From ab9653f483bd9b985524a0e2bb30cdb7d3b79e9d Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Sat, 19 Oct 2024 09:43:24 +0900
Subject: [PATCH] add test of vmovw

---
 test/misc.cpp | 52 +++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 52 insertions(+)

diff --git a/test/misc.cpp b/test/misc.cpp
index b4874a7..3ebb74c 100644
--- a/test/misc.cpp
+++ b/test/misc.cpp
@@ -2291,6 +2291,7 @@ CYBOZU_TEST_AUTO(vmovd)
 		{
 			setDefaultEncodingAVX10(PreAVX10v2Encoding);
 			vmovd(eax, xm1); // always AVX10.1
+			vmovd(xm1, eax); // always AVX10.1
 			vmovd(xm3, xm1); // always AVX10.2
 			// AVX-512 (AVX10.1)
 			vmovd(ptr[rax+128], xm1);
@@ -2300,6 +2301,7 @@ CYBOZU_TEST_AUTO(vmovd)
 
 			setDefaultEncodingAVX10(AVX10v2Encoding);
 			vmovd(eax, xm1); // always AVX10.1
+			vmovd(xm1, eax); // always AVX10.1
 			vmovd(xm3, xm1); // always AVX10.2
 			// AVX10.2
 			vmovd(ptr[rax+128], xm1);
@@ -2310,6 +2312,7 @@ CYBOZU_TEST_AUTO(vmovd)
 	} c;
 	const uint8_t tbl[] = {
 		0xc5, 0xf9, 0x7e, 0xc8, // avx10.1
+		0xc5, 0xf9, 0x6e, 0xc8, // avx10.1
 		0x62, 0xf1, 0x7e, 0x08, 0x7e, 0xd9, // avx10.2
 		0xc5, 0xf9, 0x7e, 0x88, 0x80, 0x00, 0x00, 0x00, // avx
 		0xc5, 0xf9, 0x6e, 0x88, 0x80, 0x00, 0x00, 0x00, // avx
@@ -2317,6 +2320,7 @@ CYBOZU_TEST_AUTO(vmovd)
 		0x62, 0x61, 0x7d, 0x08, 0x6e, 0x70, 0x20, // avx10.1
 
 		0xc5, 0xf9, 0x7e, 0xc8, // avx10.1
+		0xc5, 0xf9, 0x6e, 0xc8, // avx10.1
 		0x62, 0xf1, 0x7e, 0x08, 0x7e, 0xd9, // avx10.2
 		0x62, 0xf1, 0x7d, 0x08, 0xd6, 0x48, 0x20, // avx10.2
 		0x62, 0xf1, 0x7e, 0x08, 0x7e, 0x48, 0x20, // avx10.2
@@ -2328,4 +2332,52 @@ CYBOZU_TEST_AUTO(vmovd)
 	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
 }
 
+CYBOZU_TEST_AUTO(vmovw)
+{
+	struct Code : Xbyak::CodeGenerator {
+		Code()
+		{
+			setDefaultEncodingAVX10(PreAVX10v2Encoding);
+			vmovw(eax, xm1); // always avx10.1
+			vmovw(xm1, eax); // always avx10.1
+			vmovw(xm3, xm1); // always avx10.2
+			// AVX10.1
+			vmovw(ptr[rax+128], xm1);
+			vmovw(xm1, ptr[rax+128]);
+			vmovw(ptr[rax+128], xm30);
+			vmovw(xm30, ptr[rax+128]);
+
+			setDefaultEncodingAVX10(AVX10v2Encoding);
+			vmovw(eax, xm1); // always avx10.1
+			vmovw(xm1, eax); // always avx10.1
+			vmovw(xm3, xm1); // always avx10.2
+			// AVX10.2
+			vmovw(ptr[rax+128], xm1);
+			vmovw(xm1, ptr[rax+128]);
+			vmovw(ptr[rax+128], xm30);
+			vmovw(xm30, ptr[rax+128]);
+		}
+	} c;
+	const uint8_t tbl[] = {
+		0x62, 0xf5, 0x7d, 0x08, 0x7e, 0xc8,
+		0x62, 0xf5, 0x7d, 0x08, 0x6e, 0xc8,
+		0x62, 0xf5, 0x7e, 0x08, 0x6e, 0xd9,
+		0x62, 0xf5, 0x7d, 0x08, 0x7e, 0x48, 0x40,
+		0x62, 0xf5, 0x7d, 0x08, 0x6e, 0x48, 0x40,
+		0x62, 0x65, 0x7d, 0x08, 0x7e, 0x70, 0x40,
+		0x62, 0x65, 0x7d, 0x08, 0x6e, 0x70, 0x40,
+
+		0x62, 0xf5, 0x7d, 0x08, 0x7e, 0xc8,
+		0x62, 0xf5, 0x7d, 0x08, 0x6e, 0xc8,
+		0x62, 0xf5, 0x7e, 0x08, 0x6e, 0xd9,
+		0x62, 0xf5, 0x7e, 0x08, 0x7e, 0x48, 0x40,
+		0x62, 0xf5, 0x7e, 0x08, 0x6e, 0x48, 0x40,
+		0x62, 0x65, 0x7e, 0x08, 0x7e, 0x70, 0x40,
+		0x62, 0x65, 0x7e, 0x08, 0x6e, 0x70, 0x40,
+	};
+	const size_t n = sizeof(tbl) / sizeof(tbl[0]);
+	CYBOZU_TEST_EQUAL(c.getSize(), n);
+	CYBOZU_TEST_EQUAL_ARRAY(c.getCode(), tbl, n);
+}
+
 #endif