finish up long mode avx2

2 files changed, 228 insertions, 37 deletions

diff --git a/src/long_mode/vex.rs b/src/long_mode/vex.rs
index 401c61f..d7e2257 100644
--- a/src/long_mode/vex.rs
+++ b/src/long_mode/vex.rs
@@ -74,6 +74,7 @@ enum VEXOperandCode {
     G_xmm_E_xmm,
     G_xmm_E_ymm,
     G_ymm_E_xmm,
+    G_ymm_M_xmm,
     G_ymm_E_ymm,
     G_V_ymm_E_xmm,
     M_V_G_xmm,
@@ -913,12 +914,18 @@ fn read_vex_operands<T: Iterator<Item=u8>>(bytes: &mut T, instruction: &mut Inst
             instruction.operand_count = 2;
             Ok(())
         }
-        _op @ VEXOperandCode::G_ymm_E_xmm => {
+        op @ VEXOperandCode::G_ymm_M_xmm |
+        op @ VEXOperandCode::G_ymm_E_xmm => {
             if instruction.vex_reg.num != 0 {
                 instruction.opcode = Opcode::Invalid;
                 return Err(DecodeError::InvalidOperand);
             }
             let modrm = read_modrm(bytes, length)?;
+            if modrm & 0xc0 == 0xc0 {
+                if let VEXOperandCode::G_ymm_M_xmm = op {
+                    return Err(DecodeError::InvalidOperand);
+                }
+            }
             instruction.modrm_rrr =
                 RegSpec::from_parts((modrm >> 3) & 7, instruction.prefixes.vex().r(), RegisterBank::Y);
             let mem_oper = read_E_xmm(bytes, instruction, modrm, length)?;
@@ -1001,9 +1008,14 @@ fn read_vex_operands<T: Iterator<Item=u8>>(bytes: &mut T, instruction: &mut Inst
             instruction.operand_count = 2;
             Ok(())
         }
-        _op @ VEXOperandCode::G_V_E_ymm |
-        _op @ VEXOperandCode::G_V_M_ymm => {
+        op @ VEXOperandCode::G_V_E_ymm |
+        op @ VEXOperandCode::G_V_M_ymm => {
             let modrm = read_modrm(bytes, length)?;
+            if let VEXOperandCode::G_V_M_ymm = op {
+                if modrm & 0xc0 == 0xc0 {
+                    return Err(DecodeError::InvalidOperand);
+                }
+            }
             instruction.modrm_rrr =
                 RegSpec::from_parts((modrm >> 3) & 7, instruction.prefixes.vex().r(), RegisterBank::Y);
             instruction.vex_reg.bank = RegisterBank::Y;
@@ -1028,9 +1040,14 @@ fn read_vex_operands<T: Iterator<Item=u8>>(bytes: &mut T, instruction: &mut Inst
             instruction.operand_count = 4;
             Ok(())
         }
-        _op @ VEXOperandCode::E_V_G_ymm |
-        _op @ VEXOperandCode::M_V_G_ymm => {
+        op @ VEXOperandCode::E_V_G_ymm |
+        op @ VEXOperandCode::M_V_G_ymm => {
             let modrm = read_modrm(bytes, length)?;
+            if let VEXOperandCode::M_V_G_ymm = op {
+                if modrm & 0xc0 == 0xc0 {
+                    return Err(DecodeError::InvalidOperand);
+                }
+            }
             instruction.modrm_rrr =
                 RegSpec::from_parts((modrm >> 3) & 7, instruction.prefixes.vex().r(), RegisterBank::Y);
             instruction.vex_reg.bank = RegisterBank::Y;
@@ -1043,6 +1060,9 @@ fn read_vex_operands<T: Iterator<Item=u8>>(bytes: &mut T, instruction: &mut Inst
         }
         VEXOperandCode::G_V_M_xmm => {
             let modrm = read_modrm(bytes, length)?;
+            if modrm & 0xc0 == 0xc0 {
+                return Err(DecodeError::InvalidOperand);
+            }
             instruction.modrm_rrr =
                 RegSpec::from_parts((modrm >> 3) & 7, instruction.prefixes.vex().r(), RegisterBank::X);
             let mem_oper = read_E_xmm(bytes, instruction, modrm, length)?;
@@ -1129,9 +1149,14 @@ fn read_vex_operands<T: Iterator<Item=u8>>(bytes: &mut T, instruction: &mut Inst
             instruction.operand_count = 4;
             Ok(())
         }
-        _op @ VEXOperandCode::E_V_G_xmm |
-        _op @ VEXOperandCode::M_V_G_xmm => {
+        op @ VEXOperandCode::E_V_G_xmm |
+        op @ VEXOperandCode::M_V_G_xmm => {
             let modrm = read_modrm(bytes, length)?;
+            if let VEXOperandCode::M_V_G_xmm = op {
+                if modrm & 0xc0 == 0xc0 {
+                    return Err(DecodeError::InvalidOperand);
+                }
+            }
             instruction.modrm_rrr =
                 RegSpec::from_parts((modrm >> 3) & 7, instruction.prefixes.vex().r(), RegisterBank::X);
             let mem_oper = read_E_xmm(bytes, instruction, modrm, length)?;
@@ -2288,7 +2313,10 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         VEXOperandCode::G_E_xmm
                     }),
                     0x16 => (Opcode::VPERMPS, if L {
-                        VEXOperandCode::G_V_E_xmm
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
+                        VEXOperandCode::G_V_E_ymm
                     } else {
                         instruction.opcode = Opcode::Invalid;
                         return Err(DecodeError::InvalidOpcode);
@@ -2303,7 +2331,7 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         return Err(DecodeError::InvalidOpcode);
                     } else {
                         (Opcode::VBROADCASTSS, if L {
-                            VEXOperandCode::G_E_ymm
+                            VEXOperandCode::G_ymm_E_xmm
                         } else {
                             VEXOperandCode::G_E_xmm
                         })
@@ -2313,13 +2341,13 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         return Err(DecodeError::InvalidOpcode);
                     } else {
                         (Opcode::VBROADCASTSD, if L {
-                            VEXOperandCode::G_E_ymm
+                            VEXOperandCode::G_ymm_E_xmm
                         } else {
                             VEXOperandCode::G_E_xmm
                         })
                     }
                     0x1A => (Opcode::VBROADCASTF128, if L {
-                        VEXOperandCode::G_M_ymm
+                        VEXOperandCode::G_ymm_M_xmm
                     } else {
                         instruction.opcode = Opcode::Invalid;
                         return Err(DecodeError::InvalidOpcode);
@@ -2515,8 +2543,14 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         })
                     },
                     0x46 => (Opcode::VPSRAVD, if L {
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
                         VEXOperandCode::G_V_E_ymm
                     } else {
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
                         VEXOperandCode::G_V_E_xmm
                     }),
                     0x47 => if instruction.prefixes.vex().w() {
@@ -2543,7 +2577,10 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         VEXOperandCode::G_E_xmm
                     }),
                     0x5A => (Opcode::VBROADCASTI128, if L {
-                        VEXOperandCode::G_E_ymm
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
+                        VEXOperandCode::G_ymm_M_xmm
                     } else {
                         instruction.opcode = Opcode::Invalid;
                         return Err(DecodeError::InvalidOpcode);
@@ -2560,31 +2597,31 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                     }),
                     0x8C => {
                         if instruction.prefixes.vex().w() {
-                            (Opcode::VPMASKMOVD, if L {
-                                VEXOperandCode::G_V_E_ymm
+                            (Opcode::VPMASKMOVQ, if L {
+                                VEXOperandCode::G_V_M_ymm
                             } else {
-                                VEXOperandCode::G_V_E_ymm
+                                VEXOperandCode::G_V_M_xmm
                             })
                         } else {
-                            (Opcode::VPMASKMOVQ, if L {
-                                VEXOperandCode::G_V_E_ymm
+                            (Opcode::VPMASKMOVD, if L {
+                                VEXOperandCode::G_V_M_ymm
                             } else {
-                                VEXOperandCode::G_V_E_ymm
+                                VEXOperandCode::G_V_M_xmm
                             })
                         }
                     },
                     0x8E => {
                         if instruction.prefixes.vex().w() {
-                            (Opcode::VPMASKMOVD, if L {
-                                VEXOperandCode::E_V_G_ymm
+                            (Opcode::VPMASKMOVQ, if L {
+                                VEXOperandCode::M_V_G_ymm
                             } else {
-                                VEXOperandCode::E_V_G_xmm
+                                VEXOperandCode::M_V_G_xmm
                             })
                         } else {
-                            (Opcode::VPMASKMOVQ, if L {
-                                VEXOperandCode::E_V_G_ymm
+                            (Opcode::VPMASKMOVD, if L {
+                                VEXOperandCode::M_V_G_ymm
                             } else {
-                                VEXOperandCode::E_V_G_xmm
+                                VEXOperandCode::M_V_G_xmm
                             })
                         }
                     },
@@ -2621,30 +2658,30 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                     0x92 => {
                         if instruction.prefixes.vex().w() {
                             (Opcode::VGATHERDPD, if L {
-                                VEXOperandCode::G_E_ymm_imm8
+                                VEXOperandCode::G_Ey_V_ymm
                             } else {
-                                VEXOperandCode::G_E_xmm_imm8
+                                VEXOperandCode::G_Ex_V_xmm
                             })
                         } else {
                             (Opcode::VGATHERDPS, if L {
-                                VEXOperandCode::G_E_ymm_imm8
+                                VEXOperandCode::G_Ey_V_ymm
                             } else {
-                                VEXOperandCode::G_E_xmm_imm8
+                                VEXOperandCode::G_Ex_V_xmm
                             })
                         }
                     },
                     0x93 => {
                         if instruction.prefixes.vex().w() {
                             (Opcode::VGATHERQPD, if L {
-                                VEXOperandCode::G_E_ymm_imm8
+                                VEXOperandCode::G_Ey_V_ymm
                             } else {
-                                VEXOperandCode::G_E_xmm_imm8
+                                VEXOperandCode::G_Ex_V_xmm
                             })
                         } else {
                             (Opcode::VGATHERQPS, if L {
-                                VEXOperandCode::G_E_ymm_imm8
+                                VEXOperandCode::G_Ey_V_ymm
                             } else {
-                                VEXOperandCode::G_E_xmm_imm8
+                                VEXOperandCode::G_Ex_V_xmm
                             })
                         }
                     },
@@ -3097,21 +3134,33 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                 // possibly valid!
                 match opc {
                     0x00 => (Opcode::VPERMQ, if L {
+                        if !instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
                         VEXOperandCode::G_E_ymm_imm8
                     } else {
                         instruction.opcode = Opcode::Invalid;
                         return Err(DecodeError::InvalidOpcode);
                     }),
                     0x01 => (Opcode::VPERMPD, if L {
+                        if !instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
                         VEXOperandCode::G_E_ymm_imm8
                     } else {
                         instruction.opcode = Opcode::Invalid;
                         return Err(DecodeError::InvalidOpcode);
                     }),
                     0x02 => (Opcode::VPBLENDD, if L {
-                        VEXOperandCode::G_V_E_ymm
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
+                        VEXOperandCode::G_V_E_ymm_imm8
                     } else {
-                        VEXOperandCode::G_V_E_xmm
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
+                        VEXOperandCode::G_V_E_xmm_imm8
                     }),
                     0x04 => (Opcode::VPERMILPS, if L {
                         VEXOperandCode::G_E_ymm_imm8
@@ -3124,6 +3173,9 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         VEXOperandCode::G_E_xmm_imm8
                     }),
                     0x06 => (Opcode::VPERM2F128, if L {
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
                         VEXOperandCode::G_V_E_ymm_imm8
                     } else {
                         instruction.opcode = Opcode::Invalid;
@@ -3264,7 +3316,7 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         return Err(DecodeError::InvalidOpcode);
                     }),
                     0x39 => (Opcode::VEXTRACTI128, if L {
-                        VEXOperandCode::V_xmm_G_ymm_E_ymm_imm8
+                        VEXOperandCode::E_xmm_G_ymm_imm8
                     } else {
                         instruction.opcode = Opcode::Invalid;
                         return Err(DecodeError::InvalidOpcode);
@@ -3292,6 +3344,9 @@ fn read_vex_instruction<T: Iterator<Item=u8>>(opcode_map: VEXOpcodeMap, bytes: &
                         VEXOperandCode::G_V_E_xmm_imm8
                     }),
                     0x46 => (Opcode::VPERM2I128, if L {
+                        if instruction.prefixes.vex().w() {
+                            return Err(DecodeError::InvalidOpcode);
+                        }
                         VEXOperandCode::G_V_E_ymm_imm8
                     } else {
                         instruction.opcode = Opcode::Invalid;
diff --git a/test/long_mode/mod.rs b/test/long_mode/mod.rs
index c6b895c..3c2861d 100644
--- a/test/long_mode/mod.rs
+++ b/test/long_mode/mod.rs
@@ -1430,12 +1430,29 @@ fn test_vex() {
     }
 
     // prefix 03
+    test_invalid(&[0xc4, 0b000_00011, 0b1_1111_001, 0x00, 0b11_001_010, 0x77]);
+    test_invalid(&[0xc4, 0b000_00011, 0b0_1111_101, 0x00, 0b11_001_010, 0x77]);
+    test_avx2(&[0xc4, 0b000_00011, 0b1_1111_101, 0x00, 0b11_001_010, 0x77], "vpermq ymm9, ymm10, 0x77");
+    test_invalid(&[0xc4, 0b000_00011, 0b1_1111_001, 0x01, 0b11_001_010, 0x77]);
+    test_invalid(&[0xc4, 0b000_00011, 0b0_1111_101, 0x01, 0b11_001_010, 0x77]);
+    test_avx2(&[0xc4, 0b000_00011, 0b1_1111_101, 0x01, 0b11_001_010, 0x77], "vpermpd ymm9, ymm10, 0x77");
+    test_invalid(&[0xc4, 0b000_00011, 0b1_1111_001, 0x02, 0b11_001_010, 0x77]);
+    test_invalid(&[0xc4, 0b000_00011, 0b1_1111_101, 0x02, 0b11_001_010, 0x77]);
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_001, 0x02, 0b11_001_010, 0x77], "vpblendd xmm9, xmm0, xmm10, 0x77");
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_001, 0x02, 0b00_001_010, 0x77], "vpblendd xmm9, xmm0, [r10], 0x77");
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_101, 0x02, 0b11_001_010, 0x77], "vpblendd ymm9, ymm0, ymm10, 0x77");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_101, 0x02, 0b00_001_010, 0x77], "vpblendd ymm9, ymm0, [r10], 0x77");
+
     test_instr(&[0xc4, 0b000_00011, 0b0_1111_001, 0x04, 0b11_001_010, 0x77], "vpermilps xmm9, xmm10, 0x77");
     test_instr(&[0xc4, 0b000_00011, 0b0_1111_101, 0x04, 0b11_001_010, 0x77], "vpermilps ymm9, ymm10, 0x77");
     test_instr(&[0xc4, 0b000_00011, 0b0_1111_001, 0x05, 0b11_001_010, 0x77], "vpermilpd xmm9, xmm10, 0x77");
     test_instr(&[0xc4, 0b000_00011, 0b0_1111_101, 0x05, 0b11_001_010, 0x77], "vpermilpd ymm9, ymm10, 0x77");
     test_invalid(&[0xc4, 0b000_00011, 0b0_1111_001, 0x06, 0b11_001_010, 0x77]);
     test_instr(&[0xc4, 0b000_00011, 0b0_1111_101, 0x06, 0b11_001_010, 0x77], "vperm2f128 ymm9, ymm0, ymm10, 0x77");
+    // TODO: ymmword ptr
+    test_instr(&[0xc4, 0b000_00011, 0b0_1111_101, 0x06, 0b00_001_010, 0x77], "vperm2f128 ymm9, ymm0, [r10], 0x77");
 
     test_instr(&[0xc4, 0b000_00011, 0b0_0111_001, 0x0c, 0b11_001_010, 0x77], "vblendps xmm9, xmm8, xmm10, 0x77");
     test_instr(&[0xc4, 0b000_00011, 0b0_0111_101, 0x0c, 0b11_001_010, 0x77], "vblendps ymm9, ymm8, ymm10, 0x77");
@@ -1487,6 +1504,13 @@ fn test_vex() {
     test_invalid(&[0xc4, 0b000_00011, 0b0_1111_001, 0x19, 0b11_001_010, 0x77]);
     test_invalid(&[0xc4, 0b000_00011, 0b1_1111_101, 0x19, 0b11_001_010, 0x77]);
 
+    test_invalid(&[0xc4, 0b000_00011, 0b1_0111_001, 0x18, 0b11_001_010, 0x77]);
+    test_avx2(&[0xc4, 0b000_00011, 0b0_0111_101, 0x38, 0b11_001_010, 0x77], "vinserti128 ymm9, ymm8, xmm10, 0x77");
+    test_invalid(&[0xc4, 0b000_00011, 0b1_0111_101, 0x18, 0b11_001_010, 0x77]);
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_101, 0x39, 0b11_001_010, 0x77], "vextracti128 xmm10, ymm9, 0x77");
+    test_invalid(&[0xc4, 0b000_00011, 0b0_1111_001, 0x19, 0b11_001_010, 0x77]);
+    test_invalid(&[0xc4, 0b000_00011, 0b1_1111_101, 0x19, 0b11_001_010, 0x77]);
+
     test_instr(&[0xc4, 0b000_00011, 0b0_0111_001, 0x20, 0b11_001_010, 0x77], "vpinsrb xmm9, xmm8, r10d, 0x77");
     // TODO: byte ptr
     test_instr(&[0xc4, 0b000_00011, 0b0_0111_001, 0x20, 0b00_001_010, 0x77], "vpinsrb xmm9, xmm8, [r10], 0x77");
@@ -1508,6 +1532,13 @@ fn test_vex() {
     test_invalid(&[0xc4, 0b000_00011, 0b0_0111_101, 0x41, 0b11_001_010, 0x77]);
     test_instr(&[0xc4, 0b000_00011, 0b0_0111_001, 0x42, 0b11_001_010, 0x77], "vmpsadbw xmm9, xmm8, xmm10, 0x77");
     test_avx2(&[0xc4, 0b000_00011, 0b0_0111_101, 0x42, 0b11_001_010, 0x77], "vmpsadbw ymm9, ymm8, ymm10, 0x77");
+
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_101, 0x46, 0b11_001_010, 0x77], "vperm2i128 ymm9, ymm0, ymm10, 0x77");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00011, 0b0_1111_101, 0x46, 0b00_001_010, 0x77], "vperm2i128 ymm9, ymm0, [r10], 0x77");
+    test_invalid(&[0xc4, 0b000_00011, 0b0_1111_001, 0x46, 0b11_001_010, 0x77]);
+    test_invalid(&[0xc4, 0b000_00011, 0b1_1111_101, 0x46, 0b11_001_010, 0x77]);
+
     test_instr(&[0xc4, 0b000_00011, 0b0_0111_001, 0x4c, 0b11_001_010, 0x77], "vpblendvb xmm9, xmm8, xmm10, xmm7");
     test_avx2(&[0xc4, 0b000_00011, 0b0_0111_101, 0x4c, 0b11_001_010, 0x77], "vpblendvb ymm9, ymm8, ymm10, ymm7");
     test_invalid(&[0xc4, 0b000_00011, 0b1_0111_001, 0x4c, 0b11_001_010, 0x77]);
@@ -1570,6 +1601,12 @@ fn test_vex() {
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_001, 0x0f, 0b11_001_010], "vtestpd xmm9, xmm10");
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x0f, 0b11_001_010], "vtestpd ymm9, ymm10");
 
+    test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x16, 0b11_001_010], "vpermps ymm9, ymm0, ymm10");
+    // TODO: ymmword ptr
+    test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x16, 0b00_001_010], "vpermps ymm9, ymm0, [r10]");
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_001, 0x16, 0b00_011_010]);
+    test_invalid(&[0xc4, 0b000_00010, 0b1_1111_101, 0x16, 0b00_011_010]);
+
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_001, 0x17, 0b11_001_010], "vptest xmm9, xmm10");
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x17, 0b11_001_010], "vptest ymm9, ymm10");
     test_invalid(&[0xc4, 0b000_00010, 0b0_0111_001, 0x17, 0b11_001_010]);
@@ -1589,12 +1626,18 @@ fn test_vex() {
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x1a, 0b00_001_010], "vbroadcastf128 ymm9, [r10]");
     test_invalid(&[0xc4, 0b000_00010, 0b1_0111_101, 0x1a, 0b00_001_010]);
     test_invalid(&[0xc4, 0b000_00010, 0b1_0111_001, 0x1a, 0b00_001_010]);
-    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x18, 0b11_001_010], "vbroadcastss xmm9, xmm10");
+    // TODO: xmmword ptr
+    test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x5a, 0b00_001_010], "vbroadcasti128 ymm9, [r10]");
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_101, 0x5a, 0b11_001_010]);
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_001, 0x5a, 0b00_001_010]);
+    test_invalid(&[0xc4, 0b000_00010, 0b1_1111_101, 0x5a, 0b00_001_010]);
+    test_instr(&[0xc4, 0b000_00010, 0b0_1111_001, 0x18, 0b11_001_010], "vbroadcastss xmm9, xmm10");
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x18, 0b11_001_010], "vbroadcastss ymm9, xmm10");
     // TODO: dword ptr
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x18, 0b00_001_010], "vbroadcastss ymm9, [r10]");
     test_invalid(&[0xc4, 0b000_00010, 0b1_0111_001, 0x18, 0b11_001_010]);
     test_invalid(&[0xc4, 0b000_00010, 0b1_0111_101, 0x18, 0b11_001_010]);
-    test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x19, 0b11_001_010], "vbroadcastsd ymm9, ymm10");
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x19, 0b11_001_010], "vbroadcastsd ymm9, xmm10");
     // TODO: qword ptr
     test_instr(&[0xc4, 0b000_00010, 0b0_1111_101, 0x19, 0b00_001_010], "vbroadcastsd ymm9, [r10]");
     test_invalid(&[0xc4, 0b000_00010, 0b1_1111_101, 0x19, 0b11_001_010]);
@@ -1675,6 +1718,8 @@ fn test_vex() {
     test_invalid(&[0xc4, 0b000_00010, 0b0_0111_001, 0x35, 0b11_001_010]);
     test_invalid(&[0xc4, 0b000_00010, 0b0_0111_101, 0x35, 0b11_001_010]);
 
+    test_invalid(&[0xc4, 0b000_00010, 0b0_0111_001, 0x36, 0b11_001_010]);
+    test_avx2(&[0xc4, 0b000_00010, 0b0_0111_101, 0x36, 0b11_001_010], "vpermd ymm9, ymm8, ymm10");
     test_instr(&[0xc4, 0b000_00010, 0b0_0111_001, 0x37, 0b11_001_010], "vpcmpgtq xmm9, xmm8, xmm10");
     test_avx2(&[0xc4, 0b000_00010, 0b0_0111_101, 0x37, 0b11_001_010], "vpcmpgtq ymm9, ymm8, ymm10");
     test_instr(&[0xc4, 0b000_00010, 0b0_0111_001, 0x38, 0b11_001_010], "vpminsb xmm9, xmm8, xmm10");
@@ -1703,6 +1748,97 @@ fn test_vex() {
 //    test_instr(&[0xc4, 0b000_00010, 0b1_0111_001, 0x42, 0b11_001_010], "vphminposuw xmm");
 //    test_invalid(&[0xc4, 0b000_00010, 0b1_0111_101, 0x41, 0b11_001_010]);
 
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x45, 0b00_001_010], "vpsrlvd xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x45, 0b00_001_010], "vpsrlvd ymm9, ymm0, [r10]");
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x45, 0b11_001_010], "vpsrlvd xmm9, xmm0, xmm10");
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x45, 0b11_001_010], "vpsrlvd ymm9, ymm0, ymm10");
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x45, 0b00_001_010], "vpsrlvq xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x45, 0b00_001_010], "vpsrlvq ymm9, ymm0, [r10]");
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x45, 0b11_001_010], "vpsrlvq xmm9, xmm0, xmm10");
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x45, 0b11_001_010], "vpsrlvq ymm9, ymm0, ymm10");
+
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x46, 0b00_001_010], "vpsravd xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x46, 0b00_001_010], "vpsravd ymm9, ymm0, [r10]");
+    test_invalid(&[0xc4, 0b000_00010, 0b1_1111_001, 0x46, 0b00_001_010]);
+    test_invalid(&[0xc4, 0b000_00010, 0b1_1111_101, 0x46, 0b00_001_010]);
+
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x47, 0b00_001_010], "vpsllvd xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x47, 0b00_001_010], "vpsllvd ymm9, ymm0, [r10]");
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x47, 0b11_001_010], "vpsllvd xmm9, xmm0, xmm10");
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x47, 0b11_001_010], "vpsllvd ymm9, ymm0, ymm10");
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x47, 0b00_001_010], "vpsllvq xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x47, 0b00_001_010], "vpsllvq ymm9, ymm0, [r10]");
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x47, 0b11_001_010], "vpsllvq xmm9, xmm0, xmm10");
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x47, 0b11_001_010], "vpsllvq ymm9, ymm0, ymm10");
+
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x8c, 0b00_001_010], "vpmaskmovd xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x8c, 0b00_001_010], "vpmaskmovd ymm9, ymm0, [r10]");
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_001, 0x8c, 0b11_001_010]);
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x8c, 0b00_001_010], "vpmaskmovq xmm9, xmm0, [r10]");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x8c, 0b00_001_010], "vpmaskmovq ymm9, ymm0, [r10]");
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_001, 0x8c, 0b11_001_010]);
+
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x8e, 0b00_001_010], "vpmaskmovd [r10], xmm0, xmm9");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x8e, 0b00_001_010], "vpmaskmovd [r10], ymm0, ymm9");
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_001, 0x8e, 0b11_001_010]);
+    // TODO: xmmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x8e, 0b00_001_010], "vpmaskmovq [r10], xmm0, xmm9");
+    // TODO: ymmword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x8e, 0b00_001_010], "vpmaskmovq [r10], ymm0, ymm9");
+    test_invalid(&[0xc4, 0b000_00010, 0b0_1111_001, 0x8e, 0b11_001_010]);
+
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x90, 0b00_000_100, 0xa1], "vpgatherdd xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x90, 0b00_000_100, 0xa1], "vpgatherdd ymm8, [r9 + ymm12 * 4], ymm0");
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x90, 0b00_000_100, 0xa1], "vpgatherdq xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x90, 0b00_000_100, 0xa1], "vpgatherdq ymm8, [r9 + ymm12 * 4], ymm0");
+
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x91, 0b00_000_100, 0xa1], "vpgatherqd xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x91, 0b00_000_100, 0xa1], "vpgatherqd ymm8, [r9 + ymm12 * 4], ymm0");
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x91, 0b00_000_100, 0xa1], "vpgatherqq xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x91, 0b00_000_100, 0xa1], "vpgatherqq ymm8, [r9 + ymm12 * 4], ymm0");
+
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x92, 0b00_000_100, 0xa1], "vgatherdps xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x92, 0b00_000_100, 0xa1], "vgatherdps ymm8, [r9 + ymm12 * 4], ymm0");
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x92, 0b00_000_100, 0xa1], "vgatherdpd xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x92, 0b00_000_100, 0xa1], "vgatherdpd ymm8, [r9 + ymm12 * 4], ymm0");
+
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_001, 0x93, 0b00_000_100, 0xa1], "vgatherqps xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b0_1111_101, 0x93, 0b00_000_100, 0xa1], "vgatherqps ymm8, [r9 + ymm12 * 4], ymm0");
+    // TODO: dword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_001, 0x93, 0b00_000_100, 0xa1], "vgatherqpd xmm8, [r9 + xmm12 * 4], xmm0");
+    // TODO: qword ptr
+    test_avx2(&[0xc4, 0b000_00010, 0b1_1111_101, 0x93, 0b00_000_100, 0xa1], "vgatherqpd ymm8, [r9 + ymm12 * 4], ymm0");
+
     test_instr_vex_aesni(&[0xc4, 0b000_00010, 0b0_1111_001, 0xdb, 0b11_001_010], "vaesimc xmm9, xmm10");
     test_invalid(&[0xc4, 0b000_00010, 0b0_0111_101, 0xdb, 0b11_001_010]);
     test_invalid(&[0xc4, 0b000_00010, 0b1_0111_101, 0xdb, 0b11_001_010]);