From 0551aa5998226769e64b9e47ba1906c7a0cc3270 Mon Sep 17 00:00:00 2001
From: iximeow <me@iximeow.net>
Date: Sun, 22 Mar 2020 13:41:05 -0700
Subject: yaxpeax-x86 decodes in 32-bit mode now

---
 test/long_mode/mod.rs          | 1094 +++++++++++++++++++++++++++++++++++++++
 test/long_mode/regspec.rs      |   22 +
 test/protected_mode/mod.rs     | 1035 +++++++++++++++++++++++++++++++++++++
 test/protected_mode/regspec.rs |   20 +
 test/regspec.rs                |   22 -
 test/test.rs                   | 1096 +---------------------------------------
 6 files changed, 2173 insertions(+), 1116 deletions(-)
 create mode 100644 test/long_mode/mod.rs
 create mode 100644 test/long_mode/regspec.rs
 create mode 100644 test/protected_mode/mod.rs
 create mode 100644 test/protected_mode/regspec.rs
 delete mode 100644 test/regspec.rs

(limited to 'test')

diff --git a/test/long_mode/mod.rs b/test/long_mode/mod.rs
new file mode 100644
index 0000000..33de87f
--- /dev/null
+++ b/test/long_mode/mod.rs
@@ -0,0 +1,1094 @@
+mod regspec;
+
+use std::fmt::Write;
+
+use yaxpeax_arch::{Decoder, LengthedInstruction};
+use yaxpeax_x86::long_mode::{DecodeError, InstDecoder, Opcode};
+
+fn test_invalid(data: &[u8]) {
+    test_invalid_under(&InstDecoder::default(), data);
+}
+
+fn test_invalid_under(decoder: &InstDecoder, data: &[u8]) {
+    if let Ok(inst) = decoder.decode(data.into_iter().cloned()) {
+        assert_eq!(inst.opcode, Opcode::Invalid, "decoded {:?} from {:02x?} under decoder {}", inst.opcode, data, decoder);
+    } else {
+        // this is fine
+    }
+}
+
+fn test_display(data: &[u8], expected: &'static str) {
+    test_display_under(&InstDecoder::default(), data, expected);
+}
+
+fn test_display_under(decoder: &InstDecoder, data: &[u8], expected: &'static str) {
+    let mut hex = String::new();
+    for b in data {
+        write!(hex, "{:02x}", b).unwrap();
+    }
+    match decoder.decode(data.into_iter().map(|x| *x)) {
+        Ok(instr) => {
+            let text = format!("{}", instr);
+            assert!(
+                text == expected,
+                "display error for {}:\n  decoded: {:?} under decoder {}\n displayed: {}\n expected: {}\n",
+                hex,
+                instr,
+                decoder,
+                text,
+                expected
+            );
+            // while we're at it, test that the instruction is as long, and no longer, than its
+            // input
+            assert_eq!(instr.len() as usize, data.len(), "instruction length is incorrect, wanted instruction {}", expected);
+        },
+        Err(e) => {
+            assert!(false, "decode error ({}) for {} under decoder {}:\n  expected: {}\n", e, hex, decoder, expected);
+        }
+    }
+}
+
+#[test]
+fn test_mmx() {
+    test_display(&[0x4f, 0x0f, 0x7e, 0xcf], "movd r15, mm1");
+    test_display(&[0x41, 0x0f, 0x7e, 0xcf], "movd r15d, mm1");
+    test_display(&[0x4f, 0x0f, 0x7f, 0xcf], "movq mm7, mm1");
+    test_display(&[0x4f, 0x0f, 0x7f, 0x0f], "movq [r15], mm1");
+    test_display(&[0x0f, 0xc4, 0xc0, 0x14], "pinsrw mm0, eax, 0x14");
+    test_display(&[0x4f, 0x0f, 0xc4, 0xc0, 0x14], "pinsrw mm0, r8d, 0x14");
+    test_display(&[0x4f, 0x0f, 0xc4, 0x00, 0x14], "pinsrw mm0, [r8], 0x14");
+    test_display(&[0x4f, 0x0f, 0xd1, 0xcf], "psrlw mm1, mm7");
+    test_display(&[0x4f, 0x0f, 0xd1, 0x00], "psrlw mm0, [r8]");
+    test_invalid(&[0x4f, 0x0f, 0xd7, 0x00]);
+    test_display(&[0x4f, 0x0f, 0xd7, 0xcf], "pmovmskb r9d, mm7");
+}
+
+#[test]
+fn test_cvt() {
+    test_display(&[0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm7");
+    test_display(&[0x48, 0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm7");
+    test_display(&[0x4f, 0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm15");
+    test_display(&[0x4f, 0x0f, 0x2a, 0xcf], "cvtpi2ps xmm9, mm7");
+    test_display(&[0x4f, 0x0f, 0x2a, 0x00], "cvtpi2ps xmm8, [r8]");
+    test_display(&[0x4f, 0x66, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm1, mm7");
+    test_display(&[0x66, 0x4f, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm9, mm7");
+    test_display(&[0x4f, 0xf3, 0x0f, 0x2a, 0xcf], "cvtsi2ss xmm1, edi");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0xcf], "cvtsi2ss xmm9, r15");
+    test_display(&[0x4f, 0xf2, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm1, edi");
+    test_display(&[0xf2, 0x4f, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm9, r15");
+    test_display(&[0x4f, 0xf2, 0x0f, 0x2a, 0x00], "cvtsi2sd xmm0, [rax]");
+    test_display(&[0xf2, 0x4f, 0x0f, 0x2a, 0x00], "cvtsi2sd xmm8, [r8]");
+    test_display(&[0x4f, 0xf3, 0x0f, 0x2a, 0x00], "cvtsi2ss xmm0, [rax]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0x00], "cvtsi2ss xmm8, [r8]");
+    test_display(&[0x4f, 0x66, 0x0f, 0x2a, 0x00], "cvtpi2pd xmm0, [rax]");
+    test_display(&[0x66, 0x4f, 0x0f, 0x2a, 0x00], "cvtpi2pd xmm8, [r8]");
+}
+
+#[test]
+fn test_aesni() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_aesni(), bytes, text);
+        test_display_under(&InstDecoder::default(), bytes, text);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+    }
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdb, 0x0f], "aesimc xmm1, [rdi]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdb, 0xcf], "aesimc xmm9, xmm15");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdc, 0x0f], "aesenc xmm1, [rdi]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdc, 0xcf], "aesenc xmm9, xmm15");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdd, 0x0f], "aesenclast xmm1, [rdi]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdd, 0xcf], "aesenclast xmm9, xmm15");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xde, 0x0f], "aesdec xmm1, [rdi]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xde, 0xcf], "aesdec xmm9, xmm15");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdf, 0x0f], "aesdeclast xmm1, [rdi]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdf, 0xcf], "aesdeclast xmm9, xmm15");
+
+    test_instr(&[0x66, 0x0f, 0x3a, 0xdf, 0x0f, 0xaa], "aeskeygenassist xmm1, [rdi], 0xaa");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x3a, 0xdf, 0xcf, 0xaa], "aeskeygenassist xmm9, xmm15, 0xaa");
+}
+
+#[test]
+fn test_sse2() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        // sse and sse2 are part of amd64, so x86_64, meaning even the minimal decoder must support
+        // them.
+        test_display_under(&InstDecoder::minimal(), bytes, text);
+    }
+
+    test_instr(&[0xf2, 0x0f, 0x10, 0x0c, 0xc7], "movsd xmm1, [rdi + rax * 8]");
+    test_instr(&[0xf2, 0x0f, 0x11, 0x0c, 0xc7], "movsd [rdi + rax * 8], xmm1");
+    test_instr(&[0x66, 0x0f, 0x11, 0x0c, 0xc7], "movupd [rdi + rax * 8], xmm1");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x12, 0xc3], "movhlps xmm8, xmm11"); // reg-reg form is movhlps
+    test_instr(&[0x66, 0x4f, 0x0f, 0x12, 0x03], "movlpd xmm8, [r11]"); // reg-mem is movlpd
+    test_instr(&[0x66, 0x4f, 0x0f, 0x13, 0x03], "movlpd [r11], xmm8");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x13, 0xc3]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x14, 0x03], "unpcklpd xmm8, [r11]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x14, 0xc3], "unpcklpd xmm8, xmm11");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x15, 0x03], "unpckhpd xmm8, [r11]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x15, 0xc3], "unpckhpd xmm8, xmm11");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x16, 0x03], "movhpd xmm8, [r11]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x16, 0xc3], "movlhps xmm8, xmm11");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x17, 0x03], "movhpd [r11], xmm8");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x17, 0xc3]);
+
+    test_instr(&[0x66, 0x4f, 0x0f, 0x28, 0xd0], "movapd xmm10, xmm8");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8]");
+
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm9, mm7");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2a, 0x0f], "cvtpi2pd xmm9, [r15]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm9, r15");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x2a, 0x0f], "cvtsi2sd xmm9, [r15]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2b, 0x0f], "movntpd [r15], xmm9");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2c, 0xcf], "cvttpd2pi mm1, xmm15");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2c, 0x0f], "cvttpd2pi mm1, [r15]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x2c, 0xcf], "cvttsd2si xmm9, xmm15");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x2c, 0x0f], "cvttsd2si xmm9, [r15]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2d, 0xcf], "cvtpd2pi mm1, xmm15");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2d, 0x0f], "cvtpd2pi mm1, [r15]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x2d, 0xcf], "cvtsd2si xmm9, xmm15");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x2d, 0x0f], "cvtsd2si xmm9, [r15]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2e, 0xcf], "ucomisd xmm9, xmm15");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2e, 0x0f], "ucomisd xmm9, [r15]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2f, 0xcf], "comisd xmm9, xmm15");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x2f, 0x0f], "comisd xmm9, [r15]");
+
+    /*
+     * .... 660f38
+     * .... 660f7f
+     */
+
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x50, 0x01]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x50, 0xc1], "movmskpd r8d, xmm9");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x51, 0x01], "sqrtpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x51, 0x01], "sqrtsd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x52, 0x01], "rsqrtps xmm8, [r9]"); // note: NOT "rsqrtpd" - no such instruction exists, so fall back to just 0f52 parse.
+    test_instr(&[0x66, 0x4f, 0x0f, 0x53, 0x01], "rcpps xmm8, [r9]"); // note: NOT "rcppd" - no such instruction exists, so fall back to just 0f53 parse.
+    test_instr(&[0x66, 0x4f, 0x0f, 0x54, 0x01], "andpd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x55, 0x01], "andnpd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x56, 0x01], "orpd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x57, 0x01], "xorpd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x58, 0x01], "addpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x58, 0x01], "addsd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x59, 0x01], "mulpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x59, 0x01], "mulsd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x5a, 0x01], "cvtpd2ps xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x5a, 0x01], "cvtsd2ss xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x5b, 0x01], "cvtps2dq xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x5c, 0x01], "subpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x5c, 0x01], "subsd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x5d, 0x01], "minpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x5d, 0x01], "minsd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x5e, 0x01], "divpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x5e, 0x01], "divsd xmm8, [r9]");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x5f, 0x01], "maxpd xmm8, [r9]");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x5f, 0x01], "maxsd xmm8, [r9]");
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x60, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpcklbw xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x61, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpcklwd xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x62, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckldq xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x63, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "packsswb xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x64, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "pcmpgtb xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x65, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "pcmpgtw xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x66, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "pcmpgtd xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x67, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "packuswb xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x68, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhbw xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x69, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhwd xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x6a, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhdq xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x6b, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "packssdw xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x6c, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpcklqdq xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x6d, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhqdq xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    // TODO: this needs to be clear that the operand is `dword`
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x6e, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movq xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x6f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqa xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+
+    test_instr(&[0x66, 0x48, 0x0f, 0x6e, 0xc0], "movq xmm0, rax");
+    test_instr(&[0x66, 0x0f, 0x70, 0xc0, 0x4e], "pshufd xmm0, xmm0, 0x4e");
+    test_instr(&[0xf2, 0x0f, 0x70, 0xc0, 0x4e], "pshuflw xmm0, xmm0, 0x4e");
+    test_instr(&[0xf3, 0x0f, 0x70, 0xc0, 0x4e], "pshufhw xmm0, xmm0, 0x4e");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x71, 0x10, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x71, 0xd0, 0x8f], "psrlw xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x71, 0x20, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x71, 0xe0, 0x8f], "psraw xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x71, 0x30, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x71, 0xf0, 0x8f], "psllw xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x72, 0x10, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x72, 0xd0, 0x8f], "psrld xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x72, 0x20, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x72, 0xe0, 0x8f], "psrad xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x72, 0x30, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x72, 0xf0, 0x8f], "pslld xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x10, 0x8f]);
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x18, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xd0, 0x8f], "psrlq xmm0, 0x8f");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xd8, 0x8f], "psrldq xmm0, 0x8f");
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x30, 0x8f]);
+    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x38, 0x8f]);
+    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xf0, 0x8f], "psllq xmm0, 0x8f");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xf8, 0x8f], "pslldq xmm0, 0x8f");
+    test_instr(&[0x66, 0x0f, 0x7e, 0xc1], "movd ecx, xmm0");
+    test_instr(&[0x66, 0x48, 0x0f, 0x7e, 0xc1], "movq rcx, xmm0");
+    test_instr(&[0x66, 0x48, 0x0f, 0x7e, 0x01], "movd [rcx], xmm0");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x7e, 0xc1], "movq r9, xmm8");
+    test_instr(
+        &[0x66, 0x4f, 0x0f, 0x7f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqa [r12 + r11 * 4 - 0x334455cc], xmm11"
+    );
+
+    test_instr(&[0x66, 0x0f, 0xc2, 0xc3, 0x08], "cmppd xmm0, xmm3, 0x8");
+    test_instr(&[0x66, 0x4f, 0x0f, 0xc2, 0xc3, 0x08], "cmppd xmm8, xmm11, 0x8");
+    test_instr(&[0x66, 0x4f, 0x0f, 0xc2, 0x03, 0x08], "cmppd xmm8, [r11], 0x8");
+    test_instr(&[0xf2, 0x0f, 0xc2, 0xc3, 0x08], "cmpsd xmm0, xmm3, 0x8");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0xc2, 0xc3, 0x08], "cmpsd xmm8, xmm11, 0x8");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0xc2, 0x03, 0x08], "cmpsd xmm8, [r11], 0x8");
+
+    test_instr(&[0x66, 0x0f, 0xc4, 0xc3, 0x08], "pinsrw xmm0, ebx, 0x8");
+    test_instr(&[0x66, 0x4f, 0x0f, 0xc4, 0xc3, 0x08], "pinsrw xmm8, r11d, 0x8");
+
+    test_instr(&[0x66, 0x0f, 0xc4, 0x03, 0x08], "pinsrw xmm0, [rbx], 0x8");
+    test_instr(&[0x66, 0x4f, 0x0f, 0xc4, 0x03, 0x08], "pinsrw xmm8, [r11], 0x8");
+
+//    test_instr(&[0x66, 0x0f, 0xc5, 0xc3, 0x08], "pextrw eax, xmm3, 0x8");
+//    test_instr(&[0x66, 0x4f, 0x0f, 0xc5, 0xc3, 0x08], "pextrw r8d, xmm11, 0x8");
+//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x03, 0x08]);
+//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x40, 0x08]);
+//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x80, 0x08]);
+
+    test_instr(&[0x66, 0x4f, 0x0f, 0xc6, 0x03, 0x08], "shufpd xmm8, [r11], 0x8");
+    test_instr(&[0x66, 0x0f, 0xc6, 0x03, 0x08], "shufpd xmm0, [rbx], 0x8");
+    test_instr(&[0x66, 0x0f, 0xc6, 0xc3, 0x08], "shufpd xmm0, xmm3, 0x8");
+    test_instr(&[0x66, 0x0f, 0xd1, 0xc1], "psrlw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd1, 0x01], "psrlw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xd2, 0xc1], "psrld xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd2, 0x01], "psrld xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xd3, 0xc1], "psrlq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd3, 0x01], "psrlq xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xd4, 0xc1], "paddq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd4, 0x01], "paddq xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xd5, 0xc1], "pmullw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd5, 0x01], "pmullw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xd6, 0xc1], "movq xmm1, xmm0");
+    test_instr(&[0x66, 0x0f, 0xd6, 0x01], "movq [rcx], xmm0");
+    test_invalid(&[0xf3, 0x4f, 0x0f, 0xd6, 0x03]);
+    test_instr(&[0xf3, 0x4f, 0x0f, 0xd6, 0xc3], "movq2dq xmm8, mm3");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0xd6, 0xc3], "movdq2q mm0, xmm11");
+    test_instr(&[0x66, 0x0f, 0xd7, 0xc1], "pmovmskb eax, xmm1");
+    test_instr(&[0x66, 0x4f, 0x0f, 0xd7, 0xc1], "pmovmskb r8d, xmm9");
+    test_invalid(&[0x66, 0x0f, 0xd7, 0x01]);
+    test_instr(&[0x66, 0x0f, 0xd8, 0xc1], "psubusb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd8, 0x01], "psubusb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xd9, 0xc1], "psubusw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd9, 0x01], "psubusw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xda, 0xc1], "pminub xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xda, 0x01], "pminub xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xdb, 0xc1], "pand xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdb, 0x01], "pand xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xdc, 0xc1], "paddusb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdc, 0x01], "paddusb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xdd, 0xc1], "paddusw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdd, 0x01], "paddusw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xde, 0xc1], "pmaxub xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xde, 0x01], "pmaxub xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xdf, 0xc1], "pandn xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdf, 0x01], "pandn xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe0, 0xc1], "pavgb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe0, 0x01], "pavgb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe1, 0xc1], "psraw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe1, 0x01], "psraw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe2, 0xc1], "psrad xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe2, 0x01], "psrad xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe3, 0xc1], "pavgw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe3, 0x01], "pavgw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe4, 0xc1], "pmulhuw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe4, 0x01], "pmulhuw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe5, 0xc1], "pmulhw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe5, 0x01], "pmulhw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe6, 0xc1], "cvttpd2dq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe6, 0x01], "cvttpd2dq xmm0, [rcx]");
+    test_invalid(&[0x66, 0x0f, 0xe7, 0xc1]);
+    test_instr(&[0x66, 0x0f, 0xe7, 0x01], "movntdq [rcx], xmm0");
+    test_instr(&[0x66, 0x0f, 0xe8, 0xc1], "psubsb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe8, 0x01], "psubsb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xe9, 0xc1], "psubsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe9, 0x01], "psubsw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xea, 0xc1], "pminsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xea, 0x01], "pminsw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xc3], "por xmm0, xmm3");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xc4], "por xmm0, xmm4");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xd3], "por xmm2, xmm3");
+    test_instr(&[0x66, 0x0f, 0xeb, 0x12], "por xmm2, [rdx]");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xc1], "por xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xeb, 0x01], "por xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xec, 0xc1], "paddsb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xec, 0x01], "paddsb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xed, 0xc1], "paddsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xed, 0x01], "paddsw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xee, 0xc1], "pmaxsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xee, 0x01], "pmaxsw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xef, 0xc1], "pxor xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xef, 0x01], "pxor xmm0, [rcx]");
+    test_invalid(&[0x66, 0x0f, 0xf0, 0xc1]);
+    test_invalid(&[0x66, 0x0f, 0xf0, 0x01]);
+    test_instr(&[0x66, 0x0f, 0xf1, 0xc1], "psllw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf1, 0x01], "psllw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf2, 0xc1], "pslld xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf2, 0x01], "pslld xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf3, 0xc1], "psllq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf3, 0x01], "psllq xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf4, 0xc1], "pmuludq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf4, 0x01], "pmuludq xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf5, 0xc1], "pmaddwd xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf5, 0x01], "pmaddwd xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf6, 0xc1], "psadbw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf6, 0x01], "psadbw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf7, 0xc1], "maskmovdqu xmm0, xmm1");
+    test_invalid(&[0x66, 0x0f, 0xf7, 0x01]);
+    test_instr(&[0x66, 0x0f, 0xf8, 0xc1], "psubb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf8, 0x01], "psubb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xf9, 0xc1], "psubw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf9, 0x01], "psubw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xfa, 0xc1], "psubd xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfa, 0x01], "psubd xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xfb, 0xc1], "psubq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfb, 0x01], "psubq xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xfc, 0xc1], "paddb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfc, 0x01], "paddb xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xfd, 0xc1], "paddw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfd, 0x01], "paddw xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xfe, 0xc1], "paddd xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfe, 0x01], "paddd xmm0, [rcx]");
+    test_instr(&[0x66, 0x0f, 0xff, 0xc1], "paddq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xff, 0x01], "paddq xmm0, [rcx]");
+
+    test_instr(&[0x66, 0x0f, 0x74, 0xc1], "pcmpeqb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0x74, 0x12], "pcmpeqb xmm2, [rdx]");
+    test_instr(&[0x66, 0x0f, 0xf8, 0xc8], "psubb xmm1, xmm0");
+    test_instr(&[0x66, 0x0f, 0xf8, 0xd0], "psubb xmm2, xmm0");
+    test_instr(&[0x66, 0x0f, 0xf8, 0x12], "psubb xmm2, [rdx]");
+}
+
+#[test]
+fn test_sse3() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_sse3(), 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);
+    }
+
+    fn test_instr_invalid(bytes: &[u8]) {
+        test_invalid_under(&InstDecoder::minimal().with_sse3(), bytes);
+        test_invalid_under(&InstDecoder::default(), bytes);
+    }
+    test_instr(&[0xf2, 0x0f, 0xf0, 0x0f], "lddqu xmm1, [rdi]");
+    test_instr_invalid(&[0xf2, 0x0f, 0xf0, 0xcf]);
+    test_instr(&[0xf2, 0x0f, 0xd0, 0x0f], "addsubps xmm1, [rdi]");
+    test_instr(&[0xf2, 0x0f, 0xd0, 0xcf], "addsubps xmm1, xmm7");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0xd0, 0xcf], "addsubps xmm9, xmm15");
+    test_instr(&[0x66, 0x0f, 0xd0, 0x0f], "addsubpd xmm1, [rdi]");
+    test_instr(&[0x66, 0x0f, 0xd0, 0xcf], "addsubpd xmm1, xmm7");
+    test_instr(&[0x66, 0x4f, 0x0f, 0xd0, 0xcf], "addsubpd xmm9, xmm15");
+
+    test_instr(&[0xf2, 0x0f, 0x7c, 0x0f], "haddps xmm1, [rdi]");
+    test_instr(&[0xf2, 0x0f, 0x7c, 0xcf], "haddps xmm1, xmm7");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x7c, 0xcf], "haddps xmm9, xmm15");
+    test_instr(&[0x66, 0x0f, 0x7c, 0x0f], "haddpd xmm1, [rdi]");
+    test_instr(&[0x66, 0x0f, 0x7c, 0xcf], "haddpd xmm1, xmm7");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x7c, 0xcf], "haddpd xmm9, xmm15");
+
+    test_instr(&[0xf2, 0x0f, 0x7d, 0x0f], "hsubps xmm1, [rdi]");
+    test_instr(&[0xf2, 0x0f, 0x7d, 0xcf], "hsubps xmm1, xmm7");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x7d, 0xcf], "hsubps xmm9, xmm15");
+    test_instr(&[0x66, 0x0f, 0x7d, 0x0f], "hsubpd xmm1, [rdi]");
+    test_instr(&[0x66, 0x0f, 0x7d, 0xcf], "hsubpd xmm1, xmm7");
+    test_instr(&[0x66, 0x4f, 0x0f, 0x7d, 0xcf], "hsubpd xmm9, xmm15");
+
+    test_instr(&[0xf3, 0x0f, 0x12, 0x0f], "movsldup xmm1, [rdi]");
+    test_instr(&[0xf3, 0x0f, 0x12, 0xcf], "movsldup xmm1, xmm7");
+    test_instr(&[0xf3, 0x4f, 0x0f, 0x12, 0xcf], "movsldup xmm9, xmm15");
+    test_instr(&[0xf3, 0x0f, 0x16, 0x0f], "movshdup xmm1, [rdi]");
+    test_instr(&[0xf3, 0x0f, 0x16, 0xcf], "movshdup xmm1, xmm7");
+    test_instr(&[0xf3, 0x4f, 0x0f, 0x16, 0xcf], "movshdup xmm9, xmm15");
+
+    test_instr(&[0xf2, 0x0f, 0x12, 0x0f], "movddup xmm1, [rdi]");
+    test_instr(&[0xf2, 0x0f, 0x12, 0xcf], "movddup xmm1, xmm7");
+    test_instr(&[0xf2, 0x4f, 0x0f, 0x12, 0xcf], "movddup xmm9, xmm15");
+
+    test_instr(&[0x66, 0x0f, 0x01, 0xc8], "monitor");
+    test_instr(&[0xf2, 0x0f, 0x01, 0xc8], "monitor");
+    test_instr(&[0xf3, 0x0f, 0x01, 0xc8], "monitor");
+
+    test_instr(&[0x66, 0x0f, 0x01, 0xc9], "mwait");
+    test_instr(&[0xf2, 0x0f, 0x01, 0xc9], "mwait");
+    test_instr(&[0xf3, 0x0f, 0x01, 0xc9], "mwait");
+}
+
+#[test]
+fn test_0f01() {
+    // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
+    // Number"
+    test_display(&[0x0f, 0x01, 0x38], "invlpg [rax]");
+    test_display(&[0x0f, 0x01, 0x3f], "invlpg [rdi]");
+    test_display(&[0x0f, 0x01, 0x40, 0xff], "sgdt [rax - 0x1]");
+    test_display(&[0x0f, 0x01, 0x41, 0xff], "sgdt [rcx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x49, 0xff], "sidt [rcx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x51, 0xff], "lgdt [rcx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x59, 0xff], "lidt [rcx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x61, 0xff], "smsw [rcx - 0x1]");
+    test_display(&[0x0f, 0x01, 0xc0], "enclv");
+    test_display(&[0x0f, 0x01, 0xc1], "vmcall");
+    test_display(&[0x0f, 0x01, 0xc2], "vmlaunch");
+    test_display(&[0x0f, 0x01, 0xc3], "vmresume");
+    test_display(&[0x0f, 0x01, 0xc4], "vmxoff");
+    test_invalid(&[0x0f, 0x01, 0xc5]);
+    test_invalid(&[0x0f, 0x01, 0xc6]);
+    test_invalid(&[0x0f, 0x01, 0xc7]);
+    test_display(&[0x0f, 0x01, 0xc8], "monitor");
+    test_display(&[0x0f, 0x01, 0xc9], "mwait");
+    test_display(&[0x0f, 0x01, 0xca], "clac");
+    test_display(&[0x0f, 0x01, 0xcb], "stac");
+    test_display(&[0x0f, 0x01, 0xcf], "encls");
+    test_display(&[0x0f, 0x01, 0xd0], "xgetbv");
+    test_display(&[0x0f, 0x01, 0xd1], "xsetbv");
+    test_invalid(&[0x0f, 0x01, 0xd2]);
+    test_invalid(&[0x0f, 0x01, 0xd3]);
+    test_display(&[0x0f, 0x01, 0xd4], "vmfunc");
+    test_display(&[0x0f, 0x01, 0xd5], "xend");
+    test_display(&[0x0f, 0x01, 0xd6], "xtest");
+    test_display(&[0x0f, 0x01, 0xd7], "enclu");
+    test_invalid(&[0x0f, 0x01, 0xd8]);
+    test_invalid(&[0x0f, 0x01, 0xd9]);
+    test_invalid(&[0x0f, 0x01, 0xda]);
+    test_invalid(&[0x0f, 0x01, 0xdb]);
+    test_invalid(&[0x0f, 0x01, 0xdc]);
+    test_invalid(&[0x0f, 0x01, 0xdd]);
+    test_invalid(&[0x0f, 0x01, 0xde]);
+    test_invalid(&[0x0f, 0x01, 0xdf]);
+    test_display(&[0x0f, 0x01, 0xee], "rdpkru");
+    test_display(&[0x0f, 0x01, 0xef], "wrpkru");
+    test_display(&[0x0f, 0x01, 0xf8], "swapgs");
+    test_display(&[0x0f, 0x01, 0xf9], "rdtscp");
+}
+
+#[test]
+fn test_0fae() {
+    let intel = InstDecoder::minimal().with_intel_quirks();
+    let amd = InstDecoder::minimal().with_amd_quirks();
+    let default = InstDecoder::default();
+    let minimal = InstDecoder::minimal();
+    // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
+    // Number"
+    test_display(&[0x0f, 0xae, 0x04, 0x4f], "fxsave [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x0c, 0x4f], "fxrstor [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x14, 0x4f], "ldmxcsr [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x1c, 0x4f], "stmxcsr [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x24, 0x4f], "xsave [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x2c, 0x4f], "xrstor [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x34, 0x4f], "xsaveopt [rdi + rcx * 2]");
+    test_display(&[0x0f, 0xae, 0x3c, 0x4f], "clflush [rdi + rcx * 2]");
+
+    for (modrm, text) in &[(0xe8u8, "lfence"), (0xf0u8, "mfence"), (0xf8u8, "sfence")] {
+        test_display_under(&intel, &[0x0f, 0xae, *modrm], text);
+        test_display_under(&amd, &[0x0f, 0xae, *modrm], text);
+        test_display_under(&default, &[0x0f, 0xae, *modrm], text);
+        test_display_under(&minimal, &[0x0f, 0xae, *modrm], text);
+        // it turns out intel and amd accept m != 0 for {l,m,s}fence:
+        // from intel:
+        // ```
+        // Specification of the instruction's opcode above indicates a ModR/M byte of F0. For this
+        // instruction, the processor ignores the r/m field of the ModR/M byte. Thus, MFENCE is encoded
+        // by any opcode of the form 0F AE Fx, where x is in the range 0-7.
+        // ```
+        // whereas amd does not discuss the r/m field at all. at least as of zen, amd also accepts
+        // these encodings.
+        for m in 1u8..8u8 {
+            test_display_under(&intel, &[0x0f, 0xae, modrm | m], text);
+            test_display_under(&amd, &[0x0f, 0xae, modrm | m], text);
+            test_display_under(&default, &[0x0f, 0xae, modrm | m], text);
+            test_invalid_under(&minimal, &[0x0f, 0xae, modrm | m]);
+        }
+    }
+}
+
+#[test]
+fn test_system() {
+    test_display(&[0x66, 0x4f, 0x0f, 0xb2, 0x00], "lss r8, [r8]");
+    test_display(&[0x67, 0x4f, 0x0f, 0xb2, 0x00], "lss r8, [r8d]");
+    test_display(&[0x4f, 0x0f, 0xb2, 0x00], "lss r8, [r8]");
+    test_display(&[0x45, 0x0f, 0x22, 0xc8], "mov cr9, r8");
+    test_display(&[0x45, 0x0f, 0x20, 0xc8], "mov r8, cr9");
+    test_display(&[0x40, 0x0f, 0x22, 0xc8], "mov cr1, rax");
+    test_display(&[0x0f, 0x22, 0xc8], "mov cr1, rax");
+    test_display(&[0x44, 0x0f, 0x22, 0xcf], "mov cr9, rdi");
+    test_display(&[0x0f, 0x22, 0xcf], "mov cr1, rdi");
+    test_display(&[0x0f, 0x20, 0xc8], "mov rax, cr1");
+
+    test_display(&[0x45, 0x0f, 0x23, 0xc8], "mov dr9, r8");
+    test_display(&[0x45, 0x0f, 0x21, 0xc8], "mov r8, dr9");
+    test_display(&[0x40, 0x0f, 0x23, 0xc8], "mov dr1, rax");
+    test_display(&[0x0f, 0x23, 0xc8], "mov dr1, rax");
+    test_display(&[0x0f, 0x21, 0xc8], "mov rax, dr1");
+}
+
+#[test]
+fn test_arithmetic() {
+    test_display(&[0x81, 0xec, 0x10, 0x03, 0x00, 0x00], "sub esp, 0x310");
+    test_display(&[0x0f, 0xaf, 0xc2], "imul eax, edx");
+    test_display(&[0x4b, 0x69, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65], "imul rax, [r11 + 0x6f], 0x656c706d");
+    test_display(&[0x66, 0x0f, 0xaf, 0xd1], "imul dx, cx");
+    test_display(&[0x4b, 0x6b, 0x43, 0x6f, 0x6d], "imul al, [r11 + 0x6f], 0x6d");
+    test_display(&[0x4f, 0x4e, 0x00, 0xcc], "add spl, r9b");
+}
+
+#[test]
+#[allow(non_snake_case)]
+fn test_E_decode() {
+    test_display(&[0xff, 0x75, 0xb8], "push [rbp - 0x48]");
+    test_display(&[0xff, 0x75, 0x08], "push [rbp + 0x8]");
+}
+
+#[test]
+fn test_sse() {
+    test_display(&[0xf3, 0x0f, 0x10, 0x0c, 0xc7], "movss xmm1, [rdi + rax * 8]");
+    test_display(&[0xf3, 0x0f, 0x11, 0x0c, 0xc7], "movss [rdi + rax * 8], xmm1");
+    test_display(&[0x4f, 0x0f, 0x28, 0x00], "movaps xmm8, [r8]");
+    test_display(&[0x4f, 0x0f, 0x29, 0x00], "movaps [r8], xmm8");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0xc1], "cvtsi2ss xmm8, r9");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0x01], "cvtsi2ss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x2b, 0x00], "movntps [r8], xmm8");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2c, 0xc1], "cvttss2si r8, xmm9");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2c, 0x01], "cvttss2si r8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2d, 0xc1], "cvtss2si r8, xmm9");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x2d, 0x01], "cvtss2si r8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x2e, 0x00], "ucomiss xmm8, [r8]");
+    test_display(&[0x4f, 0x0f, 0x2f, 0x00], "comiss xmm8, [r8]");
+    test_display(&[0x0f, 0x28, 0xd0], "movaps xmm2, xmm0");
+    test_display(&[0x66, 0x0f, 0x28, 0xd0], "movapd xmm2, xmm0");
+    test_display(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [rax]");
+    test_display(&[0x4f, 0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [rax]");
+    test_display(&[0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8]");
+    test_display(&[0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8]");
+    test_display(&[0x67, 0x4f, 0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
+    test_display(&[0x67, 0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8d]");
+    test_display(&[0x66, 0x0f, 0x29, 0x00], "movapd [rax], xmm0");
+    test_invalid(&[0x4f, 0x0f, 0x50, 0x00]);
+    test_display(&[0x4f, 0x0f, 0x50, 0xc1], "movmskps r8d, xmm9");
+    test_display(&[0x4f, 0x0f, 0x51, 0x01], "sqrtps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x51, 0x01], "sqrtss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x52, 0x01], "rsqrtps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x52, 0x01], "rsqrtss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x53, 0x01], "rcpps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x53, 0x01], "rcpss xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x53, 0xc1], "rcpss xmm8, xmm9");
+    test_display(&[0x4f, 0x0f, 0x54, 0x01], "andps xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x55, 0x01], "andnps xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x56, 0x01], "orps xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x57, 0x01], "xorps xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x58, 0x01], "addps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x58, 0x01], "addss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x59, 0x01], "mulps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x59, 0x01], "mulss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x5a, 0x01], "cvtps2pd xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x5a, 0x01], "cvtss2sd xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x5b, 0x01], "cvtdq2ps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x5b, 0x01], "cvttps2dq xmm8, [r9]");
+    test_display(&[0x67, 0x4f, 0x0f, 0x5b, 0x01], "cvtdq2ps xmm8, [r9d]");
+    test_display(&[0x4f, 0x0f, 0x5c, 0x01], "subps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x5c, 0x01], "subss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x5d, 0x01], "minps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x5d, 0x01], "minss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x5e, 0x01], "divps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x5e, 0x01], "divss xmm8, [r9]");
+    test_display(&[0x4f, 0x0f, 0x5f, 0x01], "maxps xmm8, [r9]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x5f, 0x01], "maxss xmm8, [r9]");
+
+    test_display(&[0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [rax], 0x7f");
+    test_display(&[0x4f, 0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [r8], 0x7f");
+
+    test_display(&[0x66, 0x0f, 0xef, 0xc0], "pxor xmm0, xmm0");
+    test_display(&[0x66, 0x4f, 0x0f, 0xef, 0xc0], "pxor xmm8, xmm8");
+    test_display(&[0xf2, 0x0f, 0x10, 0x0c, 0xc6], "movsd xmm1, [rsi + rax * 8]");
+    test_display(&[0xf3, 0x0f, 0x10, 0x04, 0x86], "movss xmm0, [rsi + rax * 4]");
+    test_display(&[0xf2, 0x0f, 0x59, 0xc8], "mulsd xmm1, xmm0");
+    test_display(&[0xf3, 0x0f, 0x59, 0xc8], "mulss xmm1, xmm0");
+    test_display(&[0xf2, 0x4f, 0x0f, 0x59, 0xc8], "mulsd xmm9, xmm8");
+
+    test_display(
+        &[0xf3, 0x4f, 0x0f, 0x6f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqu xmm11, [r12 + r11 * 4 - 0x334455cc]"
+    );
+    test_display(&[0xf3, 0x0f, 0x70, 0xc0, 0x4e], "pshufhw xmm0, xmm0, 0x4e");
+    test_display(&[0xf3, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
+    test_display(&[0xf3, 0x4f, 0x0f, 0x7e, 0xc1], "movd r9, mm0"); // use of rex.w demotes to movd r/mm
+    test_display(&[0xf3, 0x40, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
+    test_display(&[0xf3, 0x41, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm9");
+    test_display(&[0xf3, 0x42, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
+    test_display(&[0xf3, 0x44, 0x0f, 0x7e, 0xc1], "movq xmm8, xmm1");
+    test_display(&[0xf3, 0x48, 0x0f, 0x7e, 0xc1], "movd rcx, mm0");
+    test_display(
+        &[0xf3, 0x4f, 0x0f, 0x7f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqu [r12 + r11 * 4 - 0x334455cc], xmm11"
+    );
+
+    test_display(&[0xf3, 0x0f, 0xc2, 0xc3, 0x08], "cmpss xmm0, xmm3, 0x8");
+    test_display(&[0xf3, 0x4f, 0x0f, 0xc2, 0xc3, 0x08], "cmpss xmm8, xmm11, 0x8");
+    test_display(&[0xf3, 0x4f, 0x0f, 0xc2, 0x03, 0x08], "cmpss xmm8, [r11], 0x8");
+}
+
+// SETLE, SETNG, ...
+
+#[test]
+fn test_mov() {
+    // test_display(&[0xa1, 0x93, 0x62, 0xc4, 0x00, 0x12, 0x34, 0x12, 0x34], "mov eax, [0x3412341200c46293]");
+    // RCT.exe 32bit version, TODO: FIX
+    test_display(&[0xa1, 0x93, 0x62, 0xc4, 0x00], "mov eax, [0xc46293]");
+    test_display(&[0xba, 0x01, 0x00, 0x00, 0x00], "mov edx, 0x1");
+    test_display(&[0x48, 0xc7, 0x04, 0x24, 0x00, 0x00, 0x00, 0x00], "mov [rsp], 0x0");
+    test_display(&[0x48, 0x89, 0x44, 0x24, 0x08], "mov [rsp + 0x8], rax");
+    test_display(&[0x48, 0x89, 0x43, 0x18], "mov [rbx + 0x18], rax");
+    test_display(&[0x48, 0xc7, 0x43, 0x10, 0x00, 0x00, 0x00, 0x00], "mov [rbx + 0x10], 0x0");
+    test_display(&[0x49, 0x89, 0x4e, 0x08], "mov [r14 + 0x8], rcx");
+    test_display(&[0x48, 0x8b, 0x32], "mov rsi, [rdx]");
+    test_display(&[0x4d, 0x8b, 0x4c, 0x10, 0xf8], "mov r9, [r8 + rdx * 1 - 0x8]");
+    test_display(&[0x49, 0x89, 0x46, 0x10], "mov [r14 + 0x10], rax");
+    test_display(&[0x4d, 0x0f, 0x43, 0xec], "cmovnb r13, r12");
+    test_display(&[0x0f, 0xb6, 0x06], "movzx eax, byte [rsi]");
+    test_display(&[0x0f, 0xb7, 0x06], "movzx eax, word [rsi]");
+    test_display(&[0x89, 0x55, 0x94], "mov [rbp - 0x6c], edx");
+    test_display(&[0x65, 0x4c, 0x89, 0x04, 0x25, 0xa8, 0x01, 0x00, 0x00], "mov gs:[0x1a8], r8");
+    test_display(&[0x0f, 0xbe, 0x83, 0xb4, 0x00, 0x00, 0x00], "movsx eax, byte [rbx + 0xb4]");
+    test_display(&[0x46, 0x63, 0xc1], "movsxd r8, ecx");
+    test_display(&[0x48, 0x63, 0x04, 0xba], "movsxd rax, [rdx + rdi * 4]");
+    test_display(&[0xf3, 0x0f, 0x6f, 0x07], "movdqu xmm0, [rdi]");
+    test_display(&[0xf3, 0x0f, 0x7f, 0x45, 0x00], "movdqu [rbp], xmm0");
+}
+
+#[test]
+fn test_stack() {
+    test_display(&[0x66, 0x41, 0x50], "push r8w");
+}
+
+#[test]
+fn test_prefixes() {
+    test_display(&[0x66, 0x41, 0x31, 0xc0], "xor r8w, ax");
+    test_display(&[0x66, 0x41, 0x32, 0xc0], "xor al, r8b");
+    test_display(&[0x40, 0x32, 0xc5], "xor al, bpl");
+
+    // test that WAIT doesn't blow up, at least...
+    assert_eq!(InstDecoder::default().decode([0x9b, 0xf8].iter().cloned()).err(), Some(DecodeError::IncompleteDecoder));
+}
+
+#[test]
+fn test_control_flow() {
+    test_display(&[0x73, 0x31], "jnb 0x31");
+    test_display(&[0x72, 0x5a], "jb 0x5a");
+    test_display(&[0x0f, 0x86, 0x8b, 0x01, 0x00, 0x00], "jna 0x18b");
+    test_display(&[0x74, 0x47], "jz 0x47");
+    test_display(&[0xff, 0x15, 0x7e, 0x72, 0x24, 0x00], "call [rip + 0x24727e]");
+    test_display(&[0xff, 0x24, 0xcd, 0x70, 0xa0, 0xbc, 0x01], "jmp [rcx * 8 + 0x1bca070]");
+    test_display(&[0xff, 0xe0], "jmp rax");
+    test_display(&[0x66, 0xff, 0xe0], "jmp rax");
+    test_display(&[0x67, 0xff, 0xe0], "jmp rax");
+    test_invalid(&[0xff, 0xd8]);
+    test_display(&[0xff, 0x18], "callf [rax]");
+    test_display(&[0xc3], "ret");
+}
+
+#[test]
+fn test_test_cmp() {
+    test_display(&[0xf6, 0x05, 0x2c, 0x9b, 0xff, 0xff, 0x01], "test [rip - 0x64d4], 0x1");
+    test_display(&[0x48, 0x3d, 0x01, 0xf0, 0xff, 0xff], "cmp rax, -0xfff");
+    test_display(&[0x3d, 0x01, 0xf0, 0xff, 0xff], "cmp eax, -0xfff");
+    test_display(&[0x48, 0x83, 0xf8, 0xff], "cmp rax, -0x1");
+    test_display(&[0x48, 0x39, 0xc6], "cmp rsi, rax");
+}
+
+#[test]
+fn test_push_pop() {
+    test_display(&[0x5b], "pop rbx");
+    test_display(&[0x41, 0x5e], "pop r14");
+    test_display(&[0x68, 0x7f, 0x63, 0xc4, 0x00], "push 0xc4637f");
+}
+
+#[test]
+fn test_bmi1() {
+    let bmi1 = InstDecoder::minimal().with_bmi1();
+    let no_bmi1 = InstDecoder::minimal();
+    test_display_under(&bmi1, &[0x41, 0x0f, 0xbc, 0xd3], "tzcnt edx, r11d");
+    test_display_under(&no_bmi1, &[0x41, 0x0f, 0xbc, 0xd3], "bsf edx, r11d");
+
+    test_display_under(&bmi1, &[0xf3, 0x0f, 0xb8, 0xc1], "popcnt eax, ecx");
+    test_display_under(&bmi1, &[0xf3, 0x4f, 0x0f, 0xb8, 0xc1], "popcnt r8, r9");
+}
+
+#[test]
+fn test_bitwise() {
+    test_display_under(&InstDecoder::minimal(), &[0x41, 0x0f, 0xbc, 0xd3], "bsf edx, r11d");
+    test_display(&[0x48, 0x0f, 0xa3, 0xd0], "bt rax, rdx");
+    test_display(&[0x48, 0x0f, 0xab, 0xd0], "bts rax, rdx");
+}
+
+#[test]
+fn test_misc() {
+    // TODO
+//    test_display(&[0xf2, 0x0f, 0x38, 0xf0, 0xc1], "crc32 eax, cl");
+//    test_display(&[0xf2, 0x0f, 0x38, 0xf1, 0xc1], "crc32 eax, ecx");
+    test_display(&[0xfe, 0x00], "inc [rax]"); // TODO: inc byte [rax]
+    test_display(&[0xfe, 0x08], "dec [rax]"); // TODO: dec byte [rax]
+    test_display(&[0xff, 0x00], "inc [rax]"); // TODO: inc dword [rax]
+    test_display(&[0x48, 0xff, 0x00], "inc [rax]"); // TODO: inc qword [rax]
+    test_display(&[0xe4, 0x99], "in al, 0x99");
+    test_display(&[0xe5, 0x99], "in eax, 0x99");
+    test_display(&[0x67, 0xe5, 0x99], "in eax, 0x99");
+    test_display(&[0x4f, 0xe5, 0x99], "in eax, 0x99");
+    test_display(&[0xe6, 0x99], "out 0x99, al");
+    test_display(&[0x4f, 0xe7, 0x99], "out 0x99, eax");
+    test_display(&[0xec], "in al, dx");
+    test_display(&[0xed], "in eax, dx");
+    test_display(&[0xee], "out dx, al");
+    test_display(&[0xef], "out dx, eax");
+    test_display(&[0xcd, 0x00], "int 0x0");
+    test_display(&[0xcd, 0xff], "int 0xff");
+    test_display(&[0x9c], "pushf");
+    test_display(&[0x48, 0x98], "cdqe");
+    test_display(&[0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00], "nop cs:[rax + rax * 1]");
+    test_display(&[0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00], "nop [rax + rax * 1]");
+    test_display(&[0x48, 0x8d, 0xa4, 0xc7, 0x20, 0x00, 0x00, 0x12], "lea rsp, [rdi + rax * 8 + 0x12000020]");
+    test_display(&[0x33, 0xc0], "xor eax, eax");
+    test_display(&[0x48, 0x8d, 0x53, 0x08], "lea rdx, [rbx + 0x8]");
+    test_display(&[0x31, 0xc9], "xor ecx, ecx");
+    test_display(&[0x48, 0x29, 0xc8], "sub rax, rcx");
+    test_display(&[0x48, 0x03, 0x0b], "add rcx, [rbx]");
+    test_display(&[0x48, 0x8d, 0x0c, 0x12], "lea rcx, [rdx + rdx * 1]");
+    test_display(&[0xf6, 0xc2, 0x18], "test dl, 0x18");
+    test_display(&[0xf3, 0x48, 0xab], "rep stos es:[rdi], rax");
+    test_display(&[0xf3, 0x48, 0xa5], "rep movs es:[rdi], ds:[rsi]");
+    test_display(&[0xf3, 0x45, 0x0f, 0xbc, 0xd7], "tzcnt r10d, r15d");
+
+    // this is actually vmx
+    // test_invalid(&[0x66, 0x0f, 0xc7, 0x03]);
+    test_display(&[0x66, 0x4f, 0x0f, 0xc7, 0x33], "vmclear [r11]");
+    test_display(&[0x66, 0x0f, 0xc7, 0x33], "vmclear [rbx]");
+    test_display(&[0xf3, 0x4f, 0x0f, 0xc7, 0x33], "vmxon [r11]");
+    test_display(&[0xf3, 0x0f, 0xc7, 0x33], "vmxon [rbx]");
+}
+
+#[test]
+#[ignore]
+// TODO also not supported at all
+fn evex() {
+    test_display(&[0x62, 0xf2, 0x7d, 0x48, 0x2a, 0x44, 0x40, 0x01], "vmovntdqa zmm0, [rax + rax*2 + 0x40]");
+    test_display(&[0x62, 0xf2, 0x7d, 0x08, 0x2a, 0x44, 0x40, 0x01], "vmovntdqa xmm0, [rax + rax*2 + 0x10]");
+}
+
+#[test]
+fn test_vex() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_avx(), bytes, text);
+        test_display_under(&InstDecoder::default(), bytes, text);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+    }
+
+    fn test_instr_invalid(bytes: &[u8]) {
+        test_invalid_under(&InstDecoder::minimal().with_avx(), bytes);
+        test_invalid_under(&InstDecoder::default(), bytes);
+    }
+
+    test_instr(&[0xc5, 0xf8, 0x10, 0x00], "vmovups xmm0, [rax]");
+    test_instr(&[0xc5, 0xf8, 0x10, 0x01], "vmovups xmm0, [rcx]");
+    test_instr(&[0xc5, 0x78, 0x10, 0x0f], "vmovups xmm9, [rdi]");
+    test_instr(&[0xc5, 0xf8, 0x10, 0xcf], "vmovups xmm1, xmm7");
+    test_instr(&[0xc5, 0xf9, 0x10, 0x0f], "vmovupd xmm1, [rdi]");
+    test_instr(&[0xc5, 0xfa, 0x7e, 0x10], "vmovq xmm2, [rax]");
+    test_instr(&[0xc5, 0xfc, 0x10, 0x0f], "vmovups ymm1, [rdi]");
+    test_instr(&[0xc5, 0xfd, 0x10, 0x0f], "vmovupd ymm1, [rdi]");
+    test_instr(&[0xc5, 0xfe, 0x10, 0x0f], "vmovss xmm1, [rdi]");
+    test_instr(&[0xc5, 0xff, 0x10, 0xcf], "vmovsd xmm1, xmm0, xmm7");
+    test_instr(&[0xc5, 0xff, 0x10, 0x01], "vmovsd xmm0, [rcx]");
+    test_instr(&[0xc5, 0xf9, 0x6e, 0xc6], "vmovd xmm0, esi");
+    test_instr(&[0xc5, 0xf9, 0x6e, 0x13], "vmovd xmm2, [rbx]");
+    test_instr(&[0xc5, 0xf9, 0x7e, 0xc6], "vmovd esi, xmm0");
+    test_instr(&[0xc5, 0xf9, 0x7e, 0x13], "vmovd [rbx], xmm2");
+    test_instr_invalid(&[0x4f, 0xc5, 0xf8, 0x10, 0x00]);
+    test_instr_invalid(&[0xf0, 0xc5, 0xf8, 0x10, 0x00]);
+    test_instr(&[0xc4, 0x02, 0x71, 0x00, 0x0f], "vpshufb xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x00, 0x0f], "vpshufb ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x00, 0xcd], "vpshufb xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x00, 0xcd], "vpshufb ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x01, 0x0f], "vphaddw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x01, 0x0f], "vphaddw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x01, 0xcd], "vphaddw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x01, 0xcd], "vphaddw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x02, 0x0f], "vphaddd xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x02, 0x0f], "vphaddd ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x02, 0xcd], "vphaddd xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x02, 0xcd], "vphaddd ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x03, 0x0f], "vphaddsw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x03, 0x0f], "vphaddsw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x03, 0xcd], "vphaddsw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x03, 0xcd], "vphaddsw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x04, 0x0f], "vphaddubsw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x04, 0x0f], "vphaddubsw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x04, 0xcd], "vphaddubsw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x04, 0xcd], "vphaddubsw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x05, 0x0f], "vphsubw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x05, 0x0f], "vphsubw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x05, 0xcd], "vphsubw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x05, 0xcd], "vphsubw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x06, 0x0f], "vphsubd xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x06, 0x0f], "vphsubd ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x06, 0xcd], "vphsubd xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x06, 0xcd], "vphsubd ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x07, 0x0f], "vphsubsw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x07, 0x0f], "vphsubsw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x07, 0xcd], "vphsubsw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x07, 0xcd], "vphsubsw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x08, 0x0f], "vpsignb xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x08, 0x0f], "vpsignb ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x08, 0xcd], "vpsignb xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x08, 0xcd], "vpsignb ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x09, 0x0f], "vpsignw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x09, 0x0f], "vpsignw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x09, 0xcd], "vpsignw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x09, 0xcd], "vpsignw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0a, 0x0f], "vpsignd xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0a, 0x0f], "vpsignd ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0a, 0xcd], "vpsignd xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0a, 0xcd], "vpsignd ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0b, 0x0f], "vpmulhrsw xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0b, 0x0f], "vpmulhrsw ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0b, 0xcd], "vpmulhrsw xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0b, 0xcd], "vpmulhrsw ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0c, 0x0f], "vpermilps xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0c, 0x0f], "vpermilps ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0c, 0xcd], "vpermilps xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0c, 0xcd], "vpermilps ymm9, ymm1, ymm13");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0d, 0x0f], "vpermilpd xmm9, xmm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0d, 0x0f], "vpermilpd ymm9, ymm1, [r15]");
+    test_instr(&[0xc4, 0x02, 0x71, 0x0d, 0xcd], "vpermilpd xmm9, xmm1, xmm13");
+    test_instr(&[0xc4, 0x02, 0x75, 0x0d, 0xcd], "vpermilpd ymm9, ymm1, ymm13");
+    test_instr_invalid(&[0xc4, 0x02, 0x71, 0x0e, 0x00]);
+    test_instr(&[0xc4, 0x02, 0x79, 0x0e, 0x0f], "vtestps xmm9, [r15]");
+    test_instr(&[0xc4, 0x02, 0x7d, 0x0e, 0x0f], "vtestps ymm9, [r15]");
+    test_instr(&[0xc4, 0x02, 0x79, 0x0e, 0xcd], "vtestps xmm9, xmm13");
+    test_instr(&[0xc4, 0x02, 0x7d, 0x0e, 0xcd], "vtestps ymm9, ymm13");
+    test_instr_invalid(&[0xc4, 0x02, 0x71, 0x0f, 0x00]);
+    test_instr(&[0xc4, 0x02, 0x79, 0x0f, 0x0f], "vtestpd xmm9, [r15]");
+    test_instr(&[0xc4, 0x02, 0x7d, 0x0f, 0x0f], "vtestpd ymm9, [r15]");
+    test_instr(&[0xc4, 0x02, 0x79, 0x0f, 0xcd], "vtestpd xmm9, xmm13");
+    test_instr(&[0xc4, 0x02, 0x7d, 0x0f, 0xcd], "vtestpd ymm9, ymm13");
+    test_instr(&[0xc4, 0xe2, 0x65, 0x90, 0x04, 0x51], "vpgatherdd ymm0, [rcx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xe2, 0xe5, 0x90, 0x04, 0x51], "vpgatherdq ymm0, [rcx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xe2, 0x65, 0x91, 0x04, 0x51], "vpgatherqd ymm0, [rcx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xe2, 0xe5, 0x91, 0x04, 0x51], "vpgatherqq ymm0, [rcx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0x02, 0x09, 0x9d, 0xcd], "vfnmadd132ss xmm9, xmm14, xmm13");
+    test_instr(&[0xc4, 0x02, 0x89, 0x9d, 0xcd], "vfnmadd132sd xmm9, xmm14, xmm13");
+// ...
+    test_instr(&[0xc4, 0xe3, 0x79, 0x14, 0xd0, 0x0a], "vpextrb rax, xmm2, 0xa");
+    test_instr(&[0xc4, 0xe3, 0x79, 0x14, 0x10, 0x0a], "vpextrb [rax], xmm2, 0xa");
+    test_instr_invalid(&[0xc4, 0xe3, 0xf9, 0x14, 0x00, 0xd0]);
+    test_instr_invalid(&[0xc4, 0xe3, 0xf9, 0x14, 0x00, 0x0a]);
+}
+
+#[test]
+fn strange_prefixing() {
+    test_display(&[0x45, 0x66, 0x0f, 0x21, 0xc8], "mov rax, dr1");
+    test_display(&[0x45, 0xf2, 0x0f, 0x21, 0xc8], "mov rax, dr1");
+    test_display(&[0x45, 0xf3, 0x0f, 0x21, 0xc8], "mov rax, dr1");
+}
+
+#[test]
+fn prefixed_0f() {
+    test_display(&[0x0f, 0x02, 0xc0], "lar eax, ax");
+    test_display(&[0x48, 0x0f, 0x02, 0xc0], "lar rax, ax");
+    test_display(&[0x0f, 0x03, 0xc0], "lsl eax, eax");
+    test_display(&[0x48, 0x0f, 0x03, 0xc0], "lsl rax, rax");
+    test_display(&[0x0f, 0x05], "syscall");
+    test_display(&[0x48, 0x0f, 0x05], "syscall");
+    test_display(&[0x66, 0x0f, 0x05], "syscall");
+    test_display(&[0x0f, 0x06], "clts");
+    test_display(&[0xf2, 0x0f, 0x06], "clts");
+    test_display(&[0x0f, 0x07], "sysret");
+    test_display(&[0xf2, 0x0f, 0x07], "sysret");
+    test_display(&[0x0f, 0x12, 0x0f], "movlps xmm1, [rdi]");
+    test_display(&[0x0f, 0x12, 0xcf], "movhlps xmm1, xmm7");
+    test_display(&[0x0f, 0x16, 0x0f], "movhps xmm1, [rdi]");
+    test_display(&[0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
+    test_display(&[0x0f, 0x12, 0xc0], "movhlps xmm0, xmm0");
+    test_invalid(&[0x0f, 0x13, 0xc0]);
+    test_display(&[0x0f, 0x13, 0x00], "movlps [rax], xmm0");
+    test_display(&[0x0f, 0x14, 0x08], "unpcklps xmm1, [rax]");
+    test_display(&[0x0f, 0x15, 0x08], "unpckhps xmm1, [rax]");
+    test_display(&[0x0f, 0x16, 0x0f], "movhps xmm1, [rdi]");
+    test_display(&[0x0f, 0x16, 0xc0], "movlhps xmm0, xmm0");
+    test_invalid(&[0x0f, 0x17, 0xc0]);
+    test_display(&[0x0f, 0x17, 0x00], "movhps [rax], xmm0");
+    test_invalid(&[0x0f, 0x18, 0xc0]);
+    test_display(&[0x0f, 0x18, 0x00], "prefetchnta [rax]");
+    test_display(&[0x0f, 0x18, 0x08], "prefetch0 [rax]");
+    test_display(&[0x0f, 0x18, 0x10], "prefetch1 [rax]");
+    test_display(&[0x0f, 0x18, 0x18], "prefetch2 [rax]");
+    test_display(&[0x0f, 0x18, 0x20], "nop [rax]");
+    test_display(&[0x4f, 0x0f, 0x18, 0x20], "nop [r8]");
+    test_display(&[0x0f, 0x19, 0x20], "nop [rax]");
+    test_display(&[0x0f, 0x1a, 0x20], "nop [rax]");
+    test_display(&[0x0f, 0x1b, 0x20], "nop [rax]");
+    test_display(&[0x0f, 0x1c, 0x20], "nop [rax]");
+    test_display(&[0x0f, 0x1d, 0x20], "nop [rax]");
+    test_display(&[0x0f, 0x1e, 0x20], "nop [rax]");
+    test_display(&[0x0f, 0x1f, 0x20], "nop [rax]");
+    test_display(&[0x45, 0x0f, 0x20, 0xc8], "mov r8, cr9");
+    test_display(&[0x0f, 0x20, 0xc8], "mov rax, cr1");
+    test_display(&[0x45, 0x0f, 0x21, 0xc8], "mov r8, dr9");
+    test_display(&[0x0f, 0x21, 0xc8], "mov rax, dr1");
+    test_display(&[0x45, 0x0f, 0x22, 0xc8], "mov cr9, r8");
+    test_display(&[0x40, 0x0f, 0x22, 0xc8], "mov cr1, rax");
+    test_display(&[0x0f, 0x22, 0xc8], "mov cr1, rax");
+    test_display(&[0x44, 0x0f, 0x22, 0xcf], "mov cr9, rdi");
+    test_display(&[0x0f, 0x22, 0xcf], "mov cr1, rdi");
+    test_display(&[0x45, 0x0f, 0x23, 0xc8], "mov dr9, r8");
+    test_display(&[0x40, 0x0f, 0x23, 0xc8], "mov dr1, rax");
+    test_display(&[0x0f, 0x23, 0xc8], "mov dr1, rax");
+    test_display(&[0x44, 0x0f, 0x23, 0xcf], "mov dr9, rdi");
+    test_display(&[0x0f, 0x23, 0xcf], "mov dr1, rdi");
+    test_display(&[0x0f, 0x30], "wrmsr");
+    test_display(&[0x0f, 0x31], "rdtsc");
+    test_display(&[0x0f, 0x32], "rdmsr");
+    test_display(&[0x0f, 0x33], "rdpmc");
+    test_display(&[0x0f, 0x34], "sysenter");
+    test_display(&[0x0f, 0x35], "sysexit");
+    test_invalid(&[0x0f, 0x36]);
+    test_display(&[0x0f, 0x37], "getsec");
+    test_display(&[0x0f, 0x60, 0x00], "punpcklbw mm0, [rax]");
+    test_display(&[0x0f, 0x60, 0xc2], "punpcklbw mm0, mm2");
+    test_display(&[0x0f, 0x61, 0x00], "punpcklwd mm0, [rax]");
+    test_display(&[0x0f, 0x61, 0xc2], "punpcklwd mm0, mm2");
+    test_display(&[0x0f, 0x62, 0x00], "punpckldq mm0, [rax]");
+    test_display(&[0x0f, 0x62, 0xc2], "punpckldq mm0, mm2");
+    test_display(&[0x0f, 0x63, 0x00], "packsswb mm0, [rax]");
+    test_display(&[0x0f, 0x63, 0xc2], "packsswb mm0, mm2");
+    test_display(&[0x0f, 0x64, 0x00], "pcmpgtb mm0, [rax]");
+    test_display(&[0x0f, 0x64, 0xc2], "pcmpgtb mm0, mm2");
+    test_display(&[0x0f, 0x65, 0x00], "pcmpgtw mm0, [rax]");
+    test_display(&[0x0f, 0x65, 0xc2], "pcmpgtw mm0, mm2");
+    test_display(&[0x0f, 0x66, 0x00], "pcmpgtd mm0, [rax]");
+    test_display(&[0x0f, 0x66, 0xc2], "pcmpgtd mm0, mm2");
+    test_display(&[0x0f, 0x67, 0x00], "packuswb mm0, [rax]");
+    test_display(&[0x0f, 0x67, 0xc2], "packuswb mm0, mm2");
+    test_display(&[0x0f, 0x68, 0x00], "punpckhbw mm0, [rax]");
+    test_display(&[0x0f, 0x68, 0xc2], "punpckhbw mm0, mm2");
+    test_display(&[0x0f, 0x69, 0x00], "punpckhwd mm0, [rax]");
+    test_display(&[0x0f, 0x69, 0xc2], "punpckhwd mm0, mm2");
+    test_display(&[0x0f, 0x6a, 0x00], "punpckhdq mm0, [rax]");
+    test_display(&[0x0f, 0x6a, 0xc2], "punpckhdq mm0, mm2");
+    test_display(&[0x0f, 0x6b, 0x00], "packssdw mm0, [rax]");
+    test_display(&[0x0f, 0x6b, 0xc2], "packssdw mm0, mm2");
+    test_invalid(&[0x0f, 0x6c]);
+    test_invalid(&[0x0f, 0x6d]);
+    test_display(&[0x0f, 0x6e, 0x00], "movd mm0, [rax]");
+    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, 0x70, 0x00, 0x7f], "pshufw mm0, [rax], 0x7f");
+    test_display(&[0x4f, 0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [r8], 0x7f");
+    test_invalid(&[0x0f, 0x71, 0x00, 0x7f]);
+    test_invalid(&[0x0f, 0x71, 0xc0, 0x7f]);
+    test_display(&[0x0f, 0x71, 0xd0, 0x7f], "psrlw mm0, 0x7f");
+    test_display(&[0x0f, 0x71, 0xe0, 0x7f], "psraw mm0, 0x7f");
+    test_display(&[0x0f, 0x71, 0xf0, 0x7f], "psllw mm0, 0x7f");
+    test_invalid(&[0x0f, 0x72, 0x00, 0x7f]);
+    test_invalid(&[0x0f, 0x72, 0xc0, 0x7f]);
+    test_display(&[0x0f, 0x72, 0xd0, 0x7f], "psrld mm0, 0x7f");
+    test_display(&[0x0f, 0x72, 0xe0, 0x7f], "psrad mm0, 0x7f");
+    test_display(&[0x0f, 0x72, 0xf0, 0x7f], "pslld mm0, 0x7f");
+    test_invalid(&[0x0f, 0x73, 0x00, 0x7f]);
+    test_invalid(&[0x0f, 0x73, 0xc0, 0x7f]);
+    test_display(&[0x0f, 0x73, 0xd0, 0x7f], "psrlq mm0, 0x7f");
+    test_invalid(&[0x0f, 0x73, 0xe0, 0x7f]);
+    test_display(&[0x0f, 0x73, 0xf0, 0x7f], "psllq mm0, 0x7f");
+    test_display(&[0x0f, 0xa0], "push fs");
+    test_display(&[0x0f, 0xa1], "pop fs");
+    test_display(&[0x0f, 0xa2], "cpuid");
+    test_display(&[0x0f, 0xa4, 0xc0, 0x11], "shld eax, eax, 0x11");
+    test_display(&[0x66, 0x0f, 0xa4, 0xcf, 0x11], "shld di, cx, 0x11");
+    test_display(&[0x66, 0x45, 0x0f, 0xa4, 0xcf, 0x11], "shld r15w, r9w, 0x11");
+    test_display(&[0x0f, 0xa5, 0xc0], "shld eax, eax, cl");
+    test_display(&[0x0f, 0xa5, 0xc9], "shld ecx, ecx, cl");
+}
+
+#[test]
+fn prefixed_660f() {
+    test_display(&[0x66, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
+    test_display(&[0x66, 0x48, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
+    test_display(&[0x66, 0x4a, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
+    test_display(&[0x66, 0x4b, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm8");
+    test_display(&[0x66, 0x4c, 0x0f, 0x10, 0xc0], "movupd xmm8, xmm0");
+    test_display(&[0x66, 0x4d, 0x0f, 0x10, 0xc0], "movupd xmm8, xmm8");
+    test_display(&[0xf2, 0x66, 0x66, 0x4d, 0x0f, 0x10, 0xc0], "movupd xmm8, xmm8");
+}
+
+#[test]
+fn prefixed_f20f() {
+    test_display(&[0xf2, 0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
+    test_display(&[0xf2, 0x4d, 0x0f, 0x16, 0xcf], "movlhps xmm9, xmm15");
+    test_display(&[0x40, 0x66, 0xf2, 0x66, 0x4d, 0x0f, 0x16, 0xcf], "movlhps xmm9, xmm15");
+}
+
+#[test]
+fn prefixed_f30f() {
+    test_display(&[0xf3, 0x0f, 0x16, 0xcf], "movshdup xmm1, xmm7");
+    test_display(&[0xf3, 0x4d, 0x0f, 0x16, 0xcf], "movshdup xmm9, xmm15");
+}
diff --git a/test/long_mode/regspec.rs b/test/long_mode/regspec.rs
new file mode 100644
index 0000000..914c376
--- /dev/null
+++ b/test/long_mode/regspec.rs
@@ -0,0 +1,22 @@
+use yaxpeax_x86::long_mode::RegSpec;
+use std::collections::{BTreeMap, HashMap};
+
+#[test]
+fn test_ord() {
+    let _: BTreeMap<RegSpec, u64> = BTreeMap::new();
+}
+
+#[test]
+fn test_hash() {
+    let _: HashMap<RegSpec, u64> = HashMap::new();
+}
+
+#[test]
+fn test_labels() {
+    assert_eq!(RegSpec::rip().name(), "rip");
+    assert_eq!(RegSpec::eip().name(), "eip");
+    assert_eq!(RegSpec::rflags().name(), "rflags");
+    assert_eq!(RegSpec::rbp().name(), "rbp");
+    assert_eq!(RegSpec::gs().name(), "gs");
+    assert_eq!(RegSpec::al().name(), "al");
+}
diff --git a/test/protected_mode/mod.rs b/test/protected_mode/mod.rs
new file mode 100644
index 0000000..e1fb281
--- /dev/null
+++ b/test/protected_mode/mod.rs
@@ -0,0 +1,1035 @@
+mod regspec;
+
+use std::fmt::Write;
+
+use yaxpeax_arch::{Decoder, LengthedInstruction};
+use yaxpeax_x86::protected_mode::{DecodeError, InstDecoder, Opcode};
+
+fn test_invalid(data: &[u8]) {
+    test_invalid_under(&InstDecoder::default(), data);
+}
+
+fn test_invalid_under(decoder: &InstDecoder, data: &[u8]) {
+    if let Ok(inst) = decoder.decode(data.into_iter().cloned()) {
+        assert_eq!(inst.opcode, Opcode::Invalid, "decoded {:?} from {:02x?} under decoder {}", inst.opcode, data, decoder);
+    } else {
+        // this is fine
+    }
+}
+
+fn test_display(data: &[u8], expected: &'static str) {
+    test_display_under(&InstDecoder::default(), data, expected);
+}
+
+fn test_display_under(decoder: &InstDecoder, data: &[u8], expected: &'static str) {
+    let mut hex = String::new();
+    for b in data {
+        write!(hex, "{:02x}", b).unwrap();
+    }
+    match decoder.decode(data.into_iter().map(|x| *x)) {
+        Ok(instr) => {
+            let text = format!("{}", instr);
+            assert!(
+                text == expected,
+                "display error for {}:\n  decoded: {:?} under decoder {}\n displayed: {}\n expected: {}\n",
+                hex,
+                instr,
+                decoder,
+                text,
+                expected
+            );
+            // while we're at it, test that the instruction is as long, and no longer, than its
+            // input
+            assert_eq!(instr.len() as usize, data.len(), "instruction length is incorrect, wanted instruction {}", expected);
+        },
+        Err(e) => {
+            assert!(false, "decode error ({}) for {} under decoder {}:\n  expected: {}\n", e, hex, decoder, expected);
+        }
+    }
+}
+
+#[test]
+fn test_mmx() {
+    test_display(&[0x0f, 0x7e, 0xcf], "movd edi, mm1");
+    test_display(&[0x0f, 0x7f, 0xcf], "movq mm7, mm1");
+    test_display(&[0x0f, 0x7f, 0x0f], "movq [edi], mm1");
+    test_display(&[0x0f, 0xc4, 0xc0, 0x14], "pinsrw mm0, eax, 0x14");
+    test_display(&[0x0f, 0xc4, 0x00, 0x14], "pinsrw mm0, [eax], 0x14");
+    test_display(&[0x0f, 0xd1, 0xcf], "psrlw mm1, mm7");
+    test_display(&[0x0f, 0xd1, 0x00], "psrlw mm0, [eax]");
+    test_invalid(&[0x0f, 0xd7, 0x00]);
+    test_display(&[0x0f, 0xd7, 0xcf], "pmovmskb ecx, mm7");
+}
+
+#[test]
+fn test_cvt() {
+    test_display(&[0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm7");
+    test_display(&[0x0f, 0x2a, 0xcf], "cvtpi2ps xmm1, mm7");
+    test_display(&[0x0f, 0x2a, 0x00], "cvtpi2ps xmm0, [eax]");
+    test_display(&[0x66, 0x0f, 0x2a, 0x00], "cvtpi2pd xmm0, [eax]");
+    test_display(&[0x66, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm1, mm7");
+    test_display(&[0xf2, 0x0f, 0x2a, 0x00], "cvtsi2sd xmm0, [eax]");
+    test_display(&[0xf2, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm1, edi");
+    test_display(&[0xf3, 0x0f, 0x2a, 0x00], "cvtsi2ss xmm0, [eax]");
+    test_display(&[0xf3, 0x0f, 0x2a, 0xcf], "cvtsi2ss xmm1, edi");
+}
+
+#[test]
+fn test_aesni() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_aesni(), bytes, text);
+        test_display_under(&InstDecoder::default(), bytes, text);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+    }
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdb, 0x0f], "aesimc xmm1, [edi]");
+    test_instr(&[0x67, 0x66, 0x0f, 0x38, 0xdb, 0x0f], "aesimc xmm1, [bx]");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdc, 0x0f], "aesenc xmm1, [edi]");
+    test_instr(&[0x67, 0x66, 0x0f, 0x38, 0xdc, 0x0f], "aesenc xmm1, [bx]");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdd, 0x0f], "aesenclast xmm1, [edi]");
+    test_instr(&[0x67, 0x66, 0x0f, 0x38, 0xdd, 0x0f], "aesenclast xmm1, [bx]");
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xde, 0x0f], "aesdec xmm1, [edi]");
+    test_instr(&[0x67, 0x66, 0x0f, 0x38, 0xde, 0x0f], "aesdec xmm1, [bx]");
+    test_invalid(&[0x66, 0x67, 0x0f, 0x38, 0xde, 0x0f]);
+
+    test_instr(&[0x66, 0x0f, 0x38, 0xdf, 0x0f], "aesdeclast xmm1, [edi]");
+    test_instr(&[0x67, 0x66, 0x0f, 0x38, 0xdf, 0x0f], "aesdeclast xmm1, [bx]");
+
+    test_instr(&[0x66, 0x0f, 0x3a, 0xdf, 0x0f, 0xaa], "aeskeygenassist xmm1, [edi], 0xaa");
+}
+
+#[test]
+fn test_sse2() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        // sse and sse2 are part of amd64, so x86_64, meaning even the minimal decoder must support
+        // them.
+        test_display_under(&InstDecoder::minimal(), bytes, text);
+    }
+
+    test_instr(&[0xf2, 0x0f, 0x10, 0x0c, 0xc7], "movsd xmm1, [edi + eax * 8]");
+    test_instr(&[0xf2, 0x0f, 0x11, 0x0c, 0xc7], "movsd [edi + eax * 8], xmm1");
+    test_instr(&[0x66, 0x0f, 0x11, 0x0c, 0xc7], "movupd [edi + eax * 8], xmm1");
+    test_instr(&[0x66, 0x0f, 0x12, 0xc3], "movhlps xmm0, xmm3"); // reg-reg form is movhlps
+    test_instr(&[0x66, 0x0f, 0x12, 0x03], "movlpd xmm0, [ebx]"); // reg-mem is movlpd
+    test_instr(&[0x66, 0x0f, 0x13, 0x03], "movlpd [ebx], xmm0");
+    test_invalid(&[0x66, 0x0f, 0x13, 0xc3]);
+    test_instr(&[0x66, 0x0f, 0x14, 0x03], "unpcklpd xmm0, [ebx]");
+    test_instr(&[0x66, 0x0f, 0x14, 0xc3], "unpcklpd xmm0, xmm3");
+    test_instr(&[0x66, 0x0f, 0x15, 0x03], "unpckhpd xmm0, [ebx]");
+    test_instr(&[0x66, 0x0f, 0x15, 0xc3], "unpckhpd xmm0, xmm3");
+    test_instr(&[0x66, 0x0f, 0x16, 0x03], "movhpd xmm0, [ebx]");
+    test_instr(&[0x66, 0x0f, 0x16, 0xc3], "movlhps xmm0, xmm3");
+    test_instr(&[0x66, 0x0f, 0x17, 0x03], "movhpd [ebx], xmm0");
+    test_invalid(&[0x66, 0x0f, 0x17, 0xc3]);
+
+    test_instr(&[0x66, 0x0f, 0x28, 0xd0], "movapd xmm2, xmm0");
+    test_instr(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
+
+    test_instr(&[0x66, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm1, mm7");
+    test_instr(&[0x66, 0x0f, 0x2a, 0x0f], "cvtpi2pd xmm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm1, edi");
+    test_instr(&[0xf2, 0x0f, 0x2a, 0x0f], "cvtsi2sd xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0x2b, 0x0f], "movntpd [edi], xmm1");
+    test_instr(&[0x66, 0x0f, 0x2c, 0xcf], "cvttpd2pi mm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0x2c, 0x0f], "cvttpd2pi mm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0x2c, 0xcf], "cvttsd2si xmm1, xmm7");
+    test_instr(&[0xf2, 0x0f, 0x2c, 0x0f], "cvttsd2si xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0x2d, 0xcf], "cvtpd2pi mm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0x2d, 0x0f], "cvtpd2pi mm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0x2d, 0xcf], "cvtsd2si xmm1, xmm7");
+    test_instr(&[0xf2, 0x0f, 0x2d, 0x0f], "cvtsd2si xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0x2e, 0xcf], "ucomisd xmm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0x2e, 0x0f], "ucomisd xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0x2f, 0xcf], "comisd xmm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0x2f, 0x0f], "comisd xmm1, [edi]");
+
+    /*
+     * .... 660f38
+     * .... 660f7f
+     */
+
+    test_invalid(&[0x66, 0x0f, 0x50, 0x01]);
+    test_instr(&[0x66, 0x0f, 0x50, 0xc1], "movmskpd eax, xmm1");
+    test_instr(&[0x66, 0x0f, 0x51, 0x01], "sqrtpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x51, 0x01], "sqrtsd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x52, 0x01], "rsqrtps xmm0, [ecx]"); // note: NOT "rsqrtpd" - no such instruction exists, so fall back to just 0f52 parse.
+    test_instr(&[0x66, 0x0f, 0x53, 0x01], "rcpps xmm0, [ecx]"); // note: NOT "rcppd" - no such instruction exists, so fall back to just 0f53 parse.
+    test_instr(&[0x66, 0x0f, 0x54, 0x01], "andpd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x55, 0x01], "andnpd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x56, 0x01], "orpd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x57, 0x01], "xorpd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x58, 0x01], "addpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x58, 0x01], "addsd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x59, 0x01], "mulpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x59, 0x01], "mulsd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x5a, 0x01], "cvtpd2ps xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x5a, 0x01], "cvtsd2ss xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x5b, 0x01], "cvtps2dq xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x5c, 0x01], "subpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x5c, 0x01], "subsd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x5d, 0x01], "minpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x5d, 0x01], "minsd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x5e, 0x01], "divpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x5e, 0x01], "divsd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0x5f, 0x01], "maxpd xmm0, [ecx]");
+    test_instr(&[0xf2, 0x0f, 0x5f, 0x01], "maxsd xmm0, [ecx]");
+    test_instr(
+        &[0x66, 0x0f, 0x60, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpcklbw xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x61, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpcklwd xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x62, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckldq xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x63, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "packsswb xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x64, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "pcmpgtb xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x65, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "pcmpgtw xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x66, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "pcmpgtd xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x67, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "packuswb xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x68, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhbw xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x69, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhwd xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x6a, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhdq xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x6b, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "packssdw xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x6c, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpcklqdq xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x6d, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "punpckhqdq xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    // TODO: this needs to be clear that the operand is `dword`
+    test_instr(
+        &[0x66, 0x0f, 0x6e, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movd xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_instr(
+        &[0x66, 0x0f, 0x6f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqa xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+
+    test_instr(&[0x66, 0x0f, 0x6e, 0xc0], "movd xmm0, eax");
+    test_instr(&[0x66, 0x0f, 0x70, 0xc0, 0x4e], "pshufd xmm0, xmm0, 0x4e");
+    test_instr(&[0xf2, 0x0f, 0x70, 0xc0, 0x4e], "pshuflw xmm0, xmm0, 0x4e");
+    test_instr(&[0xf3, 0x0f, 0x70, 0xc0, 0x4e], "pshufhw xmm0, xmm0, 0x4e");
+    test_invalid(&[0x66, 0x0f, 0x71, 0x10, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x71, 0xd0, 0x8f], "psrlw xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x71, 0x20, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x71, 0xe0, 0x8f], "psraw xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x71, 0x30, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x71, 0xf0, 0x8f], "psllw xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x72, 0x10, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x72, 0xd0, 0x8f], "psrld xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x72, 0x20, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x72, 0xe0, 0x8f], "psrad xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x72, 0x30, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x72, 0xf0, 0x8f], "pslld xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x73, 0x10, 0x8f]);
+    test_invalid(&[0x66, 0x0f, 0x73, 0x18, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x73, 0xd0, 0x8f], "psrlq xmm0, 0x8f");
+    test_instr(&[0x66, 0x0f, 0x73, 0xd8, 0x8f], "psrldq xmm0, 0x8f");
+    test_invalid(&[0x66, 0x0f, 0x73, 0x30, 0x8f]);
+    test_invalid(&[0x66, 0x0f, 0x73, 0x38, 0x8f]);
+    test_instr(&[0x66, 0x0f, 0x73, 0xf0, 0x8f], "psllq xmm0, 0x8f");
+    test_instr(&[0x66, 0x0f, 0x73, 0xf8, 0x8f], "pslldq xmm0, 0x8f");
+    test_instr(&[0x66, 0x0f, 0x7e, 0xc1], "movd ecx, xmm0");
+    test_instr(&[0x66, 0x0f, 0x7e, 0x01], "movd [ecx], xmm0");
+    test_instr(
+        &[0x66, 0x0f, 0x7f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqa [esp + ebx * 4 - 0x334455cc], xmm3"
+    );
+
+    test_instr(&[0x66, 0x0f, 0xc2, 0xc3, 0x08], "cmppd xmm0, xmm3, 0x8");
+    test_instr(&[0x66, 0x0f, 0xc2, 0x03, 0x08], "cmppd xmm0, [ebx], 0x8");
+    test_instr(&[0xf2, 0x0f, 0xc2, 0xc3, 0x08], "cmpsd xmm0, xmm3, 0x8");
+    test_instr(&[0xf2, 0x0f, 0xc2, 0x03, 0x08], "cmpsd xmm0, [ebx], 0x8");
+
+    test_instr(&[0x66, 0x0f, 0xc4, 0xc3, 0x08], "pinsrw xmm0, ebx, 0x8");
+    test_instr(&[0x66, 0x0f, 0xc4, 0x03, 0x08], "pinsrw xmm0, [ebx], 0x8");
+
+//    test_instr(&[0x66, 0x0f, 0xc5, 0xc3, 0x08], "pextrw eax, xmm3, 0x8");
+//    test_instr(&[0x66, 0x4f, 0x0f, 0xc5, 0xc3, 0x08], "pextrw eax, xmm11, 0x8");
+//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x03, 0x08]);
+//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x40, 0x08]);
+//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x80, 0x08]);
+
+    test_instr(&[0x66, 0x0f, 0xc6, 0x03, 0x08], "shufpd xmm0, [ebx], 0x8");
+    test_instr(&[0x66, 0x0f, 0xc6, 0xc3, 0x08], "shufpd xmm0, xmm3, 0x8");
+    test_instr(&[0x66, 0x0f, 0xd1, 0xc1], "psrlw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd1, 0x01], "psrlw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xd2, 0xc1], "psrld xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd2, 0x01], "psrld xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xd3, 0xc1], "psrlq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd3, 0x01], "psrlq xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xd4, 0xc1], "paddq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd4, 0x01], "paddq xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xd5, 0xc1], "pmullw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd5, 0x01], "pmullw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xd6, 0xc1], "movq xmm1, xmm0");
+    test_instr(&[0x66, 0x0f, 0xd6, 0x01], "movq [ecx], xmm0");
+    test_invalid(&[0xf3, 0x0f, 0xd6, 0x03]);
+    test_instr(&[0xf3, 0x0f, 0xd6, 0xc3], "movq2dq xmm0, mm3");
+    test_instr(&[0xf2, 0x0f, 0xd6, 0xc3], "movdq2q mm0, xmm3");
+    test_instr(&[0x66, 0x0f, 0xd7, 0xc1], "pmovmskb eax, xmm1");
+    test_invalid(&[0x66, 0x0f, 0xd7, 0x01]);
+    test_instr(&[0x66, 0x0f, 0xd8, 0xc1], "psubusb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd8, 0x01], "psubusb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xd9, 0xc1], "psubusw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xd9, 0x01], "psubusw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xda, 0xc1], "pminub xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xda, 0x01], "pminub xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xdb, 0xc1], "pand xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdb, 0x01], "pand xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xdc, 0xc1], "paddusb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdc, 0x01], "paddusb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xdd, 0xc1], "paddusw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdd, 0x01], "paddusw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xde, 0xc1], "pmaxub xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xde, 0x01], "pmaxub xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xdf, 0xc1], "pandn xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xdf, 0x01], "pandn xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe0, 0xc1], "pavgb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe0, 0x01], "pavgb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe1, 0xc1], "psraw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe1, 0x01], "psraw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe2, 0xc1], "psrad xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe2, 0x01], "psrad xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe3, 0xc1], "pavgw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe3, 0x01], "pavgw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe4, 0xc1], "pmulhuw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe4, 0x01], "pmulhuw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe5, 0xc1], "pmulhw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe5, 0x01], "pmulhw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe6, 0xc1], "cvttpd2dq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe6, 0x01], "cvttpd2dq xmm0, [ecx]");
+    test_invalid(&[0x66, 0x0f, 0xe7, 0xc1]);
+    test_instr(&[0x66, 0x0f, 0xe7, 0x01], "movntdq [ecx], xmm0");
+    test_instr(&[0x66, 0x0f, 0xe8, 0xc1], "psubsb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe8, 0x01], "psubsb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xe9, 0xc1], "psubsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xe9, 0x01], "psubsw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xea, 0xc1], "pminsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xea, 0x01], "pminsw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xc3], "por xmm0, xmm3");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xc4], "por xmm0, xmm4");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xd3], "por xmm2, xmm3");
+    test_instr(&[0x66, 0x0f, 0xeb, 0x12], "por xmm2, [edx]");
+    test_instr(&[0x66, 0x0f, 0xeb, 0xc1], "por xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xeb, 0x01], "por xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xec, 0xc1], "paddsb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xec, 0x01], "paddsb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xed, 0xc1], "paddsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xed, 0x01], "paddsw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xee, 0xc1], "pmaxsw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xee, 0x01], "pmaxsw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xef, 0xc1], "pxor xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xef, 0x01], "pxor xmm0, [ecx]");
+    test_invalid(&[0x66, 0x0f, 0xf0, 0xc1]);
+    test_invalid(&[0x66, 0x0f, 0xf0, 0x01]);
+    test_instr(&[0x66, 0x0f, 0xf1, 0xc1], "psllw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf1, 0x01], "psllw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf2, 0xc1], "pslld xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf2, 0x01], "pslld xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf3, 0xc1], "psllq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf3, 0x01], "psllq xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf4, 0xc1], "pmuludq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf4, 0x01], "pmuludq xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf5, 0xc1], "pmaddwd xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf5, 0x01], "pmaddwd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf6, 0xc1], "psadbw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf6, 0x01], "psadbw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf7, 0xc1], "maskmovdqu xmm0, xmm1");
+    test_invalid(&[0x66, 0x0f, 0xf7, 0x01]);
+    test_instr(&[0x66, 0x0f, 0xf8, 0xc1], "psubb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf8, 0x01], "psubb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xf9, 0xc1], "psubw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xf9, 0x01], "psubw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xfa, 0xc1], "psubd xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfa, 0x01], "psubd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xfb, 0xc1], "psubq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfb, 0x01], "psubq xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xfc, 0xc1], "paddb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfc, 0x01], "paddb xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xfd, 0xc1], "paddw xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfd, 0x01], "paddw xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xfe, 0xc1], "paddd xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xfe, 0x01], "paddd xmm0, [ecx]");
+    test_instr(&[0x66, 0x0f, 0xff, 0xc1], "paddq xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0xff, 0x01], "paddq xmm0, [ecx]");
+
+    test_instr(&[0x66, 0x0f, 0x74, 0xc1], "pcmpeqb xmm0, xmm1");
+    test_instr(&[0x66, 0x0f, 0x74, 0x12], "pcmpeqb xmm2, [edx]");
+    test_instr(&[0x66, 0x0f, 0xf8, 0xc8], "psubb xmm1, xmm0");
+    test_instr(&[0x66, 0x0f, 0xf8, 0xd0], "psubb xmm2, xmm0");
+    test_instr(&[0x66, 0x0f, 0xf8, 0x12], "psubb xmm2, [edx]");
+}
+
+#[test]
+fn test_sse3() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_sse3(), 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);
+    }
+
+    fn test_instr_invalid(bytes: &[u8]) {
+        test_invalid_under(&InstDecoder::minimal().with_sse3(), bytes);
+        test_invalid_under(&InstDecoder::default(), bytes);
+    }
+    test_instr(&[0xf2, 0x0f, 0xf0, 0x0f], "lddqu xmm1, [edi]");
+    test_instr_invalid(&[0xf2, 0x0f, 0xf0, 0xcf]);
+    test_instr(&[0xf2, 0x0f, 0xd0, 0x0f], "addsubps xmm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0xd0, 0xcf], "addsubps xmm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0xd0, 0x0f], "addsubpd xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0xd0, 0xcf], "addsubpd xmm1, xmm7");
+
+    test_instr(&[0xf2, 0x0f, 0x7c, 0x0f], "haddps xmm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0x7c, 0xcf], "haddps xmm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0x7c, 0x0f], "haddpd xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0x7c, 0xcf], "haddpd xmm1, xmm7");
+
+    test_instr(&[0xf2, 0x0f, 0x7d, 0x0f], "hsubps xmm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0x7d, 0xcf], "hsubps xmm1, xmm7");
+    test_instr(&[0x66, 0x0f, 0x7d, 0x0f], "hsubpd xmm1, [edi]");
+    test_instr(&[0x66, 0x0f, 0x7d, 0xcf], "hsubpd xmm1, xmm7");
+
+    test_instr(&[0xf3, 0x0f, 0x12, 0x0f], "movsldup xmm1, [edi]");
+    test_instr(&[0xf3, 0x0f, 0x12, 0xcf], "movsldup xmm1, xmm7");
+    test_instr(&[0xf3, 0x0f, 0x16, 0x0f], "movshdup xmm1, [edi]");
+    test_instr(&[0xf3, 0x0f, 0x16, 0xcf], "movshdup xmm1, xmm7");
+
+    test_instr(&[0xf2, 0x0f, 0x12, 0x0f], "movddup xmm1, [edi]");
+    test_instr(&[0xf2, 0x0f, 0x12, 0xcf], "movddup xmm1, xmm7");
+
+    test_instr(&[0x66, 0x0f, 0x01, 0xc8], "monitor");
+    test_instr(&[0xf2, 0x0f, 0x01, 0xc8], "monitor");
+    test_instr(&[0xf3, 0x0f, 0x01, 0xc8], "monitor");
+
+    test_instr(&[0x66, 0x0f, 0x01, 0xc9], "mwait");
+    test_instr(&[0xf2, 0x0f, 0x01, 0xc9], "mwait");
+    test_instr(&[0xf3, 0x0f, 0x01, 0xc9], "mwait");
+}
+
+#[test]
+fn test_0f01() {
+    // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
+    // Number"
+    test_display(&[0x0f, 0x01, 0x38], "invlpg [eax]");
+    test_display(&[0x0f, 0x01, 0x3f], "invlpg [edi]");
+    test_display(&[0x0f, 0x01, 0x40, 0xff], "sgdt [eax - 0x1]");
+    test_display(&[0x0f, 0x01, 0x41, 0xff], "sgdt [ecx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x49, 0xff], "sidt [ecx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x51, 0xff], "lgdt [ecx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x59, 0xff], "lidt [ecx - 0x1]");
+    test_display(&[0x0f, 0x01, 0x61, 0xff], "smsw [ecx - 0x1]");
+    test_display(&[0x0f, 0x01, 0xc0], "enclv");
+    test_display(&[0x0f, 0x01, 0xc1], "vmcall");
+    test_display(&[0x0f, 0x01, 0xc2], "vmlaunch");
+    test_display(&[0x0f, 0x01, 0xc3], "vmresume");
+    test_display(&[0x0f, 0x01, 0xc4], "vmxoff");
+    test_invalid(&[0x0f, 0x01, 0xc5]);
+    test_invalid(&[0x0f, 0x01, 0xc6]);
+    test_invalid(&[0x0f, 0x01, 0xc7]);
+    test_display(&[0x0f, 0x01, 0xc8], "monitor");
+    test_display(&[0x0f, 0x01, 0xc9], "mwait");
+    test_display(&[0x0f, 0x01, 0xca], "clac");
+    test_display(&[0x0f, 0x01, 0xcb], "stac");
+    test_display(&[0x0f, 0x01, 0xcf], "encls");
+    test_display(&[0x0f, 0x01, 0xd0], "xgetbv");
+    test_display(&[0x0f, 0x01, 0xd1], "xsetbv");
+    test_invalid(&[0x0f, 0x01, 0xd2]);
+    test_invalid(&[0x0f, 0x01, 0xd3]);
+    test_display(&[0x0f, 0x01, 0xd4], "vmfunc");
+    test_display(&[0x0f, 0x01, 0xd5], "xend");
+    test_display(&[0x0f, 0x01, 0xd6], "xtest");
+    test_display(&[0x0f, 0x01, 0xd7], "enclu");
+    test_invalid(&[0x0f, 0x01, 0xd8]);
+    test_invalid(&[0x0f, 0x01, 0xd9]);
+    test_invalid(&[0x0f, 0x01, 0xda]);
+    test_invalid(&[0x0f, 0x01, 0xdb]);
+    test_invalid(&[0x0f, 0x01, 0xdc]);
+    test_invalid(&[0x0f, 0x01, 0xdd]);
+    test_invalid(&[0x0f, 0x01, 0xde]);
+    test_invalid(&[0x0f, 0x01, 0xdf]);
+    test_display(&[0x0f, 0x01, 0xee], "rdpkru");
+    test_display(&[0x0f, 0x01, 0xef], "wrpkru");
+    test_display(&[0x0f, 0x01, 0xf8], "swapgs");
+    test_display(&[0x0f, 0x01, 0xf9], "rdtscp");
+}
+
+#[test]
+fn test_0fae() {
+    let intel = InstDecoder::minimal().with_intel_quirks();
+    let amd = InstDecoder::minimal().with_amd_quirks();
+    let default = InstDecoder::default();
+    let minimal = InstDecoder::minimal();
+    // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
+    // Number"
+    test_display(&[0x0f, 0xae, 0x04, 0x4f], "fxsave [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x0c, 0x4f], "fxrstor [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x14, 0x4f], "ldmxcsr [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x1c, 0x4f], "stmxcsr [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x24, 0x4f], "xsave [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x2c, 0x4f], "xrstor [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x34, 0x4f], "xsaveopt [edi + ecx * 2]");
+    test_display(&[0x0f, 0xae, 0x3c, 0x4f], "clflush [edi + ecx * 2]");
+
+    for (modrm, text) in &[(0xe8u8, "lfence"), (0xf0u8, "mfence"), (0xf8u8, "sfence")] {
+        test_display_under(&intel, &[0x0f, 0xae, *modrm], text);
+        test_display_under(&amd, &[0x0f, 0xae, *modrm], text);
+        test_display_under(&default, &[0x0f, 0xae, *modrm], text);
+        test_display_under(&minimal, &[0x0f, 0xae, *modrm], text);
+        // it turns out intel and amd accept m != 0 for {l,m,s}fence:
+        // from intel:
+        // ```
+        // Specification of the instruction's opcode above indicates a ModR/M byte of F0. For this
+        // instruction, the processor ignores the r/m field of the ModR/M byte. Thus, MFENCE is encoded
+        // by any opcode of the form 0F AE Fx, where x is in the range 0-7.
+        // ```
+        // whereas amd does not discuss the r/m field at all. at least as of zen, amd also accepts
+        // these encodings.
+        for m in 1u8..8u8 {
+            test_display_under(&intel, &[0x0f, 0xae, modrm | m], text);
+            test_display_under(&amd, &[0x0f, 0xae, modrm | m], text);
+            test_display_under(&default, &[0x0f, 0xae, modrm | m], text);
+            test_invalid_under(&minimal, &[0x0f, 0xae, modrm | m]);
+        }
+    }
+}
+
+#[test]
+fn test_system() {
+    test_display(&[0x63, 0xc1], "arpl cx, ax");
+    test_display(&[0x63, 0x04, 0xba], "arpl [edx + edi * 4], ax");
+}
+
+#[test]
+fn test_arithmetic() {
+    test_display(&[0x81, 0xec, 0x10, 0x03, 0x00, 0x00], "sub esp, 0x310");
+    test_display(&[0x0f, 0xaf, 0xc2], "imul eax, edx");
+    test_display(&[0x69, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65], "imul eax, [ebx + 0x6f], 0x656c706d");
+    test_display(&[0x66, 0x0f, 0xaf, 0xd1], "imul dx, cx");
+    test_display(&[0x6b, 0x43, 0x6f, 0x6d], "imul al, [ebx + 0x6f], 0x6d");
+    test_display(&[0x00, 0xcc], "add ah, cl");
+}
+
+#[test]
+#[allow(non_snake_case)]
+fn test_E_decode() {
+    test_display(&[0xff, 0x75, 0xb8], "push [ebp - 0x48]");
+    test_display(&[0xff, 0x75, 0x08], "push [ebp + 0x8]");
+}
+
+#[test]
+fn test_sse() {
+    test_display(&[0xf3, 0x0f, 0x10, 0x0c, 0xc7], "movss xmm1, [edi + eax * 8]");
+    test_display(&[0xf3, 0x0f, 0x11, 0x0c, 0xc7], "movss [edi + eax * 8], xmm1");
+    test_display(&[0x0f, 0x28, 0x00], "movaps xmm0, [eax]");
+    test_display(&[0x0f, 0x29, 0x00], "movaps [eax], xmm0");
+    test_display(&[0xf3, 0x0f, 0x2a, 0xc1], "cvtsi2ss xmm0, ecx");
+    test_display(&[0xf3, 0x0f, 0x2a, 0x01], "cvtsi2ss xmm0, [ecx]");
+    test_display(&[0x0f, 0x2b, 0x00], "movntps [eax], xmm0");
+    test_display(&[0xf3, 0x0f, 0x2c, 0xc1], "cvttss2si eax, xmm1");
+    test_display(&[0xf3, 0x0f, 0x2c, 0x01], "cvttss2si eax, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x2d, 0xc1], "cvtss2si eax, xmm1");
+    test_display(&[0xf3, 0x0f, 0x2d, 0x01], "cvtss2si eax, [ecx]");
+    test_display(&[0x0f, 0x2e, 0x00], "ucomiss xmm0, [eax]");
+    test_display(&[0x0f, 0x2f, 0x00], "comiss xmm0, [eax]");
+    test_display(&[0x0f, 0x28, 0xd0], "movaps xmm2, xmm0");
+    test_display(&[0x66, 0x0f, 0x28, 0xd0], "movapd xmm2, xmm0");
+    test_display(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
+    test_display(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
+    test_display(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
+    test_display(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
+    test_display(&[0x67, 0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [bx + si]");
+    test_display(&[0x66, 0x0f, 0x29, 0x00], "movapd [eax], xmm0");
+    test_invalid(&[0x0f, 0x50, 0x00]);
+    test_display(&[0x0f, 0x50, 0xc1], "movmskps eax, xmm1");
+    test_display(&[0x0f, 0x51, 0x01], "sqrtps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x51, 0x01], "sqrtss xmm0, [ecx]");
+    test_display(&[0x0f, 0x52, 0x01], "rsqrtps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x52, 0x01], "rsqrtss xmm0, [ecx]");
+    test_display(&[0x0f, 0x53, 0x01], "rcpps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x53, 0x01], "rcpss xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x53, 0xc1], "rcpss xmm0, xmm1");
+    test_display(&[0x0f, 0x54, 0x01], "andps xmm0, [ecx]");
+    test_display(&[0x0f, 0x55, 0x01], "andnps xmm0, [ecx]");
+    test_display(&[0x0f, 0x56, 0x01], "orps xmm0, [ecx]");
+    test_display(&[0x0f, 0x57, 0x01], "xorps xmm0, [ecx]");
+    test_display(&[0x0f, 0x58, 0x01], "addps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x58, 0x01], "addss xmm0, [ecx]");
+    test_display(&[0x0f, 0x59, 0x01], "mulps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x59, 0x01], "mulss xmm0, [ecx]");
+    test_display(&[0x0f, 0x5a, 0x01], "cvtps2pd xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x5a, 0x01], "cvtss2sd xmm0, [ecx]");
+    test_display(&[0x0f, 0x5b, 0x01], "cvtdq2ps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x5b, 0x01], "cvttps2dq xmm0, [ecx]");
+    test_display(&[0x67, 0x0f, 0x5b, 0x01], "cvtdq2ps xmm0, [bx + di]");
+    test_display(&[0x0f, 0x5c, 0x01], "subps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x5c, 0x01], "subss xmm0, [ecx]");
+    test_display(&[0x0f, 0x5d, 0x01], "minps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x5d, 0x01], "minss xmm0, [ecx]");
+    test_display(&[0x0f, 0x5e, 0x01], "divps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x5e, 0x01], "divss xmm0, [ecx]");
+    test_display(&[0x0f, 0x5f, 0x01], "maxps xmm0, [ecx]");
+    test_display(&[0xf3, 0x0f, 0x5f, 0x01], "maxss xmm0, [ecx]");
+
+    test_display(&[0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [eax], 0x7f");
+
+    test_display(&[0x66, 0x0f, 0xef, 0xc0], "pxor xmm0, xmm0");
+    test_display(&[0x66, 0x0f, 0xef, 0xc0], "pxor xmm0, xmm0");
+    test_display(&[0xf2, 0x0f, 0x10, 0x0c, 0xc6], "movsd xmm1, [esi + eax * 8]");
+    test_display(&[0xf3, 0x0f, 0x10, 0x04, 0x86], "movss xmm0, [esi + eax * 4]");
+    test_display(&[0xf2, 0x0f, 0x59, 0xc8], "mulsd xmm1, xmm0");
+    test_display(&[0xf3, 0x0f, 0x59, 0xc8], "mulss xmm1, xmm0");
+
+    test_display(
+        &[0xf3, 0x0f, 0x6f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqu xmm3, [esp + ebx * 4 - 0x334455cc]"
+    );
+    test_display(&[0xf3, 0x0f, 0x70, 0xc0, 0x4e], "pshufhw xmm0, xmm0, 0x4e");
+    test_display(&[0xf3, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
+    test_display(
+        &[0xf3, 0x0f, 0x7f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
+        "movdqu [esp + ebx * 4 - 0x334455cc], xmm3"
+    );
+
+    test_display(&[0xf3, 0x0f, 0xc2, 0xc3, 0x08], "cmpss xmm0, xmm3, 0x8");
+    test_display(&[0xf3, 0x0f, 0xc2, 0x03, 0x08], "cmpss xmm0, [ebx], 0x8");
+}
+
+// SETLE, SETNG, ...
+
+#[test]
+fn test_mov() {
+    // test_display(&[0xa1, 0x93, 0x62, 0xc4, 0x00, 0x12, 0x34, 0x12, 0x34], "mov eax, [0x3412341200c46293]");
+    // RCT.exe 32bit version, TODO: FIX
+    test_display(&[0xa1, 0x93, 0x62, 0xc4, 0x00], "mov eax, [0xc46293]");
+    test_display(&[0xba, 0x01, 0x00, 0x00, 0x00], "mov edx, 0x1");
+    test_display(&[0xc7, 0x04, 0x24, 0x00, 0x00, 0x00, 0x00], "mov [esp], 0x0");
+    test_display(&[0x89, 0x44, 0x24, 0x08], "mov [esp + 0x8], eax");
+    test_display(&[0x89, 0x43, 0x18], "mov [ebx + 0x18], eax");
+    test_display(&[0xc7, 0x43, 0x10, 0x00, 0x00, 0x00, 0x00], "mov [ebx + 0x10], 0x0");
+    test_display(&[0x89, 0x4e, 0x08], "mov [esi + 0x8], ecx");
+    test_display(&[0x8b, 0x32], "mov esi, [edx]");
+    test_display(&[0x8b, 0x4c, 0x10, 0xf8], "mov ecx, [eax + edx * 1 - 0x8]");
+    test_display(&[0x89, 0x46, 0x10], "mov [esi + 0x10], eax");
+    test_display(&[0x0f, 0x43, 0xec], "cmovnb ebp, esp");
+    test_display(&[0x0f, 0xb6, 0x06], "movzx eax, byte [esi]");
+    test_display(&[0x0f, 0xb7, 0x06], "movzx eax, word [esi]");
+    test_display(&[0x89, 0x55, 0x94], "mov [ebp - 0x6c], edx");
+    test_display(&[0x65, 0x89, 0x04, 0x25, 0xa8, 0x01, 0x00, 0x00], "mov gs:[0x1a8], eax");
+    test_display(&[0x0f, 0xbe, 0x83, 0xb4, 0x00, 0x00, 0x00], "movsx eax, byte [ebx + 0xb4]");
+    test_display(&[0xf3, 0x0f, 0x6f, 0x07], "movdqu xmm0, [edi]");
+    test_display(&[0xf3, 0x0f, 0x7f, 0x45, 0x00], "movdqu [ebp], xmm0");
+}
+
+#[test]
+fn test_stack() {
+    test_display(&[0x66, 0x50], "push ax");
+}
+
+#[test]
+fn test_prefixes() {
+    test_display(&[0x66, 0x31, 0xc0], "xor ax, ax");
+    test_display(&[0x66, 0x32, 0xc0], "xor al, al");
+    test_display(&[0x66, 0x32, 0xc5], "xor al, ch");
+
+    // test that WAIT doesn't blow up, at least...
+    assert_eq!(InstDecoder::default().decode([0x9b, 0xf8].iter().cloned()).err(), Some(DecodeError::IncompleteDecoder));
+}
+
+#[test]
+fn test_control_flow() {
+    test_display(&[0x73, 0x31], "jnb 0x31");
+    test_display(&[0x72, 0x5a], "jb 0x5a");
+    test_display(&[0x0f, 0x86, 0x8b, 0x01, 0x00, 0x00], "jna 0x18b");
+    test_display(&[0x74, 0x47], "jz 0x47");
+    test_display(&[0xff, 0x15, 0x7e, 0x72, 0x24, 0x00], "call [0x24727e]");
+    test_display(&[0xff, 0x24, 0xcd, 0x70, 0xa0, 0xbc, 0x01], "jmp [ecx * 8 + 0x1bca070]");
+    test_display(&[0xff, 0xe0], "jmp eax");
+    test_display(&[0x66, 0xff, 0xe0], "jmp eax");
+    test_display(&[0x67, 0xff, 0xe0], "jmp eax");
+    test_invalid(&[0xff, 0xd8]);
+    test_display(&[0xff, 0x18], "callf [eax]");
+    test_display(&[0xc3], "ret");
+}
+
+#[test]
+fn test_test_cmp() {
+    test_display(&[0xf6, 0x05, 0x2c, 0x9b, 0xff, 0xff, 0x01], "test [0xffff9b2c], 0x1");
+    test_display(&[0x3d, 0x01, 0xf0, 0xff, 0xff], "cmp eax, -0xfff");
+    test_display(&[0x83, 0xf8, 0xff], "cmp eax, -0x1");
+    test_display(&[0x39, 0xc6], "cmp esi, eax");
+}
+
+#[test]
+fn test_push_pop() {
+    test_display(&[0x5b], "pop ebx");
+    test_display(&[0x5e], "pop esi");
+    test_display(&[0x68, 0x7f, 0x63, 0xc4, 0x00], "push 0xc4637f");
+}
+
+#[test]
+fn test_bmi1() {
+    let bmi1 = InstDecoder::minimal().with_bmi1();
+    let no_bmi1 = InstDecoder::minimal();
+    test_display_under(&bmi1, &[0x0f, 0xbc, 0xd3], "tzcnt edx, ebx");
+    test_display_under(&no_bmi1, &[0x0f, 0xbc, 0xd3], "bsf edx, ebx");
+
+    test_display_under(&bmi1, &[0xf3, 0x0f, 0xb8, 0xc1], "popcnt eax, ecx");
+}
+
+#[test]
+fn test_bitwise() {
+    test_display_under(&InstDecoder::minimal(), &[0x0f, 0xbc, 0xd3], "bsf edx, ebx");
+    test_display(&[0x0f, 0xa3, 0xd0], "bt eax, edx");
+    test_display(&[0x0f, 0xab, 0xd0], "bts eax, edx");
+}
+
+#[test]
+fn test_misc() {
+    // TODO
+//    test_display(&[0xf2, 0x0f, 0x38, 0xf0, 0xc1], "crc32 eax, cl");
+//    test_display(&[0xf2, 0x0f, 0x38, 0xf1, 0xc1], "crc32 eax, ecx");
+    test_display(&[0xfe, 0x00], "inc [eax]"); // TODO: inc byte [eax]
+    test_display(&[0xfe, 0x08], "dec [eax]"); // TODO: dec byte [eax]
+    test_display(&[0xff, 0x00], "inc [eax]"); // TODO: inc dword [eax]
+    test_display(&[0xff, 0x08], "dec [eax]"); // TODO: inc qword [eax]
+    test_display(&[0xe4, 0x99], "in al, 0x99");
+    test_display(&[0xe5, 0x99], "in eax, 0x99");
+    test_display(&[0x67, 0xe5, 0x99], "in eax, 0x99");
+    test_display(&[0xe5, 0x99], "in eax, 0x99");
+    test_display(&[0xe6, 0x99], "out 0x99, al");
+    test_display(&[0xe7, 0x99], "out 0x99, eax");
+    test_display(&[0xec], "in al, dx");
+    test_display(&[0xed], "in eax, dx");
+    test_display(&[0xee], "out dx, al");
+    test_display(&[0xef], "out dx, eax");
+    test_display(&[0xcd, 0x00], "int 0x0");
+    test_display(&[0xcd, 0xff], "int 0xff");
+    test_display(&[0x9c], "pushf");
+    test_display(&[0x98], "cwde");
+    test_display(&[0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00], "nop cs:[eax + eax * 1]");
+    test_display(&[0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00], "nop [eax + eax * 1]");
+    test_display(&[0x8d, 0xa4, 0xc7, 0x20, 0x00, 0x00, 0x12], "lea esp, [edi + eax * 8 + 0x12000020]");
+    test_display(&[0x33, 0xc0], "xor eax, eax");
+    test_display(&[0x8d, 0x53, 0x08], "lea edx, [ebx + 0x8]");
+    test_display(&[0x31, 0xc9], "xor ecx, ecx");
+    test_display(&[0x29, 0xc8], "sub eax, ecx");
+    test_display(&[0x03, 0x0b], "add ecx, [ebx]");
+    test_display(&[0x8d, 0x0c, 0x12], "lea ecx, [edx + edx * 1]");
+    test_display(&[0xf6, 0xc2, 0x18], "test dl, 0x18");
+    test_display(&[0xf3, 0xab], "rep stos es:[edi], eax");
+    test_display(&[0xf3, 0xa5], "rep movs es:[edi], ds:[esi]");
+    test_display(&[0xf3, 0x0f, 0xbc, 0xd7], "tzcnt edx, edi");
+
+    // this is actually vmx
+    // test_invalid(&[0x66, 0x0f, 0xc7, 0x03]);
+    test_display(&[0x66, 0x0f, 0xc7, 0x33], "vmclear [ebx]");
+    test_display(&[0xf3, 0x0f, 0xc7, 0x33], "vmxon [ebx]");
+}
+
+#[test]
+#[ignore]
+// TODO also not supported at all
+fn evex() {
+    test_display(&[0x62, 0xf2, 0x7d, 0x48, 0x2a, 0x44, 0x40, 0x01], "vmovntdqa zmm0, [eax + eax*2 + 0x40]");
+    test_display(&[0x62, 0xf2, 0x7d, 0x08, 0x2a, 0x44, 0x40, 0x01], "vmovntdqa xmm0, [eax + eax*2 + 0x10]");
+}
+
+#[test]
+fn test_vex() {
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_avx(), bytes, text);
+        test_display_under(&InstDecoder::default(), bytes, text);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+    }
+
+    fn test_valid(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_avx(), bytes, text);
+        test_display_under(&InstDecoder::default(), bytes, text);
+        test_display_under(&InstDecoder::minimal(), bytes, text);
+    }
+
+    fn test_instr_invalid(bytes: &[u8]) {
+        test_invalid_under(&InstDecoder::minimal().with_avx(), bytes);
+        test_invalid_under(&InstDecoder::default(), bytes);
+    }
+
+    test_instr(&[0xc5, 0xf8, 0x10, 0x00], "vmovups xmm0, [eax]");
+    test_instr(&[0xc5, 0xf8, 0x10, 0x01], "vmovups xmm0, [ecx]");
+    test_valid(&[0xc5, 0x78, 0x10], "lds edi, [eax + 0x10]");
+    test_instr(&[0xc5, 0xf8, 0x10, 0xcf], "vmovups xmm1, xmm7");
+    test_instr(&[0xc5, 0xf9, 0x10, 0x0f], "vmovupd xmm1, [edi]");
+    test_instr(&[0xc5, 0xfa, 0x7e, 0x10], "vmovq xmm2, [eax]");
+    test_instr(&[0xc5, 0xfc, 0x10, 0x0f], "vmovups ymm1, [edi]");
+    test_instr(&[0xc5, 0xfd, 0x10, 0x0f], "vmovupd ymm1, [edi]");
+    test_instr(&[0xc5, 0xfe, 0x10, 0x0f], "vmovss xmm1, [edi]");
+    test_instr(&[0xc5, 0xff, 0x10, 0xcf], "vmovsd xmm1, xmm0, xmm7");
+    test_instr(&[0xc5, 0xff, 0x10, 0x01], "vmovsd xmm0, [ecx]");
+    test_instr(&[0xc5, 0xf9, 0x6e, 0xc6], "vmovd xmm0, esi");
+    test_instr(&[0xc5, 0xf9, 0x6e, 0x13], "vmovd xmm2, [ebx]");
+    test_instr(&[0xc5, 0xf9, 0x7e, 0xc6], "vmovd esi, xmm0");
+    test_instr(&[0xc5, 0xf9, 0x7e, 0x13], "vmovd [ebx], xmm2");
+    test_instr_invalid(&[0xf0, 0xc5, 0xf8, 0x10, 0x00]);
+    test_valid(&[0xc4, 0x02], "les eax, [edx]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x00, 0x0f], "vpshufb xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x00, 0x0f], "vpshufb ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x00, 0xcd], "vpshufb xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x00, 0xcd], "vpshufb ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x01, 0x0f], "vphaddw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x01, 0x0f], "vphaddw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x01, 0xcd], "vphaddw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x01, 0xcd], "vphaddw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x02, 0x0f], "vphaddd xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x02, 0x0f], "vphaddd ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x02, 0xcd], "vphaddd xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x02, 0xcd], "vphaddd ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x03, 0x0f], "vphaddsw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x03, 0x0f], "vphaddsw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x03, 0xcd], "vphaddsw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x03, 0xcd], "vphaddsw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x04, 0x0f], "vphaddubsw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x04, 0x0f], "vphaddubsw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x04, 0xcd], "vphaddubsw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x04, 0xcd], "vphaddubsw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x05, 0x0f], "vphsubw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x05, 0x0f], "vphsubw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x05, 0xcd], "vphsubw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x05, 0xcd], "vphsubw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x06, 0x0f], "vphsubd xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x06, 0x0f], "vphsubd ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x06, 0xcd], "vphsubd xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x06, 0xcd], "vphsubd ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x07, 0x0f], "vphsubsw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x07, 0x0f], "vphsubsw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x07, 0xcd], "vphsubsw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x07, 0xcd], "vphsubsw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x08, 0x0f], "vpsignb xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x08, 0x0f], "vpsignb ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x08, 0xcd], "vpsignb xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x08, 0xcd], "vpsignb ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x09, 0x0f], "vpsignw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x09, 0x0f], "vpsignw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x09, 0xcd], "vpsignw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x09, 0xcd], "vpsignw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0a, 0x0f], "vpsignd xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0a, 0x0f], "vpsignd ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0a, 0xcd], "vpsignd xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0a, 0xcd], "vpsignd ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0b, 0x0f], "vpmulhrsw xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0b, 0x0f], "vpmulhrsw ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0b, 0xcd], "vpmulhrsw xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0b, 0xcd], "vpmulhrsw ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0c, 0x0f], "vpermilps xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0c, 0x0f], "vpermilps ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0c, 0xcd], "vpermilps xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0c, 0xcd], "vpermilps ymm1, ymm1, ymm5");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0d, 0x0f], "vpermilpd xmm1, xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0d, 0x0f], "vpermilpd ymm1, ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x71, 0x0d, 0xcd], "vpermilpd xmm1, xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x75, 0x0d, 0xcd], "vpermilpd ymm1, ymm1, ymm5");
+    test_instr_invalid(&[0xc4, 0xc2, 0x71, 0x0e, 0x00]);
+    test_instr(&[0xc4, 0xc2, 0x79, 0x0e, 0x0f], "vtestps xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x7d, 0x0e, 0x0f], "vtestps ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x79, 0x0e, 0xcd], "vtestps xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x7d, 0x0e, 0xcd], "vtestps ymm1, ymm5");
+    test_instr_invalid(&[0xc4, 0xc2, 0x71, 0x0f, 0x00]);
+    test_instr(&[0xc4, 0xc2, 0x79, 0x0f, 0x0f], "vtestpd xmm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x7d, 0x0f, 0x0f], "vtestpd ymm1, [edi]");
+    test_instr(&[0xc4, 0xc2, 0x79, 0x0f, 0xcd], "vtestpd xmm1, xmm5");
+    test_instr(&[0xc4, 0xc2, 0x7d, 0x0f, 0xcd], "vtestpd ymm1, ymm5");
+    test_instr(&[0xc4, 0xe2, 0x65, 0x90, 0x04, 0x51], "vpgatherdd ymm0, [ecx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xe2, 0xe5, 0x90, 0x04, 0x51], "vpgatherdq ymm0, [ecx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xe2, 0x65, 0x91, 0x04, 0x51], "vpgatherqd ymm0, [ecx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xe2, 0xe5, 0x91, 0x04, 0x51], "vpgatherqq ymm0, [ecx + ymm2 * 2], ymm3");
+    test_instr(&[0xc4, 0xc2, 0x09, 0x9d, 0xcd], "vfnmadd132ss xmm1, xmm6, xmm5"); // note xmm6, not xmm14!
+    test_instr(&[0xc4, 0xc2, 0x89, 0x9d, 0xcd], "vfnmadd132sd xmm1, xmm6, xmm5");
+// ...
+    test_instr(&[0xc4, 0xe3, 0x79, 0x14, 0xd0, 0x0a], "vpextrb eax, xmm2, 0xa");
+    test_instr(&[0xc4, 0xe3, 0x79, 0x14, 0x10, 0x0a], "vpextrb [eax], xmm2, 0xa");
+    test_instr_invalid(&[0xc4, 0xe3, 0xf9, 0x14, 0x00, 0xd0]);
+    test_instr_invalid(&[0xc4, 0xe3, 0xf9, 0x14, 0x00, 0x0a]);
+}
+
+#[test]
+fn strange_prefixing() {
+    test_display(&[0x66, 0x0f, 0x21, 0xc8], "mov eax, dr1");
+    test_display(&[0xf2, 0x0f, 0x21, 0xc8], "mov eax, dr1");
+    test_display(&[0xf3, 0x0f, 0x21, 0xc8], "mov eax, dr1");
+}
+
+#[test]
+fn prefixed_0f() {
+    test_display(&[0x0f, 0x02, 0xc0], "lar eax, ax");
+    test_display(&[0x0f, 0x03, 0xc0], "lsl eax, eax");
+    test_display(&[0x0f, 0x05], "syscall");
+    test_display(&[0x66, 0x0f, 0x05], "syscall");
+    test_display(&[0x0f, 0x06], "clts");
+    test_display(&[0xf2, 0x0f, 0x06], "clts");
+    test_display(&[0x0f, 0x07], "sysret");
+    test_display(&[0xf2, 0x0f, 0x07], "sysret");
+    test_display(&[0x0f, 0x12, 0x0f], "movlps xmm1, [edi]");
+    test_display(&[0x0f, 0x12, 0xcf], "movhlps xmm1, xmm7");
+    test_display(&[0x0f, 0x16, 0x0f], "movhps xmm1, [edi]");
+    test_display(&[0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
+    test_display(&[0x0f, 0x12, 0xc0], "movhlps xmm0, xmm0");
+    test_invalid(&[0x0f, 0x13, 0xc0]);
+    test_display(&[0x0f, 0x13, 0x00], "movlps [eax], xmm0");
+    test_display(&[0x0f, 0x14, 0x08], "unpcklps xmm1, [eax]");
+    test_display(&[0x0f, 0x15, 0x08], "unpckhps xmm1, [eax]");
+    test_display(&[0x0f, 0x16, 0x0f], "movhps xmm1, [edi]");
+    test_display(&[0x0f, 0x16, 0xc0], "movlhps xmm0, xmm0");
+    test_invalid(&[0x0f, 0x17, 0xc0]);
+    test_display(&[0x0f, 0x17, 0x00], "movhps [eax], xmm0");
+    test_invalid(&[0x0f, 0x18, 0xc0]);
+    test_display(&[0x0f, 0x18, 0x00], "prefetchnta [eax]");
+    test_display(&[0x0f, 0x18, 0x08], "prefetch0 [eax]");
+    test_display(&[0x0f, 0x18, 0x10], "prefetch1 [eax]");
+    test_display(&[0x0f, 0x18, 0x18], "prefetch2 [eax]");
+    test_display(&[0x0f, 0x18, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x19, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x1a, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x1b, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x1c, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x1d, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x1e, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x1f, 0x20], "nop [eax]");
+    test_display(&[0x0f, 0x20, 0xc8], "mov eax, cr1");
+    test_display(&[0x0f, 0x21, 0xc8], "mov eax, dr1");
+    test_display(&[0x0f, 0x22, 0xc8], "mov cr1, eax");
+    test_display(&[0x0f, 0x22, 0xcf], "mov cr1, edi");
+    test_display(&[0x0f, 0x23, 0xc8], "mov dr1, eax");
+    test_display(&[0x0f, 0x23, 0xcf], "mov dr1, edi");
+    test_display(&[0x0f, 0x30], "wrmsr");
+    test_display(&[0x0f, 0x31], "rdtsc");
+    test_display(&[0x0f, 0x32], "rdmsr");
+    test_display(&[0x0f, 0x33], "rdpmc");
+    test_display(&[0x0f, 0x34], "sysenter");
+    test_display(&[0x0f, 0x35], "sysexit");
+    test_invalid(&[0x0f, 0x36]);
+    test_display(&[0x0f, 0x37], "getsec");
+    test_display(&[0x0f, 0x60, 0x00], "punpcklbw mm0, [eax]");
+    test_display(&[0x0f, 0x60, 0xc2], "punpcklbw mm0, mm2");
+    test_display(&[0x0f, 0x61, 0x00], "punpcklwd mm0, [eax]");
+    test_display(&[0x0f, 0x61, 0xc2], "punpcklwd mm0, mm2");
+    test_display(&[0x0f, 0x62, 0x00], "punpckldq mm0, [eax]");
+    test_display(&[0x0f, 0x62, 0xc2], "punpckldq mm0, mm2");
+    test_display(&[0x0f, 0x63, 0x00], "packsswb mm0, [eax]");
+    test_display(&[0x0f, 0x63, 0xc2], "packsswb mm0, mm2");
+    test_display(&[0x0f, 0x64, 0x00], "pcmpgtb mm0, [eax]");
+    test_display(&[0x0f, 0x64, 0xc2], "pcmpgtb mm0, mm2");
+    test_display(&[0x0f, 0x65, 0x00], "pcmpgtw mm0, [eax]");
+    test_display(&[0x0f, 0x65, 0xc2], "pcmpgtw mm0, mm2");
+    test_display(&[0x0f, 0x66, 0x00], "pcmpgtd mm0, [eax]");
+    test_display(&[0x0f, 0x66, 0xc2], "pcmpgtd mm0, mm2");
+    test_display(&[0x0f, 0x67, 0x00], "packuswb mm0, [eax]");
+    test_display(&[0x0f, 0x67, 0xc2], "packuswb mm0, mm2");
+    test_display(&[0x0f, 0x68, 0x00], "punpckhbw mm0, [eax]");
+    test_display(&[0x0f, 0x68, 0xc2], "punpckhbw mm0, mm2");
+    test_display(&[0x0f, 0x69, 0x00], "punpckhwd mm0, [eax]");
+    test_display(&[0x0f, 0x69, 0xc2], "punpckhwd mm0, mm2");
+    test_display(&[0x0f, 0x6a, 0x00], "punpckhdq mm0, [eax]");
+    test_display(&[0x0f, 0x6a, 0xc2], "punpckhdq mm0, mm2");
+    test_display(&[0x0f, 0x6b, 0x00], "packssdw mm0, [eax]");
+    test_display(&[0x0f, 0x6b, 0xc2], "packssdw mm0, mm2");
+    test_invalid(&[0x0f, 0x6c]);
+    test_invalid(&[0x0f, 0x6d]);
+    test_display(&[0x0f, 0x6e, 0x00], "movd mm0, [eax]");
+    test_display(&[0x0f, 0x6e, 0xc2], "movd mm0, edx");
+    test_display(&[0x0f, 0x6f, 0x00], "movq mm0, [eax]");
+    test_display(&[0x0f, 0x6f, 0xc2], "movq mm0, mm2");
+    test_display(&[0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [eax], 0x7f");
+    test_invalid(&[0x0f, 0x71, 0x00, 0x7f]);
+    test_invalid(&[0x0f, 0x71, 0xc0, 0x7f]);
+    test_display(&[0x0f, 0x71, 0xd0, 0x7f], "psrlw mm0, 0x7f");
+    test_display(&[0x0f, 0x71, 0xe0, 0x7f], "psraw mm0, 0x7f");
+    test_display(&[0x0f, 0x71, 0xf0, 0x7f], "psllw mm0, 0x7f");
+    test_invalid(&[0x0f, 0x72, 0x00, 0x7f]);
+    test_invalid(&[0x0f, 0x72, 0xc0, 0x7f]);
+    test_display(&[0x0f, 0x72, 0xd0, 0x7f], "psrld mm0, 0x7f");
+    test_display(&[0x0f, 0x72, 0xe0, 0x7f], "psrad mm0, 0x7f");
+    test_display(&[0x0f, 0x72, 0xf0, 0x7f], "pslld mm0, 0x7f");
+    test_invalid(&[0x0f, 0x73, 0x00, 0x7f]);
+    test_invalid(&[0x0f, 0x73, 0xc0, 0x7f]);
+    test_display(&[0x0f, 0x73, 0xd0, 0x7f], "psrlq mm0, 0x7f");
+    test_invalid(&[0x0f, 0x73, 0xe0, 0x7f]);
+    test_display(&[0x0f, 0x73, 0xf0, 0x7f], "psllq mm0, 0x7f");
+    test_display(&[0x0f, 0xa0], "push fs");
+    test_display(&[0x0f, 0xa1], "pop fs");
+    test_display(&[0x0f, 0xa2], "cpuid");
+    test_display(&[0x0f, 0xa4, 0xc0, 0x11], "shld eax, eax, 0x11");
+    test_display(&[0x66, 0x0f, 0xa4, 0xcf, 0x11], "shld di, cx, 0x11");
+    test_display(&[0x0f, 0xa5, 0xc0], "shld eax, eax, cl");
+    test_display(&[0x0f, 0xa5, 0xc9], "shld ecx, ecx, cl");
+}
+
+#[test]
+fn prefixed_660f() {
+    test_display(&[0x66, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
+    test_display(&[0xf2, 0x66, 0x66, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
+}
+
+#[test]
+fn prefixed_f20f() {
+    test_display(&[0xf2, 0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
+    test_display(&[0x66, 0xf2, 0x66, 0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
+}
+
+#[test]
+fn prefixed_f30f() {
+    test_display(&[0xf3, 0x0f, 0x16, 0xcf], "movshdup xmm1, xmm7");
+}
+
+#[test]
+fn only_32bit() {
+    test_display(&[0x40], "inc eax");
+    test_display(&[0x41], "inc ecx");
+    test_display(&[0x47], "inc edi");
+    test_display(&[0x48], "dec eax");
+    test_display(&[0x4f], "dec edi");
+
+    test_display(&[0xa0, 0xc0, 0xb0, 0xa0, 0x90], "mov al, [0x90a0b0c0]");
+    test_display(&[0x67, 0xa0, 0xc0, 0xb0], "mov al, [0xb0c0]");
+    test_display(&[0x67, 0xa1, 0xc0, 0xb0], "mov eax, [0xb0c0]");
+    test_display(&[0x66, 0x67, 0xa1, 0xc0, 0xb0], "mov ax, [0xb0c0]");
+}
diff --git a/test/protected_mode/regspec.rs b/test/protected_mode/regspec.rs
new file mode 100644
index 0000000..bd996cc
--- /dev/null
+++ b/test/protected_mode/regspec.rs
@@ -0,0 +1,20 @@
+use yaxpeax_x86::protected_mode::RegSpec;
+use std::collections::{BTreeMap, HashMap};
+
+#[test]
+fn test_ord() {
+    let _: BTreeMap<RegSpec, u64> = BTreeMap::new();
+}
+
+#[test]
+fn test_hash() {
+    let _: HashMap<RegSpec, u64> = HashMap::new();
+}
+
+#[test]
+fn test_labels() {
+    assert_eq!(RegSpec::eip().name(), "eip");
+    assert_eq!(RegSpec::ebp().name(), "ebp");
+    assert_eq!(RegSpec::gs().name(), "gs");
+    assert_eq!(RegSpec::al().name(), "al");
+}
diff --git a/test/regspec.rs b/test/regspec.rs
deleted file mode 100644
index 914c376..0000000
--- a/test/regspec.rs
+++ /dev/null
@@ -1,22 +0,0 @@
-use yaxpeax_x86::long_mode::RegSpec;
-use std::collections::{BTreeMap, HashMap};
-
-#[test]
-fn test_ord() {
-    let _: BTreeMap<RegSpec, u64> = BTreeMap::new();
-}
-
-#[test]
-fn test_hash() {
-    let _: HashMap<RegSpec, u64> = HashMap::new();
-}
-
-#[test]
-fn test_labels() {
-    assert_eq!(RegSpec::rip().name(), "rip");
-    assert_eq!(RegSpec::eip().name(), "eip");
-    assert_eq!(RegSpec::rflags().name(), "rflags");
-    assert_eq!(RegSpec::rbp().name(), "rbp");
-    assert_eq!(RegSpec::gs().name(), "gs");
-    assert_eq!(RegSpec::al().name(), "al");
-}
diff --git a/test/test.rs b/test/test.rs
index 2f7f779..f1a4129 100644
--- a/test/test.rs
+++ b/test/test.rs
@@ -1,1097 +1,5 @@
 extern crate yaxpeax_arch;
 extern crate yaxpeax_x86;
 
-mod regspec;
-
-use std::fmt::Write;
-
-use yaxpeax_arch::{Decoder, LengthedInstruction};
-use yaxpeax_x86::long_mode::{DecodeError, InstDecoder, Opcode};
-
-fn test_invalid(data: &[u8]) {
-    test_invalid_under(&InstDecoder::default(), data);
-}
-
-fn test_invalid_under(decoder: &InstDecoder, data: &[u8]) {
-    if let Ok(inst) = decoder.decode(data.into_iter().cloned()) {
-        assert_eq!(inst.opcode, Opcode::Invalid, "decoded {:?} from {:02x?} under decoder {}", inst.opcode, data, decoder);
-    } else {
-        // this is fine
-    }
-}
-
-fn test_display(data: &[u8], expected: &'static str) {
-    test_display_under(&InstDecoder::default(), data, expected);
-}
-
-fn test_display_under(decoder: &InstDecoder, data: &[u8], expected: &'static str) {
-    let mut hex = String::new();
-    for b in data {
-        write!(hex, "{:02x}", b).unwrap();
-    }
-    match decoder.decode(data.into_iter().map(|x| *x)) {
-        Ok(instr) => {
-            let text = format!("{}", instr);
-            assert!(
-                text == expected,
-                "display error for {}:\n  decoded: {:?} under decoder {}\n displayed: {}\n expected: {}\n",
-                hex,
-                instr,
-                decoder,
-                text,
-                expected
-            );
-            // while we're at it, test that the instruction is as long, and no longer, than its
-            // input
-            assert_eq!(instr.len() as usize, data.len(), "instruction length is incorrect, wanted instruction {}", expected);
-        },
-        Err(e) => {
-            assert!(false, "decode error ({}) for {} under decoder {}:\n  expected: {}\n", e, hex, decoder, expected);
-        }
-    }
-}
-
-#[test]
-fn test_mmx() {
-    test_display(&[0x4f, 0x0f, 0x7e, 0xcf], "movd r15, mm1");
-    test_display(&[0x41, 0x0f, 0x7e, 0xcf], "movd r15d, mm1");
-    test_display(&[0x4f, 0x0f, 0x7f, 0xcf], "movq mm7, mm1");
-    test_display(&[0x4f, 0x0f, 0x7f, 0x0f], "movq [r15], mm1");
-    test_display(&[0x0f, 0xc4, 0xc0, 0x14], "pinsrw mm0, eax, 0x14");
-    test_display(&[0x4f, 0x0f, 0xc4, 0xc0, 0x14], "pinsrw mm0, r8d, 0x14");
-    test_display(&[0x4f, 0x0f, 0xc4, 0x00, 0x14], "pinsrw mm0, [r8], 0x14");
-    test_display(&[0x4f, 0x0f, 0xd1, 0xcf], "psrlw mm1, mm7");
-    test_display(&[0x4f, 0x0f, 0xd1, 0x00], "psrlw mm0, [r8]");
-    test_invalid(&[0x4f, 0x0f, 0xd7, 0x00]);
-    test_display(&[0x4f, 0x0f, 0xd7, 0xcf], "pmovmskb r9d, mm7");
-}
-
-#[test]
-fn test_cvt() {
-    test_display(&[0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm7");
-    test_display(&[0x48, 0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm7");
-    test_display(&[0x4f, 0x0f, 0x2c, 0xcf], "cvttps2pi mm1, xmm15");
-    test_display(&[0x4f, 0x0f, 0x2a, 0xcf], "cvtpi2ps xmm9, mm7");
-    test_display(&[0x4f, 0x0f, 0x2a, 0x00], "cvtpi2ps xmm8, [r8]");
-    test_display(&[0x4f, 0x66, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm1, mm7");
-    test_display(&[0x66, 0x4f, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm9, mm7");
-    test_display(&[0x4f, 0xf3, 0x0f, 0x2a, 0xcf], "cvtsi2ss xmm1, edi");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0xcf], "cvtsi2ss xmm9, r15");
-    test_display(&[0x4f, 0xf2, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm1, edi");
-    test_display(&[0xf2, 0x4f, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm9, r15");
-    test_display(&[0x4f, 0xf2, 0x0f, 0x2a, 0x00], "cvtsi2sd xmm0, [rax]");
-    test_display(&[0xf2, 0x4f, 0x0f, 0x2a, 0x00], "cvtsi2sd xmm8, [r8]");
-    test_display(&[0x4f, 0xf3, 0x0f, 0x2a, 0x00], "cvtsi2ss xmm0, [rax]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0x00], "cvtsi2ss xmm8, [r8]");
-    test_display(&[0x4f, 0x66, 0x0f, 0x2a, 0x00], "cvtpi2pd xmm0, [rax]");
-    test_display(&[0x66, 0x4f, 0x0f, 0x2a, 0x00], "cvtpi2pd xmm8, [r8]");
-}
-
-#[test]
-fn test_aesni() {
-    fn test_instr(bytes: &[u8], text: &'static str) {
-        test_display_under(&InstDecoder::minimal().with_aesni(), bytes, text);
-        test_display_under(&InstDecoder::default(), bytes, text);
-        test_invalid_under(&InstDecoder::minimal(), bytes);
-    }
-
-    test_instr(&[0x66, 0x0f, 0x38, 0xdb, 0x0f], "aesimc xmm1, [rdi]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdb, 0xcf], "aesimc xmm9, xmm15");
-
-    test_instr(&[0x66, 0x0f, 0x38, 0xdc, 0x0f], "aesenc xmm1, [rdi]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdc, 0xcf], "aesenc xmm9, xmm15");
-
-    test_instr(&[0x66, 0x0f, 0x38, 0xdd, 0x0f], "aesenclast xmm1, [rdi]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdd, 0xcf], "aesenclast xmm9, xmm15");
-
-    test_instr(&[0x66, 0x0f, 0x38, 0xde, 0x0f], "aesdec xmm1, [rdi]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xde, 0xcf], "aesdec xmm9, xmm15");
-
-    test_instr(&[0x66, 0x0f, 0x38, 0xdf, 0x0f], "aesdeclast xmm1, [rdi]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x38, 0xdf, 0xcf], "aesdeclast xmm9, xmm15");
-
-    test_instr(&[0x66, 0x0f, 0x3a, 0xdf, 0x0f, 0xaa], "aeskeygenassist xmm1, [rdi], 0xaa");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x3a, 0xdf, 0xcf, 0xaa], "aeskeygenassist xmm9, xmm15, 0xaa");
-}
-
-#[test]
-fn test_sse2() {
-    fn test_instr(bytes: &[u8], text: &'static str) {
-        // sse and sse2 are part of amd64, so x86_64, meaning even the minimal decoder must support
-        // them.
-        test_display_under(&InstDecoder::minimal(), bytes, text);
-    }
-
-    test_instr(&[0xf2, 0x0f, 0x10, 0x0c, 0xc7], "movsd xmm1, [rdi + rax * 8]");
-    test_instr(&[0xf2, 0x0f, 0x11, 0x0c, 0xc7], "movsd [rdi + rax * 8], xmm1");
-    test_instr(&[0x66, 0x0f, 0x11, 0x0c, 0xc7], "movupd [rdi + rax * 8], xmm1");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x12, 0xc3], "movhlps xmm8, xmm11"); // reg-reg form is movhlps
-    test_instr(&[0x66, 0x4f, 0x0f, 0x12, 0x03], "movlpd xmm8, [r11]"); // reg-mem is movlpd
-    test_instr(&[0x66, 0x4f, 0x0f, 0x13, 0x03], "movlpd [r11], xmm8");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x13, 0xc3]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x14, 0x03], "unpcklpd xmm8, [r11]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x14, 0xc3], "unpcklpd xmm8, xmm11");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x15, 0x03], "unpckhpd xmm8, [r11]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x15, 0xc3], "unpckhpd xmm8, xmm11");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x16, 0x03], "movhpd xmm8, [r11]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x16, 0xc3], "movlhps xmm8, xmm11");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x17, 0x03], "movhpd [r11], xmm8");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x17, 0xc3]);
-
-    test_instr(&[0x66, 0x4f, 0x0f, 0x28, 0xd0], "movapd xmm10, xmm8");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8]");
-
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2a, 0xcf], "cvtpi2pd xmm9, mm7");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2a, 0x0f], "cvtpi2pd xmm9, [r15]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x2a, 0xcf], "cvtsi2sd xmm9, r15");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x2a, 0x0f], "cvtsi2sd xmm9, [r15]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2b, 0x0f], "movntpd [r15], xmm9");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2c, 0xcf], "cvttpd2pi mm1, xmm15");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2c, 0x0f], "cvttpd2pi mm1, [r15]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x2c, 0xcf], "cvttsd2si xmm9, xmm15");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x2c, 0x0f], "cvttsd2si xmm9, [r15]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2d, 0xcf], "cvtpd2pi mm1, xmm15");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2d, 0x0f], "cvtpd2pi mm1, [r15]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x2d, 0xcf], "cvtsd2si xmm9, xmm15");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x2d, 0x0f], "cvtsd2si xmm9, [r15]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2e, 0xcf], "ucomisd xmm9, xmm15");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2e, 0x0f], "ucomisd xmm9, [r15]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2f, 0xcf], "comisd xmm9, xmm15");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x2f, 0x0f], "comisd xmm9, [r15]");
-
-    /*
-     * .... 660f38
-     * .... 660f7f
-     */
-
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x50, 0x01]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x50, 0xc1], "movmskpd r8d, xmm9");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x51, 0x01], "sqrtpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x51, 0x01], "sqrtsd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x52, 0x01], "rsqrtps xmm8, [r9]"); // note: NOT "rsqrtpd" - no such instruction exists, so fall back to just 0f52 parse.
-    test_instr(&[0x66, 0x4f, 0x0f, 0x53, 0x01], "rcpps xmm8, [r9]"); // note: NOT "rcppd" - no such instruction exists, so fall back to just 0f53 parse.
-    test_instr(&[0x66, 0x4f, 0x0f, 0x54, 0x01], "andpd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x55, 0x01], "andnpd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x56, 0x01], "orpd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x57, 0x01], "xorpd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x58, 0x01], "addpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x58, 0x01], "addsd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x59, 0x01], "mulpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x59, 0x01], "mulsd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x5a, 0x01], "cvtpd2ps xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x5a, 0x01], "cvtsd2ss xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x5b, 0x01], "cvtps2dq xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x5c, 0x01], "subpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x5c, 0x01], "subsd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x5d, 0x01], "minpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x5d, 0x01], "minsd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x5e, 0x01], "divpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x5e, 0x01], "divsd xmm8, [r9]");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x5f, 0x01], "maxpd xmm8, [r9]");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x5f, 0x01], "maxsd xmm8, [r9]");
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x60, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpcklbw xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x61, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpcklwd xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x62, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpckldq xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x63, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "packsswb xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x64, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "pcmpgtb xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x65, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "pcmpgtw xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x66, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "pcmpgtd xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x67, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "packuswb xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x68, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpckhbw xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x69, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpckhwd xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x6a, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpckhdq xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x6b, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "packssdw xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x6c, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpcklqdq xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x6d, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "punpckhqdq xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    // TODO: this needs to be clear that the operand is `dword`
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x6e, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "movq xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x6f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "movdqa xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-
-    test_instr(&[0x66, 0x48, 0x0f, 0x6e, 0xc0], "movq xmm0, rax");
-    test_instr(&[0x66, 0x0f, 0x70, 0xc0, 0x4e], "pshufd xmm0, xmm0, 0x4e");
-    test_instr(&[0xf2, 0x0f, 0x70, 0xc0, 0x4e], "pshuflw xmm0, xmm0, 0x4e");
-    test_instr(&[0xf3, 0x0f, 0x70, 0xc0, 0x4e], "pshufhw xmm0, xmm0, 0x4e");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x71, 0x10, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x71, 0xd0, 0x8f], "psrlw xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x71, 0x20, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x71, 0xe0, 0x8f], "psraw xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x71, 0x30, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x71, 0xf0, 0x8f], "psllw xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x72, 0x10, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x72, 0xd0, 0x8f], "psrld xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x72, 0x20, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x72, 0xe0, 0x8f], "psrad xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x72, 0x30, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x72, 0xf0, 0x8f], "pslld xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x10, 0x8f]);
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x18, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xd0, 0x8f], "psrlq xmm0, 0x8f");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xd8, 0x8f], "psrldq xmm0, 0x8f");
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x30, 0x8f]);
-    test_invalid(&[0x66, 0x4f, 0x0f, 0x73, 0x38, 0x8f]);
-    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xf0, 0x8f], "psllq xmm0, 0x8f");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x73, 0xf8, 0x8f], "pslldq xmm0, 0x8f");
-    test_instr(&[0x66, 0x0f, 0x7e, 0xc1], "movd ecx, xmm0");
-    test_instr(&[0x66, 0x48, 0x0f, 0x7e, 0xc1], "movq rcx, xmm0");
-    test_instr(&[0x66, 0x48, 0x0f, 0x7e, 0x01], "movd [rcx], xmm0");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x7e, 0xc1], "movq r9, xmm8");
-    test_instr(
-        &[0x66, 0x4f, 0x0f, 0x7f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "movdqa [r12 + r11 * 4 - 0x334455cc], xmm11"
-    );
-
-    test_instr(&[0x66, 0x0f, 0xc2, 0xc3, 0x08], "cmppd xmm0, xmm3, 0x8");
-    test_instr(&[0x66, 0x4f, 0x0f, 0xc2, 0xc3, 0x08], "cmppd xmm8, xmm11, 0x8");
-    test_instr(&[0x66, 0x4f, 0x0f, 0xc2, 0x03, 0x08], "cmppd xmm8, [r11], 0x8");
-    test_instr(&[0xf2, 0x0f, 0xc2, 0xc3, 0x08], "cmpsd xmm0, xmm3, 0x8");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0xc2, 0xc3, 0x08], "cmpsd xmm8, xmm11, 0x8");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0xc2, 0x03, 0x08], "cmpsd xmm8, [r11], 0x8");
-
-    test_instr(&[0x66, 0x0f, 0xc4, 0xc3, 0x08], "pinsrw xmm0, ebx, 0x8");
-    test_instr(&[0x66, 0x4f, 0x0f, 0xc4, 0xc3, 0x08], "pinsrw xmm8, r11d, 0x8");
-
-    test_instr(&[0x66, 0x0f, 0xc4, 0x03, 0x08], "pinsrw xmm0, [rbx], 0x8");
-    test_instr(&[0x66, 0x4f, 0x0f, 0xc4, 0x03, 0x08], "pinsrw xmm8, [r11], 0x8");
-
-//    test_instr(&[0x66, 0x0f, 0xc5, 0xc3, 0x08], "pextrw eax, xmm3, 0x8");
-//    test_instr(&[0x66, 0x4f, 0x0f, 0xc5, 0xc3, 0x08], "pextrw r8d, xmm11, 0x8");
-//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x03, 0x08]);
-//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x40, 0x08]);
-//    test_instr_invalid(&[0x66, 0x0f, 0xc5, 0x80, 0x08]);
-
-    test_instr(&[0x66, 0x4f, 0x0f, 0xc6, 0x03, 0x08], "shufpd xmm8, [r11], 0x8");
-    test_instr(&[0x66, 0x0f, 0xc6, 0x03, 0x08], "shufpd xmm0, [rbx], 0x8");
-    test_instr(&[0x66, 0x0f, 0xc6, 0xc3, 0x08], "shufpd xmm0, xmm3, 0x8");
-    test_instr(&[0x66, 0x0f, 0xd1, 0xc1], "psrlw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd1, 0x01], "psrlw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xd2, 0xc1], "psrld xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd2, 0x01], "psrld xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xd3, 0xc1], "psrlq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd3, 0x01], "psrlq xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xd4, 0xc1], "paddq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd4, 0x01], "paddq xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xd5, 0xc1], "pmullw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd5, 0x01], "pmullw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xd6, 0xc1], "movq xmm1, xmm0");
-    test_instr(&[0x66, 0x0f, 0xd6, 0x01], "movq [rcx], xmm0");
-    test_invalid(&[0xf3, 0x4f, 0x0f, 0xd6, 0x03]);
-    test_instr(&[0xf3, 0x4f, 0x0f, 0xd6, 0xc3], "movq2dq xmm8, mm3");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0xd6, 0xc3], "movdq2q mm0, xmm11");
-    test_instr(&[0x66, 0x0f, 0xd7, 0xc1], "pmovmskb eax, xmm1");
-    test_instr(&[0x66, 0x4f, 0x0f, 0xd7, 0xc1], "pmovmskb r8d, xmm9");
-    test_invalid(&[0x66, 0x0f, 0xd7, 0x01]);
-    test_instr(&[0x66, 0x0f, 0xd8, 0xc1], "psubusb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd8, 0x01], "psubusb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xd9, 0xc1], "psubusw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xd9, 0x01], "psubusw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xda, 0xc1], "pminub xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xda, 0x01], "pminub xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xdb, 0xc1], "pand xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xdb, 0x01], "pand xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xdc, 0xc1], "paddusb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xdc, 0x01], "paddusb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xdd, 0xc1], "paddusw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xdd, 0x01], "paddusw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xde, 0xc1], "pmaxub xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xde, 0x01], "pmaxub xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xdf, 0xc1], "pandn xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xdf, 0x01], "pandn xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe0, 0xc1], "pavgb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe0, 0x01], "pavgb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe1, 0xc1], "psraw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe1, 0x01], "psraw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe2, 0xc1], "psrad xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe2, 0x01], "psrad xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe3, 0xc1], "pavgw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe3, 0x01], "pavgw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe4, 0xc1], "pmulhuw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe4, 0x01], "pmulhuw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe5, 0xc1], "pmulhw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe5, 0x01], "pmulhw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe6, 0xc1], "cvttpd2dq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe6, 0x01], "cvttpd2dq xmm0, [rcx]");
-    test_invalid(&[0x66, 0x0f, 0xe7, 0xc1]);
-    test_instr(&[0x66, 0x0f, 0xe7, 0x01], "movntdq [rcx], xmm0");
-    test_instr(&[0x66, 0x0f, 0xe8, 0xc1], "psubsb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe8, 0x01], "psubsb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xe9, 0xc1], "psubsw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xe9, 0x01], "psubsw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xea, 0xc1], "pminsw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xea, 0x01], "pminsw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xeb, 0xc3], "por xmm0, xmm3");
-    test_instr(&[0x66, 0x0f, 0xeb, 0xc4], "por xmm0, xmm4");
-    test_instr(&[0x66, 0x0f, 0xeb, 0xd3], "por xmm2, xmm3");
-    test_instr(&[0x66, 0x0f, 0xeb, 0x12], "por xmm2, [rdx]");
-    test_instr(&[0x66, 0x0f, 0xeb, 0xc1], "por xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xeb, 0x01], "por xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xec, 0xc1], "paddsb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xec, 0x01], "paddsb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xed, 0xc1], "paddsw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xed, 0x01], "paddsw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xee, 0xc1], "pmaxsw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xee, 0x01], "pmaxsw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xef, 0xc1], "pxor xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xef, 0x01], "pxor xmm0, [rcx]");
-    test_invalid(&[0x66, 0x0f, 0xf0, 0xc1]);
-    test_invalid(&[0x66, 0x0f, 0xf0, 0x01]);
-    test_instr(&[0x66, 0x0f, 0xf1, 0xc1], "psllw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf1, 0x01], "psllw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf2, 0xc1], "pslld xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf2, 0x01], "pslld xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf3, 0xc1], "psllq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf3, 0x01], "psllq xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf4, 0xc1], "pmuludq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf4, 0x01], "pmuludq xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf5, 0xc1], "pmaddwd xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf5, 0x01], "pmaddwd xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf6, 0xc1], "psadbw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf6, 0x01], "psadbw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf7, 0xc1], "maskmovdqu xmm0, xmm1");
-    test_invalid(&[0x66, 0x0f, 0xf7, 0x01]);
-    test_instr(&[0x66, 0x0f, 0xf8, 0xc1], "psubb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf8, 0x01], "psubb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xf9, 0xc1], "psubw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xf9, 0x01], "psubw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xfa, 0xc1], "psubd xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xfa, 0x01], "psubd xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xfb, 0xc1], "psubq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xfb, 0x01], "psubq xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xfc, 0xc1], "paddb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xfc, 0x01], "paddb xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xfd, 0xc1], "paddw xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xfd, 0x01], "paddw xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xfe, 0xc1], "paddd xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xfe, 0x01], "paddd xmm0, [rcx]");
-    test_instr(&[0x66, 0x0f, 0xff, 0xc1], "paddq xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0xff, 0x01], "paddq xmm0, [rcx]");
-
-    test_instr(&[0x66, 0x0f, 0x74, 0xc1], "pcmpeqb xmm0, xmm1");
-    test_instr(&[0x66, 0x0f, 0x74, 0x12], "pcmpeqb xmm2, [rdx]");
-    test_instr(&[0x66, 0x0f, 0xf8, 0xc8], "psubb xmm1, xmm0");
-    test_instr(&[0x66, 0x0f, 0xf8, 0xd0], "psubb xmm2, xmm0");
-    test_instr(&[0x66, 0x0f, 0xf8, 0x12], "psubb xmm2, [rdx]");
-}
-
-#[test]
-fn test_sse3() {
-    fn test_instr(bytes: &[u8], text: &'static str) {
-        test_display_under(&InstDecoder::minimal().with_sse3(), 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);
-    }
-
-    fn test_instr_invalid(bytes: &[u8]) {
-        test_invalid_under(&InstDecoder::minimal().with_sse3(), bytes);
-        test_invalid_under(&InstDecoder::default(), bytes);
-    }
-    test_instr(&[0xf2, 0x0f, 0xf0, 0x0f], "lddqu xmm1, [rdi]");
-    test_instr_invalid(&[0xf2, 0x0f, 0xf0, 0xcf]);
-    test_instr(&[0xf2, 0x0f, 0xd0, 0x0f], "addsubps xmm1, [rdi]");
-    test_instr(&[0xf2, 0x0f, 0xd0, 0xcf], "addsubps xmm1, xmm7");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0xd0, 0xcf], "addsubps xmm9, xmm15");
-    test_instr(&[0x66, 0x0f, 0xd0, 0x0f], "addsubpd xmm1, [rdi]");
-    test_instr(&[0x66, 0x0f, 0xd0, 0xcf], "addsubpd xmm1, xmm7");
-    test_instr(&[0x66, 0x4f, 0x0f, 0xd0, 0xcf], "addsubpd xmm9, xmm15");
-
-    test_instr(&[0xf2, 0x0f, 0x7c, 0x0f], "haddps xmm1, [rdi]");
-    test_instr(&[0xf2, 0x0f, 0x7c, 0xcf], "haddps xmm1, xmm7");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x7c, 0xcf], "haddps xmm9, xmm15");
-    test_instr(&[0x66, 0x0f, 0x7c, 0x0f], "haddpd xmm1, [rdi]");
-    test_instr(&[0x66, 0x0f, 0x7c, 0xcf], "haddpd xmm1, xmm7");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x7c, 0xcf], "haddpd xmm9, xmm15");
-
-    test_instr(&[0xf2, 0x0f, 0x7d, 0x0f], "hsubps xmm1, [rdi]");
-    test_instr(&[0xf2, 0x0f, 0x7d, 0xcf], "hsubps xmm1, xmm7");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x7d, 0xcf], "hsubps xmm9, xmm15");
-    test_instr(&[0x66, 0x0f, 0x7d, 0x0f], "hsubpd xmm1, [rdi]");
-    test_instr(&[0x66, 0x0f, 0x7d, 0xcf], "hsubpd xmm1, xmm7");
-    test_instr(&[0x66, 0x4f, 0x0f, 0x7d, 0xcf], "hsubpd xmm9, xmm15");
-
-    test_instr(&[0xf3, 0x0f, 0x12, 0x0f], "movsldup xmm1, [rdi]");
-    test_instr(&[0xf3, 0x0f, 0x12, 0xcf], "movsldup xmm1, xmm7");
-    test_instr(&[0xf3, 0x4f, 0x0f, 0x12, 0xcf], "movsldup xmm9, xmm15");
-    test_instr(&[0xf3, 0x0f, 0x16, 0x0f], "movshdup xmm1, [rdi]");
-    test_instr(&[0xf3, 0x0f, 0x16, 0xcf], "movshdup xmm1, xmm7");
-    test_instr(&[0xf3, 0x4f, 0x0f, 0x16, 0xcf], "movshdup xmm9, xmm15");
-
-    test_instr(&[0xf2, 0x0f, 0x12, 0x0f], "movddup xmm1, [rdi]");
-    test_instr(&[0xf2, 0x0f, 0x12, 0xcf], "movddup xmm1, xmm7");
-    test_instr(&[0xf2, 0x4f, 0x0f, 0x12, 0xcf], "movddup xmm9, xmm15");
-
-    test_instr(&[0x66, 0x0f, 0x01, 0xc8], "monitor");
-    test_instr(&[0xf2, 0x0f, 0x01, 0xc8], "monitor");
-    test_instr(&[0xf3, 0x0f, 0x01, 0xc8], "monitor");
-
-    test_instr(&[0x66, 0x0f, 0x01, 0xc9], "mwait");
-    test_instr(&[0xf2, 0x0f, 0x01, 0xc9], "mwait");
-    test_instr(&[0xf3, 0x0f, 0x01, 0xc9], "mwait");
-}
-
-#[test]
-fn test_0f01() {
-    // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
-    // Number"
-    test_display(&[0x0f, 0x01, 0x38], "invlpg [rax]");
-    test_display(&[0x0f, 0x01, 0x3f], "invlpg [rdi]");
-    test_display(&[0x0f, 0x01, 0x40, 0xff], "sgdt [rax - 0x1]");
-    test_display(&[0x0f, 0x01, 0x41, 0xff], "sgdt [rcx - 0x1]");
-    test_display(&[0x0f, 0x01, 0x49, 0xff], "sidt [rcx - 0x1]");
-    test_display(&[0x0f, 0x01, 0x51, 0xff], "lgdt [rcx - 0x1]");
-    test_display(&[0x0f, 0x01, 0x59, 0xff], "lidt [rcx - 0x1]");
-    test_display(&[0x0f, 0x01, 0x61, 0xff], "smsw [rcx - 0x1]");
-    test_display(&[0x0f, 0x01, 0xc0], "enclv");
-    test_display(&[0x0f, 0x01, 0xc1], "vmcall");
-    test_display(&[0x0f, 0x01, 0xc2], "vmlaunch");
-    test_display(&[0x0f, 0x01, 0xc3], "vmresume");
-    test_display(&[0x0f, 0x01, 0xc4], "vmxoff");
-    test_invalid(&[0x0f, 0x01, 0xc5]);
-    test_invalid(&[0x0f, 0x01, 0xc6]);
-    test_invalid(&[0x0f, 0x01, 0xc7]);
-    test_display(&[0x0f, 0x01, 0xc8], "monitor");
-    test_display(&[0x0f, 0x01, 0xc9], "mwait");
-    test_display(&[0x0f, 0x01, 0xca], "clac");
-    test_display(&[0x0f, 0x01, 0xcb], "stac");
-    test_display(&[0x0f, 0x01, 0xcf], "encls");
-    test_display(&[0x0f, 0x01, 0xd0], "xgetbv");
-    test_display(&[0x0f, 0x01, 0xd1], "xsetbv");
-    test_invalid(&[0x0f, 0x01, 0xd2]);
-    test_invalid(&[0x0f, 0x01, 0xd3]);
-    test_display(&[0x0f, 0x01, 0xd4], "vmfunc");
-    test_display(&[0x0f, 0x01, 0xd5], "xend");
-    test_display(&[0x0f, 0x01, 0xd6], "xtest");
-    test_display(&[0x0f, 0x01, 0xd7], "enclu");
-    test_invalid(&[0x0f, 0x01, 0xd8]);
-    test_invalid(&[0x0f, 0x01, 0xd9]);
-    test_invalid(&[0x0f, 0x01, 0xda]);
-    test_invalid(&[0x0f, 0x01, 0xdb]);
-    test_invalid(&[0x0f, 0x01, 0xdc]);
-    test_invalid(&[0x0f, 0x01, 0xdd]);
-    test_invalid(&[0x0f, 0x01, 0xde]);
-    test_invalid(&[0x0f, 0x01, 0xdf]);
-    test_display(&[0x0f, 0x01, 0xee], "rdpkru");
-    test_display(&[0x0f, 0x01, 0xef], "wrpkru");
-    test_display(&[0x0f, 0x01, 0xf8], "swapgs");
-    test_display(&[0x0f, 0x01, 0xf9], "rdtscp");
-}
-
-#[test]
-fn test_0fae() {
-    let intel = InstDecoder::minimal().with_intel_quirks();
-    let amd = InstDecoder::minimal().with_amd_quirks();
-    let default = InstDecoder::default();
-    let minimal = InstDecoder::minimal();
-    // drawn heavily from "Table A-6.  Opcode Extensions for One- and Two-byte Opcodes by Group
-    // Number"
-    test_display(&[0x0f, 0xae, 0x04, 0x4f], "fxsave [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x0c, 0x4f], "fxrstor [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x14, 0x4f], "ldmxcsr [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x1c, 0x4f], "stmxcsr [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x24, 0x4f], "xsave [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x2c, 0x4f], "xrstor [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x34, 0x4f], "xsaveopt [rdi + rcx * 2]");
-    test_display(&[0x0f, 0xae, 0x3c, 0x4f], "clflush [rdi + rcx * 2]");
-
-    for (modrm, text) in &[(0xe8u8, "lfence"), (0xf0u8, "mfence"), (0xf8u8, "sfence")] {
-        test_display_under(&intel, &[0x0f, 0xae, *modrm], text);
-        test_display_under(&amd, &[0x0f, 0xae, *modrm], text);
-        test_display_under(&default, &[0x0f, 0xae, *modrm], text);
-        test_display_under(&minimal, &[0x0f, 0xae, *modrm], text);
-        // it turns out intel and amd accept m != 0 for {l,m,s}fence:
-        // from intel:
-        // ```
-        // Specification of the instruction's opcode above indicates a ModR/M byte of F0. For this
-        // instruction, the processor ignores the r/m field of the ModR/M byte. Thus, MFENCE is encoded
-        // by any opcode of the form 0F AE Fx, where x is in the range 0-7.
-        // ```
-        // whereas amd does not discuss the r/m field at all. at least as of zen, amd also accepts
-        // these encodings.
-        for m in 1u8..8u8 {
-            test_display_under(&intel, &[0x0f, 0xae, modrm | m], text);
-            test_display_under(&amd, &[0x0f, 0xae, modrm | m], text);
-            test_display_under(&default, &[0x0f, 0xae, modrm | m], text);
-            test_invalid_under(&minimal, &[0x0f, 0xae, modrm | m]);
-        }
-    }
-}
-
-#[test]
-fn test_system() {
-    test_display(&[0x66, 0x4f, 0x0f, 0xb2, 0x00], "lss r8, [r8]");
-    test_display(&[0x67, 0x4f, 0x0f, 0xb2, 0x00], "lss r8, [r8d]");
-    test_display(&[0x4f, 0x0f, 0xb2, 0x00], "lss r8, [r8]");
-    test_display(&[0x45, 0x0f, 0x22, 0xc8], "mov cr9, r8");
-    test_display(&[0x45, 0x0f, 0x20, 0xc8], "mov r8, cr9");
-    test_display(&[0x40, 0x0f, 0x22, 0xc8], "mov cr1, rax");
-    test_display(&[0x0f, 0x22, 0xc8], "mov cr1, rax");
-    test_display(&[0x44, 0x0f, 0x22, 0xcf], "mov cr9, rdi");
-    test_display(&[0x0f, 0x22, 0xcf], "mov cr1, rdi");
-    test_display(&[0x0f, 0x20, 0xc8], "mov rax, cr1");
-
-    test_display(&[0x45, 0x0f, 0x23, 0xc8], "mov dr9, r8");
-    test_display(&[0x45, 0x0f, 0x21, 0xc8], "mov r8, dr9");
-    test_display(&[0x40, 0x0f, 0x23, 0xc8], "mov dr1, rax");
-    test_display(&[0x0f, 0x23, 0xc8], "mov dr1, rax");
-    test_display(&[0x0f, 0x21, 0xc8], "mov rax, dr1");
-}
-
-#[test]
-fn test_arithmetic() {
-    test_display(&[0x81, 0xec, 0x10, 0x03, 0x00, 0x00], "sub esp, 0x310");
-    test_display(&[0x0f, 0xaf, 0xc2], "imul eax, edx");
-    test_display(&[0x4b, 0x69, 0x43, 0x6f, 0x6d, 0x70, 0x6c, 0x65], "imul rax, [r11 + 0x6f], 0x656c706d");
-    test_display(&[0x66, 0x0f, 0xaf, 0xd1], "imul dx, cx");
-    test_display(&[0x4b, 0x6b, 0x43, 0x6f, 0x6d], "imul al, [r11 + 0x6f], 0x6d");
-    test_display(&[0x4f, 0x4e, 0x00, 0xcc], "add spl, r9b");
-}
-
-#[test]
-#[allow(non_snake_case)]
-fn test_E_decode() {
-    test_display(&[0xff, 0x75, 0xb8], "push [rbp - 0x48]");
-    test_display(&[0xff, 0x75, 0x08], "push [rbp + 0x8]");
-}
-
-#[test]
-fn test_sse() {
-    test_display(&[0xf3, 0x0f, 0x10, 0x0c, 0xc7], "movss xmm1, [rdi + rax * 8]");
-    test_display(&[0xf3, 0x0f, 0x11, 0x0c, 0xc7], "movss [rdi + rax * 8], xmm1");
-    test_display(&[0x4f, 0x0f, 0x28, 0x00], "movaps xmm8, [r8]");
-    test_display(&[0x4f, 0x0f, 0x29, 0x00], "movaps [r8], xmm8");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0xc1], "cvtsi2ss xmm8, r9");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2a, 0x01], "cvtsi2ss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x2b, 0x00], "movntps [r8], xmm8");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2c, 0xc1], "cvttss2si r8, xmm9");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2c, 0x01], "cvttss2si r8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2d, 0xc1], "cvtss2si r8, xmm9");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x2d, 0x01], "cvtss2si r8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x2e, 0x00], "ucomiss xmm8, [r8]");
-    test_display(&[0x4f, 0x0f, 0x2f, 0x00], "comiss xmm8, [r8]");
-    test_display(&[0x0f, 0x28, 0xd0], "movaps xmm2, xmm0");
-    test_display(&[0x66, 0x0f, 0x28, 0xd0], "movapd xmm2, xmm0");
-    test_display(&[0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [rax]");
-    test_display(&[0x4f, 0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [rax]");
-    test_display(&[0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8]");
-    test_display(&[0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8]");
-    test_display(&[0x67, 0x4f, 0x66, 0x0f, 0x28, 0x00], "movapd xmm0, [eax]");
-    test_display(&[0x67, 0x66, 0x4f, 0x0f, 0x28, 0x00], "movapd xmm8, [r8d]");
-    test_display(&[0x66, 0x0f, 0x29, 0x00], "movapd [rax], xmm0");
-    test_invalid(&[0x4f, 0x0f, 0x50, 0x00]);
-    test_display(&[0x4f, 0x0f, 0x50, 0xc1], "movmskps r8d, xmm9");
-    test_display(&[0x4f, 0x0f, 0x51, 0x01], "sqrtps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x51, 0x01], "sqrtss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x52, 0x01], "rsqrtps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x52, 0x01], "rsqrtss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x53, 0x01], "rcpps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x53, 0x01], "rcpss xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x53, 0xc1], "rcpss xmm8, xmm9");
-    test_display(&[0x4f, 0x0f, 0x54, 0x01], "andps xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x55, 0x01], "andnps xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x56, 0x01], "orps xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x57, 0x01], "xorps xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x58, 0x01], "addps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x58, 0x01], "addss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x59, 0x01], "mulps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x59, 0x01], "mulss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x5a, 0x01], "cvtps2pd xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x5a, 0x01], "cvtss2sd xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x5b, 0x01], "cvtdq2ps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x5b, 0x01], "cvttps2dq xmm8, [r9]");
-    test_display(&[0x67, 0x4f, 0x0f, 0x5b, 0x01], "cvtdq2ps xmm8, [r9d]");
-    test_display(&[0x4f, 0x0f, 0x5c, 0x01], "subps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x5c, 0x01], "subss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x5d, 0x01], "minps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x5d, 0x01], "minss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x5e, 0x01], "divps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x5e, 0x01], "divss xmm8, [r9]");
-    test_display(&[0x4f, 0x0f, 0x5f, 0x01], "maxps xmm8, [r9]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x5f, 0x01], "maxss xmm8, [r9]");
-
-    test_display(&[0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [rax], 0x7f");
-    test_display(&[0x4f, 0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [r8], 0x7f");
-
-    test_display(&[0x66, 0x0f, 0xef, 0xc0], "pxor xmm0, xmm0");
-    test_display(&[0x66, 0x4f, 0x0f, 0xef, 0xc0], "pxor xmm8, xmm8");
-    test_display(&[0xf2, 0x0f, 0x10, 0x0c, 0xc6], "movsd xmm1, [rsi + rax * 8]");
-    test_display(&[0xf3, 0x0f, 0x10, 0x04, 0x86], "movss xmm0, [rsi + rax * 4]");
-    test_display(&[0xf2, 0x0f, 0x59, 0xc8], "mulsd xmm1, xmm0");
-    test_display(&[0xf3, 0x0f, 0x59, 0xc8], "mulss xmm1, xmm0");
-    test_display(&[0xf2, 0x4f, 0x0f, 0x59, 0xc8], "mulsd xmm9, xmm8");
-
-    test_display(
-        &[0xf3, 0x4f, 0x0f, 0x6f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "movdqu xmm11, [r12 + r11 * 4 - 0x334455cc]"
-    );
-    test_display(&[0xf3, 0x0f, 0x70, 0xc0, 0x4e], "pshufhw xmm0, xmm0, 0x4e");
-    test_display(&[0xf3, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
-    test_display(&[0xf3, 0x4f, 0x0f, 0x7e, 0xc1], "movd r9, mm0"); // use of rex.w demotes to movd r/mm
-    test_display(&[0xf3, 0x40, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
-    test_display(&[0xf3, 0x41, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm9");
-    test_display(&[0xf3, 0x42, 0x0f, 0x7e, 0xc1], "movq xmm0, xmm1");
-    test_display(&[0xf3, 0x44, 0x0f, 0x7e, 0xc1], "movq xmm8, xmm1");
-    test_display(&[0xf3, 0x48, 0x0f, 0x7e, 0xc1], "movd rcx, mm0");
-    test_display(
-        &[0xf3, 0x4f, 0x0f, 0x7f, 0x9c, 0x9c, 0x34, 0xaa, 0xbb, 0xcc],
-        "movdqu [r12 + r11 * 4 - 0x334455cc], xmm11"
-    );
-
-    test_display(&[0xf3, 0x0f, 0xc2, 0xc3, 0x08], "cmpss xmm0, xmm3, 0x8");
-    test_display(&[0xf3, 0x4f, 0x0f, 0xc2, 0xc3, 0x08], "cmpss xmm8, xmm11, 0x8");
-    test_display(&[0xf3, 0x4f, 0x0f, 0xc2, 0x03, 0x08], "cmpss xmm8, [r11], 0x8");
-}
-
-// SETLE, SETNG, ...
-
-#[test]
-fn test_mov() {
-    // test_display(&[0xa1, 0x93, 0x62, 0xc4, 0x00, 0x12, 0x34, 0x12, 0x34], "mov eax, [0x3412341200c46293]");
-    // RCT.exe 32bit version, TODO: FIX
-    test_display(&[0xa1, 0x93, 0x62, 0xc4, 0x00], "mov eax, [0xc46293]");
-    test_display(&[0xba, 0x01, 0x00, 0x00, 0x00], "mov edx, 0x1");
-    test_display(&[0x48, 0xc7, 0x04, 0x24, 0x00, 0x00, 0x00, 0x00], "mov [rsp], 0x0");
-    test_display(&[0x48, 0x89, 0x44, 0x24, 0x08], "mov [rsp + 0x8], rax");
-    test_display(&[0x48, 0x89, 0x43, 0x18], "mov [rbx + 0x18], rax");
-    test_display(&[0x48, 0xc7, 0x43, 0x10, 0x00, 0x00, 0x00, 0x00], "mov [rbx + 0x10], 0x0");
-    test_display(&[0x49, 0x89, 0x4e, 0x08], "mov [r14 + 0x8], rcx");
-    test_display(&[0x48, 0x8b, 0x32], "mov rsi, [rdx]");
-    test_display(&[0x4d, 0x8b, 0x4c, 0x10, 0xf8], "mov r9, [r8 + rdx * 1 - 0x8]");
-    test_display(&[0x49, 0x89, 0x46, 0x10], "mov [r14 + 0x10], rax");
-    test_display(&[0x4d, 0x0f, 0x43, 0xec], "cmovnb r13, r12");
-    test_display(&[0x0f, 0xb6, 0x06], "movzx eax, byte [rsi]");
-    test_display(&[0x0f, 0xb7, 0x06], "movzx eax, word [rsi]");
-    test_display(&[0x89, 0x55, 0x94], "mov [rbp - 0x6c], edx");
-    test_display(&[0x65, 0x4c, 0x89, 0x04, 0x25, 0xa8, 0x01, 0x00, 0x00], "mov gs:[0x1a8], r8");
-    test_display(&[0x0f, 0xbe, 0x83, 0xb4, 0x00, 0x00, 0x00], "movsx eax, byte [rbx + 0xb4]");
-    test_display(&[0x46, 0x63, 0xc1], "movsxd r8, ecx");
-    test_display(&[0x48, 0x63, 0x04, 0xba], "movsxd rax, [rdx + rdi * 4]");
-    test_display(&[0xf3, 0x0f, 0x6f, 0x07], "movdqu xmm0, [rdi]");
-    test_display(&[0xf3, 0x0f, 0x7f, 0x45, 0x00], "movdqu [rbp], xmm0");
-}
-
-#[test]
-fn test_stack() {
-    test_display(&[0x66, 0x41, 0x50], "push r8w");
-}
-
-#[test]
-fn test_prefixes() {
-    test_display(&[0x66, 0x41, 0x31, 0xc0], "xor r8w, ax");
-    test_display(&[0x66, 0x41, 0x32, 0xc0], "xor al, r8b");
-    test_display(&[0x40, 0x32, 0xc5], "xor al, bpl");
-
-    // test that WAIT doesn't blow up, at least...
-    assert_eq!(InstDecoder::default().decode([0x9b, 0xf8].iter().cloned()).err(), Some(DecodeError::IncompleteDecoder));
-}
-
-#[test]
-fn test_control_flow() {
-    test_display(&[0x73, 0x31], "jnb 0x31");
-    test_display(&[0x72, 0x5a], "jb 0x5a");
-    test_display(&[0x0f, 0x86, 0x8b, 0x01, 0x00, 0x00], "jna 0x18b");
-    test_display(&[0x74, 0x47], "jz 0x47");
-    test_display(&[0xff, 0x15, 0x7e, 0x72, 0x24, 0x00], "call [rip + 0x24727e]");
-    test_display(&[0xff, 0x24, 0xcd, 0x70, 0xa0, 0xbc, 0x01], "jmp [rcx * 8 + 0x1bca070]");
-    test_display(&[0xff, 0xe0], "jmp rax");
-    test_display(&[0x66, 0xff, 0xe0], "jmp rax");
-    test_display(&[0x67, 0xff, 0xe0], "jmp rax");
-    test_invalid(&[0xff, 0xd8]);
-    test_display(&[0xff, 0x18], "callf [rax]");
-    test_display(&[0xc3], "ret");
-}
-
-#[test]
-fn test_test_cmp() {
-    test_display(&[0xf6, 0x05, 0x2c, 0x9b, 0xff, 0xff, 0x01], "test [rip - 0x64d4], 0x1");
-    test_display(&[0x48, 0x3d, 0x01, 0xf0, 0xff, 0xff], "cmp rax, -0xfff");
-    test_display(&[0x3d, 0x01, 0xf0, 0xff, 0xff], "cmp eax, -0xfff");
-    test_display(&[0x48, 0x83, 0xf8, 0xff], "cmp rax, -0x1");
-    test_display(&[0x48, 0x39, 0xc6], "cmp rsi, rax");
-}
-
-#[test]
-fn test_push_pop() {
-    test_display(&[0x5b], "pop rbx");
-    test_display(&[0x41, 0x5e], "pop r14");
-    test_display(&[0x68, 0x7f, 0x63, 0xc4, 0x00], "push 0xc4637f");
-}
-
-#[test]
-fn test_bmi1() {
-    let bmi1 = InstDecoder::minimal().with_bmi1();
-    let no_bmi1 = InstDecoder::minimal();
-    test_display_under(&bmi1, &[0x41, 0x0f, 0xbc, 0xd3], "tzcnt edx, r11d");
-    test_display_under(&no_bmi1, &[0x41, 0x0f, 0xbc, 0xd3], "bsf edx, r11d");
-
-    test_display_under(&bmi1, &[0xf3, 0x0f, 0xb8, 0xc1], "popcnt eax, ecx");
-    test_display_under(&bmi1, &[0xf3, 0x4f, 0x0f, 0xb8, 0xc1], "popcnt r8, r9");
-}
-
-#[test]
-fn test_bitwise() {
-    test_display_under(&InstDecoder::minimal(), &[0x41, 0x0f, 0xbc, 0xd3], "bsf edx, r11d");
-    test_display(&[0x48, 0x0f, 0xa3, 0xd0], "bt rax, rdx");
-    test_display(&[0x48, 0x0f, 0xab, 0xd0], "bts rax, rdx");
-}
-
-#[test]
-fn test_misc() {
-    // TODO
-//    test_display(&[0xf2, 0x0f, 0x38, 0xf0, 0xc1], "crc32 eax, cl");
-//    test_display(&[0xf2, 0x0f, 0x38, 0xf1, 0xc1], "crc32 eax, ecx");
-    test_display(&[0xfe, 0x00], "inc [rax]"); // TODO: inc byte [rax]
-    test_display(&[0xfe, 0x08], "dec [rax]"); // TODO: dec byte [rax]
-    test_display(&[0xff, 0x00], "inc [rax]"); // TODO: inc dword [rax]
-    test_display(&[0x48, 0xff, 0x00], "inc [rax]"); // TODO: inc qword [rax]
-    test_display(&[0xe4, 0x99], "in al, 0x99");
-    test_display(&[0xe5, 0x99], "in eax, 0x99");
-    test_display(&[0x67, 0xe5, 0x99], "in eax, 0x99");
-    test_display(&[0x4f, 0xe5, 0x99], "in eax, 0x99");
-    test_display(&[0xe6, 0x99], "out 0x99, al");
-    test_display(&[0x4f, 0xe7, 0x99], "out 0x99, eax");
-    test_display(&[0xec], "in al, dx");
-    test_display(&[0xed], "in eax, dx");
-    test_display(&[0xee], "out dx, al");
-    test_display(&[0xef], "out dx, eax");
-    test_display(&[0xcd, 0x00], "int 0x0");
-    test_display(&[0xcd, 0xff], "int 0xff");
-    test_display(&[0x9c], "pushf");
-    test_display(&[0x48, 0x98], "cdqe");
-    test_display(&[0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00], "nop cs:[rax + rax * 1]");
-    test_display(&[0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00], "nop [rax + rax * 1]");
-    test_display(&[0x48, 0x8d, 0xa4, 0xc7, 0x20, 0x00, 0x00, 0x12], "lea rsp, [rdi + rax * 8 + 0x12000020]");
-    test_display(&[0x33, 0xc0], "xor eax, eax");
-    test_display(&[0x48, 0x8d, 0x53, 0x08], "lea rdx, [rbx + 0x8]");
-    test_display(&[0x31, 0xc9], "xor ecx, ecx");
-    test_display(&[0x48, 0x29, 0xc8], "sub rax, rcx");
-    test_display(&[0x48, 0x03, 0x0b], "add rcx, [rbx]");
-    test_display(&[0x48, 0x8d, 0x0c, 0x12], "lea rcx, [rdx + rdx * 1]");
-    test_display(&[0xf6, 0xc2, 0x18], "test dl, 0x18");
-    test_display(&[0xf3, 0x48, 0xab], "rep stos es:[rdi], rax");
-    test_display(&[0xf3, 0x48, 0xa5], "rep movs es:[rdi], ds:[rsi]");
-    test_display(&[0xf3, 0x45, 0x0f, 0xbc, 0xd7], "tzcnt r10d, r15d");
-
-    // this is actually vmx
-    // test_invalid(&[0x66, 0x0f, 0xc7, 0x03]);
-    test_display(&[0x66, 0x4f, 0x0f, 0xc7, 0x33], "vmclear [r11]");
-    test_display(&[0x66, 0x0f, 0xc7, 0x33], "vmclear [rbx]");
-    test_display(&[0xf3, 0x4f, 0x0f, 0xc7, 0x33], "vmxon [r11]");
-    test_display(&[0xf3, 0x0f, 0xc7, 0x33], "vmxon [rbx]");
-}
-
-#[test]
-#[ignore]
-// TODO also not supported at all
-fn evex() {
-    test_display(&[0x62, 0xf2, 0x7d, 0x48, 0x2a, 0x44, 0x40, 0x01], "vmovntdqa zmm0, [rax + rax*2 + 0x40]");
-    test_display(&[0x62, 0xf2, 0x7d, 0x08, 0x2a, 0x44, 0x40, 0x01], "vmovntdqa xmm0, [rax + rax*2 + 0x10]");
-}
-
-#[test]
-fn test_vex() {
-    fn test_instr(bytes: &[u8], text: &'static str) {
-        test_display_under(&InstDecoder::minimal().with_avx(), bytes, text);
-        test_display_under(&InstDecoder::default(), bytes, text);
-        test_invalid_under(&InstDecoder::minimal(), bytes);
-    }
-
-    fn test_instr_invalid(bytes: &[u8]) {
-        test_invalid_under(&InstDecoder::minimal().with_avx(), bytes);
-        test_invalid_under(&InstDecoder::default(), bytes);
-    }
-
-    test_instr(&[0xc5, 0xf8, 0x10, 0x00], "vmovups xmm0, [rax]");
-    test_instr(&[0xc5, 0xf8, 0x10, 0x01], "vmovups xmm0, [rcx]");
-    test_instr(&[0xc5, 0x78, 0x10, 0x0f], "vmovups xmm9, [rdi]");
-    test_instr(&[0xc5, 0xf8, 0x10, 0xcf], "vmovups xmm1, xmm7");
-    test_instr(&[0xc5, 0xf9, 0x10, 0x0f], "vmovupd xmm1, [rdi]");
-    test_instr(&[0xc5, 0xfa, 0x7e, 0x10], "vmovq xmm2, [rax]");
-    test_instr(&[0xc5, 0xfc, 0x10, 0x0f], "vmovups ymm1, [rdi]");
-    test_instr(&[0xc5, 0xfd, 0x10, 0x0f], "vmovupd ymm1, [rdi]");
-    test_instr(&[0xc5, 0xfe, 0x10, 0x0f], "vmovss xmm1, [rdi]");
-    test_instr(&[0xc5, 0xff, 0x10, 0xcf], "vmovsd xmm1, xmm0, xmm7");
-    test_instr(&[0xc5, 0xff, 0x10, 0x01], "vmovsd xmm0, [rcx]");
-    test_instr(&[0xc5, 0xf9, 0x6e, 0xc6], "vmovd xmm0, esi");
-    test_instr(&[0xc5, 0xf9, 0x6e, 0x13], "vmovd xmm2, [rbx]");
-    test_instr(&[0xc5, 0xf9, 0x7e, 0xc6], "vmovd esi, xmm0");
-    test_instr(&[0xc5, 0xf9, 0x7e, 0x13], "vmovd [rbx], xmm2");
-    test_instr_invalid(&[0x4f, 0xc5, 0xf8, 0x10, 0x00]);
-    test_instr_invalid(&[0xf0, 0xc5, 0xf8, 0x10, 0x00]);
-    test_instr(&[0xc4, 0x02, 0x71, 0x00, 0x0f], "vpshufb xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x00, 0x0f], "vpshufb ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x00, 0xcd], "vpshufb xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x00, 0xcd], "vpshufb ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x01, 0x0f], "vphaddw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x01, 0x0f], "vphaddw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x01, 0xcd], "vphaddw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x01, 0xcd], "vphaddw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x02, 0x0f], "vphaddd xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x02, 0x0f], "vphaddd ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x02, 0xcd], "vphaddd xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x02, 0xcd], "vphaddd ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x03, 0x0f], "vphaddsw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x03, 0x0f], "vphaddsw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x03, 0xcd], "vphaddsw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x03, 0xcd], "vphaddsw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x04, 0x0f], "vphaddubsw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x04, 0x0f], "vphaddubsw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x04, 0xcd], "vphaddubsw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x04, 0xcd], "vphaddubsw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x05, 0x0f], "vphsubw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x05, 0x0f], "vphsubw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x05, 0xcd], "vphsubw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x05, 0xcd], "vphsubw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x06, 0x0f], "vphsubd xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x06, 0x0f], "vphsubd ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x06, 0xcd], "vphsubd xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x06, 0xcd], "vphsubd ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x07, 0x0f], "vphsubsw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x07, 0x0f], "vphsubsw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x07, 0xcd], "vphsubsw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x07, 0xcd], "vphsubsw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x08, 0x0f], "vpsignb xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x08, 0x0f], "vpsignb ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x08, 0xcd], "vpsignb xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x08, 0xcd], "vpsignb ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x09, 0x0f], "vpsignw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x09, 0x0f], "vpsignw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x09, 0xcd], "vpsignw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x09, 0xcd], "vpsignw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0a, 0x0f], "vpsignd xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0a, 0x0f], "vpsignd ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0a, 0xcd], "vpsignd xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0a, 0xcd], "vpsignd ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0b, 0x0f], "vpmulhrsw xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0b, 0x0f], "vpmulhrsw ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0b, 0xcd], "vpmulhrsw xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0b, 0xcd], "vpmulhrsw ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0c, 0x0f], "vpermilps xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0c, 0x0f], "vpermilps ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0c, 0xcd], "vpermilps xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0c, 0xcd], "vpermilps ymm9, ymm1, ymm13");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0d, 0x0f], "vpermilpd xmm9, xmm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0d, 0x0f], "vpermilpd ymm9, ymm1, [r15]");
-    test_instr(&[0xc4, 0x02, 0x71, 0x0d, 0xcd], "vpermilpd xmm9, xmm1, xmm13");
-    test_instr(&[0xc4, 0x02, 0x75, 0x0d, 0xcd], "vpermilpd ymm9, ymm1, ymm13");
-    test_instr_invalid(&[0xc4, 0x02, 0x71, 0x0e, 0x00]);
-    test_instr(&[0xc4, 0x02, 0x79, 0x0e, 0x0f], "vtestps xmm9, [r15]");
-    test_instr(&[0xc4, 0x02, 0x7d, 0x0e, 0x0f], "vtestps ymm9, [r15]");
-    test_instr(&[0xc4, 0x02, 0x79, 0x0e, 0xcd], "vtestps xmm9, xmm13");
-    test_instr(&[0xc4, 0x02, 0x7d, 0x0e, 0xcd], "vtestps ymm9, ymm13");
-    test_instr_invalid(&[0xc4, 0x02, 0x71, 0x0f, 0x00]);
-    test_instr(&[0xc4, 0x02, 0x79, 0x0f, 0x0f], "vtestpd xmm9, [r15]");
-    test_instr(&[0xc4, 0x02, 0x7d, 0x0f, 0x0f], "vtestpd ymm9, [r15]");
-    test_instr(&[0xc4, 0x02, 0x79, 0x0f, 0xcd], "vtestpd xmm9, xmm13");
-    test_instr(&[0xc4, 0x02, 0x7d, 0x0f, 0xcd], "vtestpd ymm9, ymm13");
-    test_instr(&[0xc4, 0xe2, 0x65, 0x90, 0x04, 0x51], "vpgatherdd ymm0, [rcx + ymm2 * 2], ymm3");
-    test_instr(&[0xc4, 0xe2, 0xe5, 0x90, 0x04, 0x51], "vpgatherdq ymm0, [rcx + ymm2 * 2], ymm3");
-    test_instr(&[0xc4, 0xe2, 0x65, 0x91, 0x04, 0x51], "vpgatherqd ymm0, [rcx + ymm2 * 2], ymm3");
-    test_instr(&[0xc4, 0xe2, 0xe5, 0x91, 0x04, 0x51], "vpgatherqq ymm0, [rcx + ymm2 * 2], ymm3");
-    test_instr(&[0xc4, 0x02, 0x09, 0x9d, 0xcd], "vfnmadd132ss xmm9, xmm14, xmm13");
-    test_instr(&[0xc4, 0x02, 0x89, 0x9d, 0xcd], "vfnmadd132sd xmm9, xmm14, xmm13");
-// ...
-    test_instr(&[0xc4, 0xe3, 0x79, 0x14, 0xd0, 0x0a], "vpextrb rax, xmm2, 0xa");
-    test_instr(&[0xc4, 0xe3, 0x79, 0x14, 0x10, 0x0a], "vpextrb [rax], xmm2, 0xa");
-    test_instr_invalid(&[0xc4, 0xe3, 0xf9, 0x14, 0x00, 0xd0]);
-    test_instr_invalid(&[0xc4, 0xe3, 0xf9, 0x14, 0x00, 0x0a]);
-}
-
-#[test]
-fn strange_prefixing() {
-    test_display(&[0x45, 0x66, 0x0f, 0x21, 0xc8], "mov rax, dr1");
-    test_display(&[0x45, 0xf2, 0x0f, 0x21, 0xc8], "mov rax, dr1");
-    test_display(&[0x45, 0xf3, 0x0f, 0x21, 0xc8], "mov rax, dr1");
-}
-
-#[test]
-fn prefixed_0f() {
-    test_display(&[0x0f, 0x02, 0xc0], "lar eax, ax");
-    test_display(&[0x48, 0x0f, 0x02, 0xc0], "lar rax, ax");
-    test_display(&[0x0f, 0x03, 0xc0], "lsl eax, eax");
-    test_display(&[0x48, 0x0f, 0x03, 0xc0], "lsl rax, rax");
-    test_display(&[0x0f, 0x05], "syscall");
-    test_display(&[0x48, 0x0f, 0x05], "syscall");
-    test_display(&[0x66, 0x0f, 0x05], "syscall");
-    test_display(&[0x0f, 0x06], "clts");
-    test_display(&[0xf2, 0x0f, 0x06], "clts");
-    test_display(&[0x0f, 0x07], "sysret");
-    test_display(&[0xf2, 0x0f, 0x07], "sysret");
-    test_display(&[0x0f, 0x12, 0x0f], "movlps xmm1, [rdi]");
-    test_display(&[0x0f, 0x12, 0xcf], "movhlps xmm1, xmm7");
-    test_display(&[0x0f, 0x16, 0x0f], "movhps xmm1, [rdi]");
-    test_display(&[0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
-    test_display(&[0x0f, 0x12, 0xc0], "movhlps xmm0, xmm0");
-    test_invalid(&[0x0f, 0x13, 0xc0]);
-    test_display(&[0x0f, 0x13, 0x00], "movlps [rax], xmm0");
-    test_display(&[0x0f, 0x14, 0x08], "unpcklps xmm1, [rax]");
-    test_display(&[0x0f, 0x15, 0x08], "unpckhps xmm1, [rax]");
-    test_display(&[0x0f, 0x16, 0x0f], "movhps xmm1, [rdi]");
-    test_display(&[0x0f, 0x16, 0xc0], "movlhps xmm0, xmm0");
-    test_invalid(&[0x0f, 0x17, 0xc0]);
-    test_display(&[0x0f, 0x17, 0x00], "movhps [rax], xmm0");
-    test_invalid(&[0x0f, 0x18, 0xc0]);
-    test_display(&[0x0f, 0x18, 0x00], "prefetchnta [rax]");
-    test_display(&[0x0f, 0x18, 0x08], "prefetch0 [rax]");
-    test_display(&[0x0f, 0x18, 0x10], "prefetch1 [rax]");
-    test_display(&[0x0f, 0x18, 0x18], "prefetch2 [rax]");
-    test_display(&[0x0f, 0x18, 0x20], "nop [rax]");
-    test_display(&[0x4f, 0x0f, 0x18, 0x20], "nop [r8]");
-    test_display(&[0x0f, 0x19, 0x20], "nop [rax]");
-    test_display(&[0x0f, 0x1a, 0x20], "nop [rax]");
-    test_display(&[0x0f, 0x1b, 0x20], "nop [rax]");
-    test_display(&[0x0f, 0x1c, 0x20], "nop [rax]");
-    test_display(&[0x0f, 0x1d, 0x20], "nop [rax]");
-    test_display(&[0x0f, 0x1e, 0x20], "nop [rax]");
-    test_display(&[0x0f, 0x1f, 0x20], "nop [rax]");
-    test_display(&[0x45, 0x0f, 0x20, 0xc8], "mov r8, cr9");
-    test_display(&[0x0f, 0x20, 0xc8], "mov rax, cr1");
-    test_display(&[0x45, 0x0f, 0x21, 0xc8], "mov r8, dr9");
-    test_display(&[0x0f, 0x21, 0xc8], "mov rax, dr1");
-    test_display(&[0x45, 0x0f, 0x22, 0xc8], "mov cr9, r8");
-    test_display(&[0x40, 0x0f, 0x22, 0xc8], "mov cr1, rax");
-    test_display(&[0x0f, 0x22, 0xc8], "mov cr1, rax");
-    test_display(&[0x44, 0x0f, 0x22, 0xcf], "mov cr9, rdi");
-    test_display(&[0x0f, 0x22, 0xcf], "mov cr1, rdi");
-    test_display(&[0x45, 0x0f, 0x23, 0xc8], "mov dr9, r8");
-    test_display(&[0x40, 0x0f, 0x23, 0xc8], "mov dr1, rax");
-    test_display(&[0x0f, 0x23, 0xc8], "mov dr1, rax");
-    test_display(&[0x44, 0x0f, 0x23, 0xcf], "mov dr9, rdi");
-    test_display(&[0x0f, 0x23, 0xcf], "mov dr1, rdi");
-    test_display(&[0x0f, 0x30], "wrmsr");
-    test_display(&[0x0f, 0x31], "rdtsc");
-    test_display(&[0x0f, 0x32], "rdmsr");
-    test_display(&[0x0f, 0x33], "rdpmc");
-    test_display(&[0x0f, 0x34], "sysenter");
-    test_display(&[0x0f, 0x35], "sysexit");
-    test_invalid(&[0x0f, 0x36]);
-    test_display(&[0x0f, 0x37], "getsec");
-    test_display(&[0x0f, 0x60, 0x00], "punpcklbw mm0, [rax]");
-    test_display(&[0x0f, 0x60, 0xc2], "punpcklbw mm0, mm2");
-    test_display(&[0x0f, 0x61, 0x00], "punpcklwd mm0, [rax]");
-    test_display(&[0x0f, 0x61, 0xc2], "punpcklwd mm0, mm2");
-    test_display(&[0x0f, 0x62, 0x00], "punpckldq mm0, [rax]");
-    test_display(&[0x0f, 0x62, 0xc2], "punpckldq mm0, mm2");
-    test_display(&[0x0f, 0x63, 0x00], "packsswb mm0, [rax]");
-    test_display(&[0x0f, 0x63, 0xc2], "packsswb mm0, mm2");
-    test_display(&[0x0f, 0x64, 0x00], "pcmpgtb mm0, [rax]");
-    test_display(&[0x0f, 0x64, 0xc2], "pcmpgtb mm0, mm2");
-    test_display(&[0x0f, 0x65, 0x00], "pcmpgtw mm0, [rax]");
-    test_display(&[0x0f, 0x65, 0xc2], "pcmpgtw mm0, mm2");
-    test_display(&[0x0f, 0x66, 0x00], "pcmpgtd mm0, [rax]");
-    test_display(&[0x0f, 0x66, 0xc2], "pcmpgtd mm0, mm2");
-    test_display(&[0x0f, 0x67, 0x00], "packuswb mm0, [rax]");
-    test_display(&[0x0f, 0x67, 0xc2], "packuswb mm0, mm2");
-    test_display(&[0x0f, 0x68, 0x00], "punpckhbw mm0, [rax]");
-    test_display(&[0x0f, 0x68, 0xc2], "punpckhbw mm0, mm2");
-    test_display(&[0x0f, 0x69, 0x00], "punpckhwd mm0, [rax]");
-    test_display(&[0x0f, 0x69, 0xc2], "punpckhwd mm0, mm2");
-    test_display(&[0x0f, 0x6a, 0x00], "punpckhdq mm0, [rax]");
-    test_display(&[0x0f, 0x6a, 0xc2], "punpckhdq mm0, mm2");
-    test_display(&[0x0f, 0x6b, 0x00], "packssdw mm0, [rax]");
-    test_display(&[0x0f, 0x6b, 0xc2], "packssdw mm0, mm2");
-    test_invalid(&[0x0f, 0x6c]);
-    test_invalid(&[0x0f, 0x6d]);
-    test_display(&[0x0f, 0x6e, 0x00], "movd mm0, [rax]");
-    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, 0x70, 0x00, 0x7f], "pshufw mm0, [rax], 0x7f");
-    test_display(&[0x4f, 0x0f, 0x70, 0x00, 0x7f], "pshufw mm0, [r8], 0x7f");
-    test_invalid(&[0x0f, 0x71, 0x00, 0x7f]);
-    test_invalid(&[0x0f, 0x71, 0xc0, 0x7f]);
-    test_display(&[0x0f, 0x71, 0xd0, 0x7f], "psrlw mm0, 0x7f");
-    test_display(&[0x0f, 0x71, 0xe0, 0x7f], "psraw mm0, 0x7f");
-    test_display(&[0x0f, 0x71, 0xf0, 0x7f], "psllw mm0, 0x7f");
-    test_invalid(&[0x0f, 0x72, 0x00, 0x7f]);
-    test_invalid(&[0x0f, 0x72, 0xc0, 0x7f]);
-    test_display(&[0x0f, 0x72, 0xd0, 0x7f], "psrld mm0, 0x7f");
-    test_display(&[0x0f, 0x72, 0xe0, 0x7f], "psrad mm0, 0x7f");
-    test_display(&[0x0f, 0x72, 0xf0, 0x7f], "pslld mm0, 0x7f");
-    test_invalid(&[0x0f, 0x73, 0x00, 0x7f]);
-    test_invalid(&[0x0f, 0x73, 0xc0, 0x7f]);
-    test_display(&[0x0f, 0x73, 0xd0, 0x7f], "psrlq mm0, 0x7f");
-    test_invalid(&[0x0f, 0x73, 0xe0, 0x7f]);
-    test_display(&[0x0f, 0x73, 0xf0, 0x7f], "psllq mm0, 0x7f");
-    test_display(&[0x0f, 0xa0], "push fs");
-    test_display(&[0x0f, 0xa1], "pop fs");
-    test_display(&[0x0f, 0xa2], "cpuid");
-    test_display(&[0x0f, 0xa4, 0xc0, 0x11], "shld eax, eax, 0x11");
-    test_display(&[0x66, 0x0f, 0xa4, 0xcf, 0x11], "shld di, cx, 0x11");
-    test_display(&[0x66, 0x45, 0x0f, 0xa4, 0xcf, 0x11], "shld r15w, r9w, 0x11");
-    test_display(&[0x0f, 0xa5, 0xc0], "shld eax, eax, cl");
-    test_display(&[0x0f, 0xa5, 0xc9], "shld ecx, ecx, cl");
-}
-
-#[test]
-fn prefixed_660f() {
-    test_display(&[0x66, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
-    test_display(&[0x66, 0x48, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
-    test_display(&[0x66, 0x4a, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm0");
-    test_display(&[0x66, 0x4b, 0x0f, 0x10, 0xc0], "movupd xmm0, xmm8");
-    test_display(&[0x66, 0x4c, 0x0f, 0x10, 0xc0], "movupd xmm8, xmm0");
-    test_display(&[0x66, 0x4d, 0x0f, 0x10, 0xc0], "movupd xmm8, xmm8");
-    test_display(&[0xf2, 0x66, 0x66, 0x4d, 0x0f, 0x10, 0xc0], "movupd xmm8, xmm8");
-}
-
-#[test]
-fn prefixed_f20f() {
-    test_display(&[0xf2, 0x0f, 0x16, 0xcf], "movlhps xmm1, xmm7");
-    test_display(&[0xf2, 0x4d, 0x0f, 0x16, 0xcf], "movlhps xmm9, xmm15");
-    test_display(&[0x40, 0x66, 0xf2, 0x66, 0x4d, 0x0f, 0x16, 0xcf], "movlhps xmm9, xmm15");
-}
-
-#[test]
-fn prefixed_f30f() {
-    test_display(&[0xf3, 0x0f, 0x16, 0xcf], "movshdup xmm1, xmm7");
-    test_display(&[0xf3, 0x4d, 0x0f, 0x16, 0xcf], "movshdup xmm9, xmm15");
-}
+mod long_mode;
+mod protected_mode;
-- 
cgit v1.1