ssse3, some missing sse4.1, and pextrw operands

3 files changed, 88 insertions, 9 deletions

diff --git a/src/long_mode/display.rs b/src/long_mode/display.rs
index 807565c..9614374 100644
--- a/src/long_mode/display.rs
+++ b/src/long_mode/display.rs
@@ -1074,7 +1074,7 @@ impl fmt::Display for Opcode {
             &Opcode::PMINUD => write!(f, "pminud"),
             &Opcode::PMINUW => write!(f, "pminuw"),
             &Opcode::BLENDW => write!(f, "blendw"),
-            &Opcode::BLENDDVB => write!(f, "blenddvb"),
+            &Opcode::BLENDVB => write!(f, "blendvb"),
             &Opcode::BLENDVPS => write!(f, "blendvps"),
             &Opcode::BLENDVPD => write!(f, "blendvpd"),
             &Opcode::BLENDPS => write!(f, "blendps"),
@@ -1087,7 +1087,7 @@ impl fmt::Display for Opcode {
             &Opcode::PSIGND => write!(f, "psignd"),
             &Opcode::PSIGNB => write!(f, "psignb"),
             &Opcode::PSHUFB => write!(f, "pshufb"),
-            &Opcode::PMULHRSU => write!(f, "pmulhrsu"),
+            &Opcode::PMULHRSW => write!(f, "pmulhrsw"),
             &Opcode::PMADDUBSW => write!(f, "pmaddubsw"),
             &Opcode::PABSD => write!(f, "pabsd"),
             &Opcode::PABSW => write!(f, "pabsw"),
@@ -1244,7 +1244,7 @@ impl <T: fmt::Write, Color: fmt::Display, Y: YaxColors<Color>> Colorize<T, Color
             Opcode::VPMULUDQ |
             Opcode::PCLMULQDQ |
             Opcode::PMULDQ |
-            Opcode::PMULHRSU |
+            Opcode::PMULHRSW |
             Opcode::PMULLD |
             Opcode::VPSUBB |
             Opcode::VPSUBD |
@@ -1515,7 +1515,7 @@ impl <T: fmt::Write, Color: fmt::Display, Y: YaxColors<Color>> Colorize<T, Color
             Opcode::VBLENDPS |
             Opcode::VBLENDVPD |
             Opcode::VBLENDVPS |
-            Opcode::BLENDDVB |
+            Opcode::BLENDVB |
             Opcode::BLENDPD |
             Opcode::BLENDPS |
             Opcode::BLENDVPD |
diff --git a/src/long_mode/mod.rs b/src/long_mode/mod.rs
index ac9edf5..d709e53 100644
--- a/src/long_mode/mod.rs
+++ b/src/long_mode/mod.rs
@@ -1346,7 +1346,7 @@ pub enum Opcode {
     PMINUD,
     PMINUW,
     BLENDW,
-    BLENDDVB,
+    BLENDVB,
     BLENDVPS,
     BLENDVPD,
     BLENDPS,
@@ -1359,7 +1359,7 @@ pub enum Opcode {
     PSIGND,
     PSIGNB,
     PSHUFB,
-    PMULHRSU,
+    PMULHRSW,
     PMADDUBSW,
     PABSD,
     PABSW,
@@ -2255,7 +2255,7 @@ impl InstDecoder {
             Opcode::PABSW |
             Opcode::PABSD |
             Opcode::PMADDUBSW |
-            Opcode::PMULHRSU |
+            Opcode::PMULHRSW |
             Opcode::PSHUFB |
             Opcode::PSIGNB |
             Opcode::PSIGNW |
@@ -2274,7 +2274,7 @@ impl InstDecoder {
             Opcode::BLENDPS |
             Opcode::BLENDVPD |
             Opcode::BLENDVPS |
-            Opcode::BLENDDVB |
+            Opcode::BLENDVB |
             Opcode::BLENDW |
             Opcode::PMINUW |
             Opcode::PMINUD |
@@ -3229,7 +3229,6 @@ pub enum OperandCode {
     Yv_Xv,
     G_E_q,
     E_G_q,
-    Rv_Gmm_Ib,
     G_mm_Ew_Ib,
     Mq_Dq,
     ModRM_0x0f38,
@@ -3348,6 +3347,7 @@ pub enum OperandCode {
     G_xmm_E_mm = 0xed,
     G_xmm_U_mm = 0x1ed,
     U_mm_G_xmm = 0x2ed,
+    Rv_Gmm_Ib = 0x3ed,
     G_xmm_Edq = 0xef,
     G_U_mm = 0xf1,
     Ev_Gv_Ib = 0xf3,
@@ -5909,6 +5909,19 @@ fn unlikely_operands<T: Iterator<Item=u8>>(decoder: &InstDecoder, mut bytes_iter
                 }
             };
             instruction.opcode = match opcode {
+                0x00 => Opcode::PSHUFB,
+                0x01 => Opcode::PHADDW,
+                0x02 => Opcode::PHADDD,
+                0x03 => Opcode::PHADDSW,
+                0x04 => Opcode::PMADDUBSW,
+                0x05 => Opcode::PHSUBW,
+                0x06 => Opcode::PHSUBD,
+                0x07 => Opcode::PHSUBSW,
+                0x08 => Opcode::PSIGNB,
+                0x09 => Opcode::PSIGNW,
+                0x0a => Opcode::PSIGND,
+                0x0b => Opcode::PMULHRSW,
+
                 0xc8 => Opcode::SHA1NEXTE,
                 0xc9 => Opcode::SHA1MSG1,
                 0xca => Opcode::SHA1MSG2,
@@ -6144,6 +6157,25 @@ fn unlikely_operands<T: Iterator<Item=u8>>(decoder: &InstDecoder, mut bytes_iter
         OperandCode::ModRM_0x660f38 => {
             let op = bytes_iter.next().ok_or(DecodeError::ExhaustedInput).map(|b| { *length += 1; b })?;
             match op {
+                0x00 => { instruction.opcode = Opcode::PSHUFB; }
+                0x01 => { instruction.opcode = Opcode::PHADDW; }
+                0x02 => { instruction.opcode = Opcode::PHADDD; }
+                0x03 => { instruction.opcode = Opcode::PHADDSW; }
+                0x04 => { instruction.opcode = Opcode::PMADDUBSW; }
+                0x05 => { instruction.opcode = Opcode::PHSUBW; }
+                0x06 => { instruction.opcode = Opcode::PHSUBD; }
+                0x07 => { instruction.opcode = Opcode::PHSUBSW; }
+                0x08 => { instruction.opcode = Opcode::PSIGNB; }
+                0x09 => { instruction.opcode = Opcode::PSIGNW; }
+                0x0a => { instruction.opcode = Opcode::PSIGND; }
+                0x0b => { instruction.opcode = Opcode::PMULHRSW; }
+
+                0x1c => { instruction.opcode = Opcode::PABSB; }
+                0x1d => { instruction.opcode = Opcode::PABSW; }
+                0x1e => { instruction.opcode = Opcode::PABSD; }
+
+                0x40 => { instruction.opcode = Opcode::PMULLD; }
+
                 0xdb => { instruction.opcode = Opcode::AESIMC; }
                 0xdc => { instruction.opcode = Opcode::AESENC; }
                 0xdd => { instruction.opcode = Opcode::AESENCLAST; }
@@ -6171,6 +6203,14 @@ fn unlikely_operands<T: Iterator<Item=u8>>(decoder: &InstDecoder, mut bytes_iter
         OperandCode::ModRM_0x660f3a => {
             let op = bytes_iter.next().ok_or(DecodeError::ExhaustedInput).map(|b| { *length += 1; b })?;
             match op {
+                0x0c => { instruction.opcode = Opcode::BLENDPS; }
+                0x0d => { instruction.opcode = Opcode::BLENDPD; }
+                0x0f => { instruction.opcode = Opcode::PALIGNR; }
+                0x10 => { instruction.opcode = Opcode::BLENDVB; }
+
+                0x14 => { instruction.opcode = Opcode::BLENDVPS; }
+                0x15 => { instruction.opcode = Opcode::BLENDVPD; }
+
                 0xcc => {
                     instruction.opcode = Opcode::SHA1RNDS4;
 
@@ -6562,6 +6602,20 @@ fn unlikely_operands<T: Iterator<Item=u8>>(decoder: &InstDecoder, mut bytes_iter
                 }
             }
         },
+        OperandCode::Rv_Gmm_Ib => {
+            instruction.operands[1] = mem_oper;
+            instruction.operands[2] = OperandSpec::ImmU8;
+            instruction.imm =
+                read_num(&mut bytes_iter, 1)? as u64;
+            *length += 1;
+            instruction.modrm_rrr.bank = RegisterBank::D;
+            if mem_oper == OperandSpec::RegMMM {
+                instruction.modrm_mmm.bank = RegisterBank::MM;
+                instruction.modrm_mmm.num &= 0b111;
+            } else {
+                return Err(DecodeError::InvalidOperand);
+            }
+        }
         OperandCode::U_mm_G_xmm => {
             instruction.operands[1] = mem_oper;
             instruction.modrm_mmm.bank = RegisterBank::X;
diff --git a/test/long_mode/mod.rs b/test/long_mode/mod.rs
index 50774ff..545a2d5 100644
--- a/test/long_mode/mod.rs
+++ b/test/long_mode/mod.rs
@@ -130,10 +130,14 @@ fn test_mmx() {
     test_display(&[0x0f, 0x6f, 0xe9], "movq mm5, mm1");
     test_display(&[0x0f, 0xe5, 0x3d, 0xaa, 0xbb, 0xcc, 0x77], "pmulhw mm7, [rip + 0x77ccbbaa]");
 
+    test_display(&[0x0f, 0x38, 0x00, 0xda], "pshufb mm3, mm2");
+
     test_display(&[0x0f, 0x74, 0xc2], "pcmpeqb mm0, mm2");
     test_display(&[0x0f, 0x75, 0xc2], "pcmpeqw mm0, mm2");
     test_display(&[0x0f, 0x76, 0xc2], "pcmpeqd mm0, mm2");
 
+    test_display(&[0x0f, 0xc5, 0xd1, 0x00], "pextrw edx, mm1, 0x0");
+
     test_display(&[0x0f, 0xd8, 0xc2], "psubusb mm0, mm2");
     test_display(&[0x0f, 0xd9, 0xc2], "psubusw mm0, mm2");
     test_display(&[0x0f, 0xda, 0xc2], "pminub mm0, mm2");
@@ -582,6 +586,26 @@ fn test_sse3() {
 }
 
 #[test]
+fn test_ssse3() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_ssse3(), bytes, text);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+        // avx doesn't imply older instructions are necessarily valid
+        test_invalid_under(&InstDecoder::minimal().with_avx(), bytes);
+        // sse4 doesn't imply older instructions are necessarily valid
+        test_invalid_under(&InstDecoder::minimal().with_sse4_1(), bytes);
+        test_invalid_under(&InstDecoder::minimal().with_sse4_2(), bytes);
+    }
+
+    #[allow(unused)]
+    fn test_instr_invalid(bytes: &[u8]) {
+        test_invalid_under(&InstDecoder::minimal().with_ssse3(), bytes);
+        test_invalid_under(&InstDecoder::default(), bytes);
+    }
+    test_instr(&[0x66, 0x0f, 0x38, 0x00, 0xda], "pshufb xmm3, xmm2");
+}
+
+#[test]
 fn test_0f01() {
     // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
     // Number"
@@ -1214,6 +1238,7 @@ fn prefixed_0f() {
     test_display(&[0x0f, 0x6e, 0xc2], "movd mm0, edx");
     test_display(&[0x0f, 0x6f, 0x00], "movq mm0, [rax]");
     test_display(&[0x0f, 0x6f, 0xc2], "movq mm0, mm2");
+    test_display(&[0x0f, 0x6f, 0xfb], "movq mm7, mm3");
     test_display(&[0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [rax], 0x7f");
     test_display(&[0x4f, 0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [r8], 0x7f");
     test_invalid(&[0x0f, 0x71, 0x00, 0x7f]);