Merge branch 'dev'HEADv7.22master

3 files changed, 74 insertions, 35 deletions

diff --git a/test/Makefile b/test/Makefile
index a61895f..cf5c716 100644
--- a/test/Makefile
+++ b/test/Makefile
@@ -60,8 +60,7 @@ apx: apx.cpp $(XBYAK_INC)
 avx10_test: avx10_test.cpp $(XBYAK_INC)
 	$(CXX) $(CFLAGS) avx10_test.cpp -o $@ -DXBYAK64
 
-#TEST_FILES=old.txt new-ymm.txt bf16.txt comp.txt misc.txt convert.txt minmax.txt saturation.txt
-TEST_FILES=old.txt new-ymm.txt bf16.txt misc.txt convert.txt minmax.txt saturation.txt
+TEST_FILES=old.txt new-ymm.txt bf16.txt comp.txt misc.txt convert.txt minmax.txt saturation.txt
 xed_test:
 	@set -e; \
 	for target in $(addprefix avx10/, $(TEST_FILES)); do \
diff --git a/test/avx10/bf16.txt b/test/avx10/bf16.txt
index a387c61..c544e02 100644
--- a/test/avx10/bf16.txt
+++ b/test/avx10/bf16.txt
@@ -113,17 +113,17 @@ vfpclasspbf16(k7|k5, zword_b[rax+128], 13);
 vcomsbf16(xm2, xm3);
 vcomsbf16(xm2, ptr[rax+128]);
 
-//vgetexppbf16(xm1|k3, xmm2);
-//vgetexppbf16(xm1|k3, ptr[rax+128]);
-//vgetexppbf16(xm1|k3, ptr_b[rax+128]);
+vgetexppbf16(xm1|k3, xmm2);
+vgetexppbf16(xm1|k3, ptr[rax+128]);
+vgetexppbf16(xm1|k3, ptr_b[rax+128]);
 
-//vgetexppbf16(ym1|k3, ymm2);
-//vgetexppbf16(ym1|k3, ptr[rax+128]);
-//vgetexppbf16(ym1|k3, ptr_b[rax+128]);
+vgetexppbf16(ym1|k3, ymm2);
+vgetexppbf16(ym1|k3, ptr[rax+128]);
+vgetexppbf16(ym1|k3, ptr_b[rax+128]);
 
-//vgetexppbf16(zm1|k3, zmm2);
-//vgetexppbf16(zm1|k3, ptr[rax+128]);
-//vgetexppbf16(zm1|k3, ptr_b[rax+128]);
+vgetexppbf16(zm1|k3, zmm2);
+vgetexppbf16(zm1|k3, ptr[rax+128]);
+vgetexppbf16(zm1|k3, ptr_b[rax+128]);
 
 vgetmantpbf16(xm1|k3, xmm2, 3);
 vgetmantpbf16(xm1|k3, ptr[rax+128], 5);
diff --git a/test/cvt_test.cpp b/test/cvt_test.cpp
index e5e9748..20befef 100644
--- a/test/cvt_test.cpp
+++ b/test/cvt_test.cpp
@@ -12,38 +12,45 @@ const struct Ptn {
 	Reg16 reg16;
 	Reg32 reg32;
 	Reg64 reg64;
+	Xmm x;
+	Ymm y;
+	Zmm z;
 } tbl[] = {
-	{ &al, ax, eax, rax },
-	{ &bl, bx, ebx, rbx },
-	{ &cl, cx, ecx, rcx },
-	{ &dl, dx, edx, rdx },
-	{ &sil, si, esi, rsi },
-	{ &dil, di, edi, rdi },
-	{ &bpl, bp, ebp, rbp },
-	{ &spl, sp, esp, rsp },
-	{ &r8b, r8w, r8d, r8 },
-	{ &r9b, r9w, r9d, r9 },
-	{ &r10b, r10w, r10d, r10 },
-	{ &r11b, r11w, r11d, r11 },
-	{ &r12b, r12w, r12d, r12 },
-	{ &r13b, r13w, r13d, r13 },
-	{ &r14b, r14w, r14d, r14 },
-	{ &r15b, r15w, r15d, r15 },
+	{ &al, ax, eax, rax, xmm0, ymm0, zmm0 },
+	{ &bl, bx, ebx, rbx, xmm3, ymm3, zmm3 },
+	{ &cl, cx, ecx, rcx, xmm1, ymm1, zmm1 },
+	{ &dl, dx, edx, rdx, xmm2, ymm2, zmm2 },
+	{ &sil, si, esi, rsi, xmm6, ymm6, zmm6 },
+	{ &dil, di, edi, rdi, xmm7, ymm7, zmm7 },
+	{ &bpl, bp, ebp, rbp, xmm5, ymm5, zmm5 },
+	{ &spl, sp, esp, rsp, xmm4, ymm4, zmm4 },
+	{ &r8b, r8w, r8d, r8, xmm8, ymm8, zmm8 },
+	{ &r9b, r9w, r9d, r9, xmm9, ymm9, zmm9 },
+	{ &r10b, r10w, r10d, r10, xmm10, ymm10, zmm10 },
+	{ &r11b, r11w, r11d, r11, xmm11, ymm11, zmm11 },
+	{ &r12b, r12w, r12d, r12, xmm12, ymm12, zmm12 },
+	{ &r13b, r13w, r13d, r13, xmm13, ymm13, zmm13 },
+	{ &r14b, r14w, r14d, r14, xmm14, ymm14, zmm14 },
+	{ &r15b, r15w, r15d, r15, xmm15, ymm15, zmm15 },
+	{ &r31b, r31w, r31d, r31, xmm31, ymm31, zmm31 },
 };
 #else
 const struct Ptn {
 	const Reg8 *reg8;
 	Reg16 reg16;
 	Reg32 reg32;
+	Xmm x;
+	Ymm y;
+	Zmm z;
 } tbl[] = {
-	{ &al, ax, eax },
-	{ &bl, bx, ebx },
-	{ &cl, cx, ecx },
-	{ &dl, dx, edx },
-	{ 0, si, esi },
-	{ 0, di, edi },
-	{ 0, bp, ebp },
-	{ 0, sp, esp },
+	{ &al, ax, eax, xmm0, ymm0, zmm0 },
+	{ &bl, bx, ebx, xmm3, ymm3, zmm3 },
+	{ &cl, cx, ecx, xmm1, ymm1, zmm1 },
+	{ &dl, dx, edx, xmm2, ymm2, zmm2 },
+	{ 0, si, esi, xmm6, ymm6, zmm6 },
+	{ 0, di, edi, xmm7, ymm7, zmm7 },
+	{ 0, bp, ebp, xmm5, ymm5, zmm5 },
+	{ 0, sp, esp, xmm4, ymm4, zmm4 },
 };
 #endif
 
@@ -54,13 +61,40 @@ CYBOZU_TEST_AUTO(cvt)
 			CYBOZU_TEST_ASSERT(tbl[i].reg8->cvt8() == *tbl[i].reg8);
 			CYBOZU_TEST_ASSERT(tbl[i].reg8->cvt16() == tbl[i].reg16);
 			CYBOZU_TEST_ASSERT(tbl[i].reg8->cvt32() == tbl[i].reg32);
+			CYBOZU_TEST_ASSERT(tbl[i].reg8->cvt128() == tbl[i].x);
+			CYBOZU_TEST_ASSERT(tbl[i].reg8->cvt256() == tbl[i].y);
+			CYBOZU_TEST_ASSERT(tbl[i].reg8->cvt512() == tbl[i].z);
 			CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt8() == *tbl[i].reg8);
 			CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt8() == *tbl[i].reg8);
+			CYBOZU_TEST_ASSERT(tbl[i].x.cvt8() == *tbl[i].reg8);
+			CYBOZU_TEST_ASSERT(tbl[i].y.cvt8() == *tbl[i].reg8);
+			CYBOZU_TEST_ASSERT(tbl[i].z.cvt8() == *tbl[i].reg8);
 		}
 		CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt16() == tbl[i].reg16);
 		CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt32() == tbl[i].reg32);
+		CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt128() == tbl[i].x);
+		CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt256() == tbl[i].y);
+		CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt512() == tbl[i].z);
 		CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt16() == tbl[i].reg16);
 		CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt32() == tbl[i].reg32);
+		CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt128() == tbl[i].x);
+		CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt256() == tbl[i].y);
+		CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt512() == tbl[i].z);
+		CYBOZU_TEST_ASSERT(tbl[i].x.cvt16() == tbl[i].reg16);
+		CYBOZU_TEST_ASSERT(tbl[i].x.cvt32() == tbl[i].reg32);
+		CYBOZU_TEST_ASSERT(tbl[i].x.cvt128() == tbl[i].x);
+		CYBOZU_TEST_ASSERT(tbl[i].x.cvt256() == tbl[i].y);
+		CYBOZU_TEST_ASSERT(tbl[i].x.cvt512() == tbl[i].z);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt16() == tbl[i].reg16);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt32() == tbl[i].reg32);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt128() == tbl[i].x);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt256() == tbl[i].y);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt512() == tbl[i].z);
+		CYBOZU_TEST_ASSERT(tbl[i].z.cvt16() == tbl[i].reg16);
+		CYBOZU_TEST_ASSERT(tbl[i].z.cvt32() == tbl[i].reg32);
+		CYBOZU_TEST_ASSERT(tbl[i].z.cvt128() == tbl[i].x);
+		CYBOZU_TEST_ASSERT(tbl[i].z.cvt256() == tbl[i].y);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt512() == tbl[i].z);
 #ifdef XBYAK64
 		if (tbl[i].reg8) {
 			CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt8() == *tbl[i].reg8);
@@ -69,8 +103,14 @@ CYBOZU_TEST_AUTO(cvt)
 		CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt16() == tbl[i].reg16);
 		CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt32() == tbl[i].reg32);
 		CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt64() == tbl[i].reg64);
+		CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt128() == tbl[i].x);
+		CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt256() == tbl[i].y);
+		CYBOZU_TEST_ASSERT(tbl[i].reg64.cvt512() == tbl[i].z);
 		CYBOZU_TEST_ASSERT(tbl[i].reg16.cvt64() == tbl[i].reg64);
 		CYBOZU_TEST_ASSERT(tbl[i].reg32.cvt64() == tbl[i].reg64);
+		CYBOZU_TEST_ASSERT(tbl[i].x.cvt64() == tbl[i].reg64);
+		CYBOZU_TEST_ASSERT(tbl[i].y.cvt64() == tbl[i].reg64);
+		CYBOZU_TEST_ASSERT(tbl[i].z.cvt64() == tbl[i].reg64);
 #endif
 	}
 	{