3dnow was still supported on K8, K10. 32-bit mode should learn about uarch tweaks too

5 files changed, 271 insertions, 54 deletions

diff --git a/src/long_mode/uarch.rs b/src/long_mode/uarch.rs
index 5af0175..63fa972 100644
--- a/src/long_mode/uarch.rs
+++ b/src/long_mode/uarch.rs
@@ -35,13 +35,15 @@ pub mod amd {
     /// support - SSE2 and no later.
     pub fn k8() -> InstDecoder {
         InstDecoder::minimal()
+            .with_3dnow()
+            .with_3dnowprefetch()
+            .with_cmov()
     }
 
     /// `k10` was the successor to `k8`, launched in 2007. `k10` cores extended SSE support through
     /// to SSE4.2a, as well as consistent `cmov` support, among other features.
     pub fn k10() -> InstDecoder {
         k8()
-            .with_cmov()
             .with_cmpxchg16b()
             .with_svm()
             .with_abm()
@@ -51,9 +53,20 @@ pub mod amd {
     }
 
     /// `Bulldozer` was the successor to `K10`, launched in 2011. `Bulldozer` cores include AVX
-    /// support among other extensions, and are notable for including `AESNI`.
+    /// support among other extensions, and are notable for including `AESNI`. `Bulldozer` was also
+    /// the first microarchitecture to *remove* support for 3DNow instructions.
     pub fn bulldozer() -> InstDecoder {
-        k10()
+        InstDecoder::minimal()
+            // first, apply all the K8 extensions again, sans 3DNow
+            // .. should be sse, sse2
+            // then the K10
+            .with_cmpxchg16b()
+            .with_svm()
+            .with_abm()
+            .with_lahfsahf()
+            .with_sse3()
+            .with_sse4a()
+            // now the new extensions
             .with_ssse3()
             .with_sse4()
             .with_sse4_2()
@@ -101,9 +114,18 @@ pub mod amd {
     /// SHA, RDSEED, and other extensions.
     pub fn zen() -> InstDecoder {
         // no nice way to *un*set feature bits, but several extensions were dropped.
-        // so, start again from K10.
-        k10()
-            // first, bundle all the K10->Bulldozer features..
+        // so, start again.
+        InstDecoder::minimal()
+            // first, apply all the K8 extensions again, sans 3DNow
+            // .. should be sse, sse2
+            // then the K10
+            .with_cmpxchg16b()
+            .with_svm()
+            .with_abm()
+            .with_lahfsahf()
+            .with_sse3()
+            .with_sse4a()
+            // now, bundle all the K10->Bulldozer features..
             .with_ssse3()
             .with_sse4()
             .with_sse4_2()
@@ -114,11 +136,8 @@ pub mod amd {
             .with_avx()
             .with_xsave()
             .with_skinit()
-            // now all the Bulldozer (/Piledriver/Steamroller/Excavator)->Zen features
+            // finally all the Bulldozer (/Piledriver/Steamroller/Excavator)->Zen features
             .with_avx2()
-            .with_aesni()
-            .with_pclmulqdq()
-            .with_f16c()
             .with_movbe()
             .with_bmi2()
             .with_adx()
diff --git a/src/protected_mode/uarch.rs b/src/protected_mode/uarch.rs
index cbe3e89..6914348 100644
--- a/src/protected_mode/uarch.rs
+++ b/src/protected_mode/uarch.rs
@@ -1,12 +1,24 @@
+//! information for AMD and Intel microarchitectures in the modules below is sourced from a
+//! combination of Wikipedia (especially for dates), one-off research for particular
+//! microarchitectures, and `InstLatx64`'s CPUID dumps via [chip directory](https://github.com/iximeow/chip_directory).
+//!
+//! these microarchitecture-specific decoders are relatively rarely used, but generally should be
+//! accurate.
+
 pub mod amd {
-    //! most information about instruction set extensions for microarchitectures here was sourced
-    //! from
-    //! [https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Unit#Feature_overview](https://docs.rs/yaxpeax-x86/0.0.12/yaxpeax_x86/protected_mode/uarch/intel/index.html)
+    //! initial information for the mircoarchitecture (families) described here came from a
+    //! combination of the Wikipedia pages
+    //! [https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Unit#Feature_overview](https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Unit#Feature_overview)
     //! and
-    //! [https://en.wikipedia.org/wiki/Template:AMD_x86_CPU_features](https://docs.rs/yaxpeax-x86/0.0.12/yaxpeax_x86/protected_mode/uarch/intel/index.html).
-    //! these mappings are best-effort but fairly unused, so a critical eye should be kept towards
-    //! these decoders rejecting instructions they should not, or incorrectly accepting
-    //! instructions.
+    //! [https://en.wikipedia.org/wiki/Template:AMD_x86_CPU_features](https://en.wikipedia.org/wiki/Template:AMD_x86_CPU_features).
+    //! it has been since "augmented" by the CPUID dumps from InstLatx64, via [chip
+    //! directory](https://github.com/iximeow/chip_directory/tree/no-gods-no-/x86). scare quotes
+    //! because in several cases CPUID measurement error adds, rather than removes, ambiguity.
+    //! additionally, for some CPU features, InstLatx64 has CPUID dumps of early engineering
+    //! samples where features are not present. later production steppings of those parts do
+    //! universally have the corresponding feature, which makes it less obvious which features are
+    //! universally present in a family, standardized in a following architecture, unevenly present
+    //! due to market segmentation, and so on.
     //!
     //! microarchitectures as defined here are with respect to flags reported by CPUID. notably,
     //! `Zen` does not report `FMA4` support by `CPUID`, but instructions in that extension
@@ -23,28 +35,41 @@ pub mod amd {
     /// support - SSE2 and no later.
     pub fn k8() -> InstDecoder {
         InstDecoder::minimal()
+            .with_3dnow()
+            .with_3dnowprefetch()
+            .with_cmov()
     }
 
     /// `k10` was the successor to `k8`, launched in 2007. `k10` cores extended SSE support through
     /// to SSE4.2a, as well as consistent `cmov` support, among other features.
     pub fn k10() -> InstDecoder {
         k8()
-            .with_cmov()
             .with_cmpxchg16b()
             .with_svm()
             .with_abm()
             .with_lahfsahf()
             .with_sse3()
-            .with_ssse3()
-            .with_sse4()
-            .with_sse4_2()
             .with_sse4a()
     }
 
     /// `Bulldozer` was the successor to `K10`, launched in 2011. `Bulldozer` cores include AVX
-    /// support among other extensions, and are notable for including `AESNI`.
+    /// support among other extensions, and are notable for including `AESNI`. `Bulldozer` was also
+    /// the first microarchitecture to *remove* support for 3DNow instructions.
     pub fn bulldozer() -> InstDecoder {
-        k10()
+        InstDecoder::minimal()
+            // first, apply all the K8 extensions again, sans 3DNow
+            // .. should be sse, sse2
+            // then the K10
+            .with_cmpxchg16b()
+            .with_svm()
+            .with_abm()
+            .with_lahfsahf()
+            .with_sse3()
+            .with_sse4a()
+            // now the new extensions
+            .with_ssse3()
+            .with_sse4()
+            .with_sse4_2()
             .with_bmi1()
             .with_aesni()
             .with_pclmulqdq()
@@ -52,6 +77,8 @@ pub mod amd {
             .with_avx()
             .with_fma4()
             .with_xop()
+            .with_xsave()
+            .with_skinit()
     }
 
     /// `Piledriver` was the successor to `Bulldozer`, launched in 2012.
@@ -86,21 +113,87 @@ pub mod amd {
     /// instructions to AVX2 and discarded FMA4, TBM, and XOP extensions. they also gained ADX,
     /// SHA, RDSEED, and other extensions.
     pub fn zen() -> InstDecoder {
-        k10()
-            .with_avx()
-            .with_avx2()
+        // no nice way to *un*set feature bits, but several extensions were dropped.
+        // so, start again.
+        InstDecoder::minimal()
+            // first, apply all the K8 extensions again, sans 3DNow
+            // .. should be sse, sse2
+            // then the K10
+            .with_cmpxchg16b()
+            .with_svm()
+            .with_abm()
+            .with_lahfsahf()
+            .with_sse3()
+            .with_sse4a()
+            // now, bundle all the K10->Bulldozer features..
+            .with_ssse3()
+            .with_sse4()
+            .with_sse4_2()
             .with_bmi1()
             .with_aesni()
             .with_pclmulqdq()
             .with_f16c()
+            .with_avx()
+            .with_xsave()
+            .with_skinit()
+            // finally all the Bulldozer (/Piledriver/Steamroller/Excavator)->Zen features
+            .with_avx2()
             .with_movbe()
             .with_bmi2()
-            .with_rdrand()
             .with_adx()
             .with_sha()
+            .with_rdrand()
             .with_rdseed()
             .with_fma3()
-            // TODO: XSAVEC, XSAVES, XRSTORS, CLFLUSHOPT, CLZERO?
+
+            .with_xsavec()
+            .with_xsaves()
+            .with_xsaveopt()
+            .with_clflushopt()
+            .with_clwb()
+            .with_fsgsbase()
+            .with_monitorx()
+    }
+
+    /// `Zen 2`, launched in 2019, succeeded `Zen`/`Zen+`. there aren't many instruction set
+    /// extensions here, but `clwb`, `rdpid`, and `wbnoinvd` show up here.
+    pub fn zen2() -> InstDecoder {
+        zen()
+            .with_clwb()
+            .with_rdpid()
+            .with_wbnoinvd()
+    }
+
+    /// `Zen 3`, launched in 2020, succeeded `Zen 2`. like `Zen 2`, there aren't many instruction
+    /// set extensions here.
+    pub fn zen3() -> InstDecoder {
+        zen2()
+            .with_invpcid()
+            .with_vaes()
+            .with_vpclmulqdq()
+    }
+
+    /// `Zen 4`, launched in 2022, succeeded `Zen 3`. `Zen 4` is notable for being the first AMD
+    /// processor family supporting AVX-512.
+    pub fn zen4() -> InstDecoder {
+        zen3()
+            .with_avx512_f()
+            .with_avx512_vl()
+            .with_avx512_bw()
+            .with_avx512_cd()
+            .with_avx512_cd()
+            .with_avx512_vbmi()
+            .with_avx512_vbmi2()
+            .with_avx512_vpopcntdq()
+            .with_gfni()
+    }
+
+    /// `Zen 5`, launched in 2024, succeeded `Zen 4`. `Zen 5` adds only a few additional
+    /// instructions; some AVX-512 features, `enqcmd`, and `movdir64b`.
+    pub fn zen5() -> InstDecoder {
+        zen4()
+            .with_movdir64b()
+            .with_enqcmd()
     }
 }
 
diff --git a/src/real_mode/uarch.rs b/src/real_mode/uarch.rs
index 60bf168..8df4213 100644
--- a/src/real_mode/uarch.rs
+++ b/src/real_mode/uarch.rs
@@ -1,12 +1,24 @@
+//! information for AMD and Intel microarchitectures in the modules below is sourced from a
+//! combination of Wikipedia (especially for dates), one-off research for particular
+//! microarchitectures, and `InstLatx64`'s CPUID dumps via [chip directory](https://github.com/iximeow/chip_directory).
+//!
+//! these microarchitecture-specific decoders are relatively rarely used, but generally should be
+//! accurate.
+
 pub mod amd {
-    //! most information about instruction set extensions for microarchitectures here was sourced
-    //! from
-    //! [https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Unit#Feature_overview](https://docs.rs/yaxpeax-x86/0.0.12/yaxpeax_x86/real_mode/uarch/intel/index.html)
+    //! initial information for the mircoarchitecture (families) described here came from a
+    //! combination of the Wikipedia pages
+    //! [https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Unit#Feature_overview](https://en.wikipedia.org/wiki/AMD_Accelerated_Processing_Unit#Feature_overview)
     //! and
-    //! [https://en.wikipedia.org/wiki/Template:AMD_x86_CPU_features](https://docs.rs/yaxpeax-x86/0.0.12/yaxpeax_x86/real_mode/uarch/intel/index.html).
-    //! these mappings are best-effort but fairly unused, so a critical eye should be kept towards
-    //! these decoders rejecting instructions they should not, or incorrectly accepting
-    //! instructions.
+    //! [https://en.wikipedia.org/wiki/Template:AMD_x86_CPU_features](https://en.wikipedia.org/wiki/Template:AMD_x86_CPU_features).
+    //! it has been since "augmented" by the CPUID dumps from InstLatx64, via [chip
+    //! directory](https://github.com/iximeow/chip_directory/tree/no-gods-no-/x86). scare quotes
+    //! because in several cases CPUID measurement error adds, rather than removes, ambiguity.
+    //! additionally, for some CPU features, InstLatx64 has CPUID dumps of early engineering
+    //! samples where features are not present. later production steppings of those parts do
+    //! universally have the corresponding feature, which makes it less obvious which features are
+    //! universally present in a family, standardized in a following architecture, unevenly present
+    //! due to market segmentation, and so on.
     //!
     //! microarchitectures as defined here are with respect to flags reported by CPUID. notably,
     //! `Zen` does not report `FMA4` support by `CPUID`, but instructions in that extension
@@ -23,28 +35,41 @@ pub mod amd {
     /// support - SSE2 and no later.
     pub fn k8() -> InstDecoder {
         InstDecoder::minimal()
+            .with_3dnow()
+            .with_3dnowprefetch()
+            .with_cmov()
     }
 
     /// `k10` was the successor to `k8`, launched in 2007. `k10` cores extended SSE support through
     /// to SSE4.2a, as well as consistent `cmov` support, among other features.
     pub fn k10() -> InstDecoder {
         k8()
-            .with_cmov()
             .with_cmpxchg16b()
             .with_svm()
             .with_abm()
             .with_lahfsahf()
             .with_sse3()
-            .with_ssse3()
-            .with_sse4()
-            .with_sse4_2()
             .with_sse4a()
     }
 
     /// `Bulldozer` was the successor to `K10`, launched in 2011. `Bulldozer` cores include AVX
-    /// support among other extensions, and are notable for including `AESNI`.
+    /// support among other extensions, and are notable for including `AESNI`. `Bulldozer` was also
+    /// the first microarchitecture to *remove* support for 3DNow instructions.
     pub fn bulldozer() -> InstDecoder {
-        k10()
+        InstDecoder::minimal()
+            // first, apply all the K8 extensions again, sans 3DNow
+            // .. should be sse, sse2
+            // then the K10
+            .with_cmpxchg16b()
+            .with_svm()
+            .with_abm()
+            .with_lahfsahf()
+            .with_sse3()
+            .with_sse4a()
+            // now the new extensions
+            .with_ssse3()
+            .with_sse4()
+            .with_sse4_2()
             .with_bmi1()
             .with_aesni()
             .with_pclmulqdq()
@@ -52,6 +77,8 @@ pub mod amd {
             .with_avx()
             .with_fma4()
             .with_xop()
+            .with_xsave()
+            .with_skinit()
     }
 
     /// `Piledriver` was the successor to `Bulldozer`, launched in 2012.
@@ -86,21 +113,87 @@ pub mod amd {
     /// instructions to AVX2 and discarded FMA4, TBM, and XOP extensions. they also gained ADX,
     /// SHA, RDSEED, and other extensions.
     pub fn zen() -> InstDecoder {
-        k10()
-            .with_avx()
-            .with_avx2()
+        // no nice way to *un*set feature bits, but several extensions were dropped.
+        // so, start again.
+        InstDecoder::minimal()
+            // first, apply all the K8 extensions again, sans 3DNow
+            // .. should be sse, sse2
+            // then the K10
+            .with_cmpxchg16b()
+            .with_svm()
+            .with_abm()
+            .with_lahfsahf()
+            .with_sse3()
+            .with_sse4a()
+            // now, bundle all the K10->Bulldozer features..
+            .with_ssse3()
+            .with_sse4()
+            .with_sse4_2()
             .with_bmi1()
             .with_aesni()
             .with_pclmulqdq()
             .with_f16c()
+            .with_avx()
+            .with_xsave()
+            .with_skinit()
+            // finally all the Bulldozer (/Piledriver/Steamroller/Excavator)->Zen features
+            .with_avx2()
             .with_movbe()
             .with_bmi2()
-            .with_rdrand()
             .with_adx()
             .with_sha()
+            .with_rdrand()
             .with_rdseed()
             .with_fma3()
-            // TODO: XSAVEC, XSAVES, XRSTORS, CLFLUSHOPT, CLZERO?
+
+            .with_xsavec()
+            .with_xsaves()
+            .with_xsaveopt()
+            .with_clflushopt()
+            .with_clwb()
+            .with_fsgsbase()
+            .with_monitorx()
+    }
+
+    /// `Zen 2`, launched in 2019, succeeded `Zen`/`Zen+`. there aren't many instruction set
+    /// extensions here, but `clwb`, `rdpid`, and `wbnoinvd` show up here.
+    pub fn zen2() -> InstDecoder {
+        zen()
+            .with_clwb()
+            .with_rdpid()
+            .with_wbnoinvd()
+    }
+
+    /// `Zen 3`, launched in 2020, succeeded `Zen 2`. like `Zen 2`, there aren't many instruction
+    /// set extensions here.
+    pub fn zen3() -> InstDecoder {
+        zen2()
+            .with_invpcid()
+            .with_vaes()
+            .with_vpclmulqdq()
+    }
+
+    /// `Zen 4`, launched in 2022, succeeded `Zen 3`. `Zen 4` is notable for being the first AMD
+    /// processor family supporting AVX-512.
+    pub fn zen4() -> InstDecoder {
+        zen3()
+            .with_avx512_f()
+            .with_avx512_vl()
+            .with_avx512_bw()
+            .with_avx512_cd()
+            .with_avx512_cd()
+            .with_avx512_vbmi()
+            .with_avx512_vbmi2()
+            .with_avx512_vpopcntdq()
+            .with_gfni()
+    }
+
+    /// `Zen 5`, launched in 2024, succeeded `Zen 4`. `Zen 5` adds only a few additional
+    /// instructions; some AVX-512 features, `enqcmd`, and `movdir64b`.
+    pub fn zen5() -> InstDecoder {
+        zen4()
+            .with_movdir64b()
+            .with_enqcmd()
     }
 }
 
diff --git a/test/long_mode/mod.rs b/test/long_mode/mod.rs
index 9493a63..2c7771c 100644
--- a/test/long_mode/mod.rs
+++ b/test/long_mode/mod.rs
@@ -3454,7 +3454,8 @@ fn test_3dnow() {
         test_display_under(&InstDecoder::default(), bytes, text);
         test_invalid_under(&InstDecoder::minimal(), bytes);
         test_invalid_under(&InstDecoder::minimal(), bytes);
-        test_invalid_under(&yaxpeax_x86::long_mode::uarch::amd::k8(), bytes);
+        test_display_under(&yaxpeax_x86::long_mode::uarch::amd::k8(), bytes, text);
+        test_invalid_under(&yaxpeax_x86::long_mode::uarch::amd::bulldozer(), bytes);
         test_invalid_under(&yaxpeax_x86::long_mode::uarch::intel::netburst(), bytes);
     }
 
diff --git a/test/protected_mode/mod.rs b/test/protected_mode/mod.rs
index 8fecdab..29e9ec4 100644
--- a/test/protected_mode/mod.rs
+++ b/test/protected_mode/mod.rs
@@ -3068,13 +3068,24 @@ fn test_sse4a() {
 
 #[test]
 fn test_3dnow() {
-    test_display(&[0x0f, 0x0f, 0xe0, 0x8a], "pfnacc mm4, mm0");
-    test_display(&[0x0f, 0x0f, 0x38, 0x8e], "pfpnacc mm7, qword [eax]");
-    test_display(&[0x65, 0x67, 0x65, 0x65, 0x0f, 0x0e], "femms");
-    test_display(&[0x3e, 0xf3, 0x2e, 0xf2, 0x0f, 0x0f, 0x64, 0x93, 0x93, 0xa4], "pfmax mm4, qword cs:[ebx + edx * 4 - 0x6d]");
-    test_display(&[0x26, 0x36, 0x0f, 0x0f, 0x70, 0xfb, 0x0c], "pi2fw mm6, qword ss:[eax - 0x5]");
-    test_display(&[0x66, 0x0f, 0x0f, 0xc6, 0xb7], "pmulhrw mm0, mm6");
-    test_display(&[0x0f, 0x0f, 0xc6, 0xb7], "pmulhrw mm0, mm6");
+    fn test_instr(bytes: &[u8], text: &'static str) {
+        test_display_under(&InstDecoder::minimal().with_3dnow(), bytes, text);
+        test_display_under(&InstDecoder::default(), bytes, text);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+        test_invalid_under(&InstDecoder::minimal(), bytes);
+        test_display_under(&yaxpeax_x86::protected_mode::uarch::amd::k8(), bytes, text);
+        test_invalid_under(&yaxpeax_x86::protected_mode::uarch::amd::bulldozer(), bytes);
+        test_invalid_under(&yaxpeax_x86::protected_mode::uarch::intel::netburst(), bytes);
+    }
+
+    test_instr(&[0x0f, 0x0f, 0xe0, 0x8a], "pfnacc mm4, mm0");
+    test_instr(&[0x0f, 0x0f, 0x38, 0x8e], "pfpnacc mm7, qword [eax]");
+    test_instr(&[0x65, 0x67, 0x65, 0x65, 0x0f, 0x0e], "femms");
+    test_instr(&[0x3e, 0xf3, 0x2e, 0xf2, 0x0f, 0x0f, 0x64, 0x93, 0x93, 0xa4], "pfmax mm4, qword cs:[ebx + edx * 4 - 0x6d]");
+    test_instr(&[0x26, 0x36, 0x0f, 0x0f, 0x70, 0xfb, 0x0c], "pi2fw mm6, qword ss:[eax - 0x5]");
+    test_instr(&[0x66, 0x0f, 0x0f, 0xc6, 0xb7], "pmulhrw mm0, mm6");
+    test_instr(&[0x0f, 0x0f, 0xc6, 0xb7], "pmulhrw mm0, mm6");
+    test_instr(&[0x0f, 0x0e], "femms");
 }
 
 // first appeared in tremont