1, op2, T_MAP5 | T_F3 | T_EW0 | T_MUST_EVEX | T_ER_X | T_N2, 0x59); }