From eec7972808e751dd834043248ed7d64008faba53 Mon Sep 17 00:00:00 2001
From: lioncash <mathew1800@gmail.com>
Date: Mon, 20 Jun 2022 10:24:03 -0400
Subject: [PATCH] UContext: Add definitions for accessing XSAVE data

Adds the layout of extended state for accessing/saving ymm high lanes
---
 .../FEXCore/include/FEXCore/Core/UContext.h   | 98 ++++++++++++++++++-
 1 file changed, 96 insertions(+), 2 deletions(-)

diff --git a/External/FEXCore/include/FEXCore/Core/UContext.h b/External/FEXCore/include/FEXCore/Core/UContext.h
index 9697924e5..38adb0b99 100644
--- a/External/FEXCore/include/FEXCore/Core/UContext.h
+++ b/External/FEXCore/include/FEXCore/Core/UContext.h
@@ -27,6 +27,57 @@ namespace FEXCore {
     };
     static_assert(sizeof(FEXCore::x86_64::stack_t) == 24, "This needs to be the right size");
 
+    /**
+     * Describes the software specific bytes added at the end of the
+     * fpstate to identify whether or not an extended context area is
+     * present and what kind of extended features are present in said
+     * context area.
+     */
+    struct FEX_PACKED fpx_sw_bytes {
+      static constexpr uint32_t FP_XSTATE_MAGIC = 0x46505853;
+
+      enum FeatureFlag : uint32_t {
+        FEATURE_FP         = 1U << 0,
+        FEATURE_SSE        = 1U << 1,
+        FEATURE_YMM        = 1U << 2,
+        FEATURE_BNDREGS    = 1U << 3,
+        FEATURE_BNDCSR     = 1U << 4,
+        FEATURE_OPMASK     = 1U << 5,
+        FEATURE_ZMM_Hi256  = 1U << 6,
+        FEATURE_Hi16_ZMM   = 1U << 7,
+        FEATURE_PT_UNIMPL  = 1U << 8,
+        FEATURE_PKRU       = 1U << 9,
+        FEATURE_PASID      = 1U << 10,
+        FEATURE_RESERVED11 = 1U << 11,
+        FEATURE_RESERVED12 = 1U << 12,
+        FEATURE_RESERVED13 = 1U << 13,
+        FEATURE_RESERVED14 = 1U << 14,
+        FEATURE_LBR        = 1U << 15,
+        FEATURE_RESERVED16 = 1U << 16,
+        FEATURE_XTILE_CFG  = 1U << 17,
+        FEATURE_XTILE_DATA = 1U << 18,
+      };
+
+      // If magic1 is set to FP_XSTATE_MAGIC, then the encompassing
+      // frame is an xstate frame. If 0, then it's a legacy frame.
+      uint32_t magic1;
+
+      // Total size of the fpstate area
+      // - magic1 = 0               -> sizeof(fpstate)
+      // - magic1 = FP_XSTATE_MAGIC -> sizeof(xstate) + extensions (if any)
+      uint32_t extended_size;
+
+      // Feature bitmask describing supported features.
+      uint64_t xfeatures;
+
+      // Actual XSAVE state size, based on above xfeatures
+      uint32_t xstate_size;
+
+      // Reserved data
+      uint32_t padding[7];
+    };
+    static_assert(sizeof(fpx_sw_bytes) == 48);
+
     struct FEX_PACKED _libc_fpstate {
       // This is in FXSAVE format
       uint16_t fcw;
@@ -39,10 +90,37 @@ namespace FEXCore {
       uint32_t mxcsr_mask;
       __uint128_t _st[8];
       __uint128_t _xmm[16];
-      uint32_t _res[24];
+      uint32_t _res[12];
+
+      // Linux uses 12 of the bytes relegated for software purposes
+      // to store info describing any existing XSAVE context data.
+      fpx_sw_bytes sw_reserved;
     };
     static_assert(sizeof(FEXCore::x86_64::_libc_fpstate) == 512, "This needs to be the right size");
 
+    struct FEX_PACKED xstate_header {
+      uint64_t xfeatures;
+      uint64_t reserved1[2];
+      uint64_t reserved2[5];
+    };
+    static_assert(sizeof(xstate_header) == 64);
+
+    struct FEX_PACKED ymmh_state {
+      __uint128_t ymmh_space[16];
+    };
+    static_assert(sizeof(ymmh_state) == 256);
+
+    /**
+     * Extended state that includes both the main fpstate
+     * and the extended state.
+     */
+    struct FEX_PACKED xstate {
+      _libc_fpstate fpstate;
+      xstate_header xstate_hdr;
+      ymmh_state ymmh;
+    };
+    static_assert(sizeof(xstate) == 832);
+
     ///< The order of these must match the GNU ordering
     enum ContextRegs {
       FEX_REG_R8 = 0,
@@ -233,6 +311,11 @@ namespace FEXCore {
     };
     static_assert(sizeof(FEXCore::x86::_libc_fpreg) == 10, "This needs to be the right size");
 
+    // Same layout on both x86 and x86_64
+    using fpx_sw_bytes = x86_64::fpx_sw_bytes;
+    using xstate_header = x86_64::xstate_header;
+    using ymmh_state = x86_64::ymmh_state;
+
     enum fpstate_magic {
       // Legacy fpstate
       MAGIC_FPU = 0xFFFF'0000,
@@ -257,10 +340,21 @@ namespace FEXCore {
       __uint128_t _st_pad[8]; // Ignored st data
       __uint128_t _xmm[8]; // First 8 XMM registers
       uint32_t pad2[44]; // Second 8 XMM registers plus padding
-      uint32_t pad3[12]; // extended state encoding
+      fpx_sw_bytes sw_reserved; // extended state encoding
     };
     static_assert(sizeof(FEXCore::x86::_libc_fpstate) == 624, "This needs to be the right size");
 
+    /**
+     * Extended state that includes both the main fpstate
+     * and the extended state.
+     */
+    struct FEX_PACKED xstate {
+      _libc_fpstate fpstate;
+      xstate_header xstate_hdr;
+      ymmh_state ymmh;
+    };
+    static_assert(sizeof(xstate) == 944);
+
     struct FEX_PACKED ucontext_t {
       uint32_t uc_flags;
       uint32_t uc_link; // XXX: should be a compat_ptr<FEXCore::x86::ucontext_t>