Support for vgather

3 files changed, 40 insertions, 2 deletions

diff --git a/simplex86.c b/simplex86.c
index 8620679..900715c 100644
--- a/simplex86.c
+++ b/simplex86.c
@@ -65,6 +65,7 @@ typedef enum
     E_NP,		/* opcode only */
     E_M,		/* modrm/d */
     E_VM,               /* modrm/d with source in VEX */
+    E_RMV,
     E_D,                /* immediate data, no opcode */
     E_ANNOTATE,		/* generate an annotation */
     E_N_ENCODINGS
@@ -704,6 +705,24 @@ static const variant_t variants[] =
     { I_vmaskmovpd, { A_MEM, A_AVX, A_AVX }, F_AVX, E_MVR, 0x2f, VEX(256,0x66,0x0f38,W0) },
     { I_vzeroupper, { 0 }, F_AVX, E_NP, 0x77, VEX(128,0x00,0x0f,WIG) },
     { I_vzeroall, { 0 }, F_AVX, E_NP, 0x77, VEX(256,0x00,0x0f,WIG) },
+
+    /* AVX 2 */
+    { I_vgatherdd, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x90, VEX(128,0x66,0x0f38,W0) },
+    { I_vgatherdd, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x90, VEX(256,0x66,0x0f38,W0) },
+    { I_vgatherqd, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x91, VEX(128,0x66,0x0f38,W0) },
+    { I_vgatherqd, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x91, VEX(256,0x66,0x0f38,W0) },
+    { I_vgatherdq, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x90, VEX(128,0x66,0x0f38,W1) },
+    { I_vgatherdq, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x90, VEX(256,0x66,0x0f38,W1) },
+    { I_vgatherqq, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x91, VEX(128,0x66,0x0f38,W1) },
+    { I_vgatherqq, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x91, VEX(256,0x66,0x0f38,W1) },
+    { I_vgatherdps, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x92, VEX(128,0x66,0x0f38,W0) },
+    { I_vgatherdps, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x92, VEX(256,0x66,0x0f38,W0) },
+    { I_vgatherqps, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x93, VEX(128,0x66,0x0f38,W0) },
+    { I_vgatherqps, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x93, VEX(256,0x66,0x0f38,W0) },
+    { I_vgatherdpd, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x92, VEX(128,0x66,0x0f38,W1) },
+    { I_vgatherdpd, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x92, VEX(256,0x66,0x0f38,W1) },
+    { I_vgatherqpd, { A_SSE, A_MEM, A_SSE }, F_AVX2, E_RMV, 0x93, VEX(128,0x66,0x0f38,W1) },
+    { I_vgatherqpd, { A_AVX, A_MEM, A_AVX }, F_AVX2, E_RMV, 0x93, VEX(256,0x66,0x0f38,W1) },
     
     /* CVT16 */
     { I_vcvtph2ps, { A_SSE, A_SSEM }, F_CVT16, E_RM, 0x13, VEX(128,0x66,0x0f38,W0) },
@@ -1809,6 +1828,11 @@ emit (fragment_t *frag, const variant_t *variant, const op_t ops[4])
         vex_regno = GET_REGNO (ops[1]);
 	break;
 
+    case E_RMV:
+	c = emit_reg_regm (c, ops[0], ops[1], &r, &x, &b, &ann);
+	vex_regno = GET_REGNO (ops[2]);
+	break;
+
     case E_VM:
         c = emit_reg_regm (c, MODRM_OP (variant->info), ops[1], &r, &x, &b, &ann);
         vex_regno = GET_REGNO (ops[0]);
diff --git a/simplex86.h b/simplex86.h
index fc7ff87..bcafb24 100644
--- a/simplex86.h
+++ b/simplex86.h
@@ -504,7 +504,15 @@ typedef enum
     PROCESS_INSTRUCTION (crc32_8, 2)                                    \
     PROCESS_INSTRUCTION (crc32_16, 2)                                   \
     PROCESS_INSTRUCTION (crc32_32, 2)                                   \
-    PROCESS_INSTRUCTION (crc32_64, 2)
+    PROCESS_INSTRUCTION (crc32_64, 2)					\
+    PROCESS_INSTRUCTION (vgatherdd, 3)					\
+    PROCESS_INSTRUCTION (vgatherqd, 3)					\
+    PROCESS_INSTRUCTION (vgatherdq, 3)					\
+    PROCESS_INSTRUCTION (vgatherqq, 3)					\
+    PROCESS_INSTRUCTION (vgatherdps, 3)					\
+    PROCESS_INSTRUCTION (vgatherqps, 3)					\
+    PROCESS_INSTRUCTION (vgatherdpd, 3)					\
+    PROCESS_INSTRUCTION (vgatherqpd, 3)					\
 
 enum
 {
diff --git a/testsuite.c b/testsuite.c
index 5090514..349a718 100644
--- a/testsuite.c
+++ b/testsuite.c
@@ -42,6 +42,12 @@ test_shifts (assembler_t *as)
     fragment_t *frag = fragment_new (as);
 
     BEGIN_ASM (frag)
+	I_vgatherdd,	ymm0, INDEX(r9, 23417, ymm0, 8), ymm2,
+	I_vgatherdd,	ymm0, INDEX(eax, 23423, ymm9, 2), ymm12,
+	I_vgatherdd,	ymm0, INDEX(rax, 23423, ymm9, 2), ymm12,
+	I_vgatherdps,	ymm0, INDEX(r9, 23417, ymm0, 8), ymm2,
+	I_vgatherdps,	ymm0, INDEX(eax, 23423, ymm9, 2), ymm12,
+	I_vgatherqpd,	ymm0, INDEX(rax, 23423, ymm9, 2), ymm12,
         I_add,          PTR (rax), al,
         I_movabs,       rbx, IMM64 (0xaaaabbbbeeeeffff),
         I_stmxcsr,      PTR (rbx),
@@ -388,7 +394,7 @@ main ()
  */
     success &= run_test (as, "test_crc32", test_crc32, 0x3ce8ed36 );
     success &= run_test (as, "avx_supported", avx_supported, 0xeeaec30c);
-    success &= run_test (as, "test_shifts", test_shifts, 0x8aa3c8ae);
+    success &= run_test (as, "test_shifts", test_shifts, 0x4a641094);
     success &= run_test (as, "test_float", test_float, 0xc193030 );
     success &= run_test (as, "test_convert", test_convert, 0x9d52c83b);
     success &= run_test (as, "test_misc", test_misc, 0xc6b9051f);