Change XGETBV architecture modeling to hypercall

Author: mrexodia

include/remill/Arch/Runtime/HyperCall.h

@@ -30,6 +30,7 @@ class SyncHyperCall {
     kX86CPUID,
     kX86ReadTSC,
     kX86ReadTSCP,
+    kX86XGetBV,
     kX86LoadGlobalDescriptorTable,
     kX86LoadInterruptDescriptorTable,
     kX86LoadAccessRights,

include/remill/Arch/X86/Runtime/State.h

@@ -445,34 +445,6 @@ struct alignas(8) ArithFlags final {
 
 static_assert(16 == sizeof(ArithFlags), "Invalid packing of `ArithFlags`.");
 
-union XCR0 {
-  uint64_t flat;
-
-  struct {
-    uint32_t eax;
-    uint32_t edx;
-  } __attribute__((packed));
-
-  // Bits specify what process states should be saved.
-  struct {
-    uint64_t x87_fpu_mmx : 1;  // Must be 1; bit 0.
-    uint64_t xmm : 1;  // SSE.
-    uint64_t ymm : 1;  // AVX and AVX2.
-    uint64_t bndreg : 1;  // Part of MPX.
-    uint64_t bndcsr : 1;  // Part of MPX.
-    uint64_t opmask : 1;  // Registers k0 through k7, AVX512-only.
-    uint64_t
-        zmm_hi256 : 1;  // High 256 bits of ZMM0 through ZMM15, AVX512-only.
-    uint64_t hi16_zmm : 1;  // ZMM16 through ZMM31, AVX512-only.
-    uint64_t pkru : 1;  // Protected key stuff.
-    uint64_t _reserved0 : 53;
-    uint64_t lwp : 1;  // AMD lightweight profiling.
-    uint64_t _reserved1 : 1;
-  } __attribute__((packed));
-} __attribute__((packed));
-
-static_assert(8 == sizeof(XCR0), "Invalid packing of `XCR0`.");
-
 struct alignas(8) Segments final {
   volatile uint16_t _0;
   SegmentSelector ss;
@@ -794,7 +766,7 @@ struct alignas(16) X86State : public ArchState {
   X87Stack st;  // 128 bytes.
   MMX mmx;  // 128 bytes.
   FPUStatusFlags sw;  // 24 bytes
-  XCR0 xcr0;  // 8 bytes.
+  uint8_t _padding[8]; // 8 bytes
   FPU x87;  // 512 bytes
   SegmentCaches seg_caches;  // 96 bytes
   K_REG k_reg; // 128 bytes.

lib/Arch/Runtime/HyperCall.cpp

@@ -89,6 +89,16 @@ Memory *__remill_sync_hyper_call(State &state, Memory *mem,
                      "d"(state.gpr.rdx.aword));
       break;
 
+    case SyncHyperCall::kX86XGetBV: {
+      uint32_t eax = 0;
+      uint32_t edx = 0;
+      const auto selector = static_cast<uint32_t>(state.addr_to_load);
+      asm volatile("xgetbv" : "=a"(eax), "=d"(edx) : "c"(selector));
+      state.gpr.rax.aword = eax;
+      state.gpr.rdx.aword = edx;
+      break;
+    }
+
     case SyncHyperCall::kX86LoadGlobalDescriptorTable: {
       const auto read =
           __remill_read_memory_64(mem, static_cast<addr_t>(state.addr_to_load));

lib/Arch/X86/Semantics/XSAVE.cpp

@@ -16,33 +16,9 @@
 
 namespace {
 
-DEF_SEM(DoXGETBV, PC next_pc) {
-  switch (Read(REG_ECX)) {
-
-    // Current state of the `xcr0` register.
-    case 0:
-      WriteZExt(IF_64BIT_ELSE(REG_RAX, REG_EAX), state.xcr0.eax);
-      WriteZExt(IF_64BIT_ELSE(REG_RDX, REG_EDX), state.xcr0.edx);
-      break;
-
-    // Current state of the `xcr0` register, anded with the `xinuse` register.
-    // We fake this as saying: this is what
-    case 1: {
-      XCR0 xcr0 = {};
-      xcr0.x87_fpu_mmx = 1;
-      xcr0.xmm = 1;
-      IF_AVX(xcr0.ymm = 1;)
-      IF_AVX512(xcr0.opmask = 1;)
-      IF_AVX512(xcr0.zmm_hi256 = 1;)
-      IF_AVX512(xcr0.hi16_zmm = 1;)
-      WriteZExt(IF_64BIT_ELSE(REG_RAX, REG_EAX), xcr0.eax);
-      WriteZExt(IF_64BIT_ELSE(REG_RDX, REG_EDX), xcr0.edx);
-      break;
-    }
-
-    default: WriteZExt(REG_PC, Read(next_pc)); StopFailure();
-  }
-  return memory;
+DEF_SEM(DoXGETBV, PC) {
+  state.addr_to_load = ZExtTo<uint64_t>(Read(REG_ECX));
+  return __remill_sync_hyper_call(state, memory, SyncHyperCall::kX86XGetBV);
 }
 
 }  // namespace

tests/X86/Run.cpp

@@ -914,26 +914,6 @@ static void RunWithFlags(const test::TestInfo *info, Flags flags,
   ImportX87State(native_state);
   ResetFlags();
 
-#if defined(__x86_64__) || defined(__i386__)
-  {
-    uint32_t eax = 1;
-    uint32_t ebx = 0;
-    uint32_t ecx = 0;
-    uint32_t edx = 0;
-    asm volatile("cpuid"
-                 : "=a"(eax), "=b"(ebx), "=c"(ecx), "=d"(edx)
-                 : "a"(eax), "b"(ebx), "c"(ecx), "d"(edx));
-    if (ecx & (1U << 26U)) {
-      uint32_t xcr0_lo = 0;
-      uint32_t xcr0_hi = 0;
-      asm volatile("xgetbv" : "=a"(xcr0_lo), "=d"(xcr0_hi) : "c"(0));
-      native_state->xcr0.eax = xcr0_lo;
-      native_state->xcr0.edx = xcr0_hi;
-      lifted_state->xcr0 = native_state->xcr0;
-    }
-  }
-#endif
-
   // Set up the RIP correctly.
   lifted_state->gpr.rip.aword = static_cast<addr_t>(info->test_begin);
   native_state->gpr.rip.aword = static_cast<addr_t>(info->test_end);
@@ -1206,7 +1186,6 @@ static void RunWithFlags(const test::TestInfo *info, Flags flags,
           << "st:" << offsetof(State, st) << "\n"
           << "mmx:" << offsetof(State, mmx) << "\n"
           << "sw:" << offsetof(State, sw) << "\n"
-          << "xcr0:" << offsetof(State, xcr0) << "\n"
           << "x87:" << offsetof(State, x87) << "\n"
           << "seg_caches:" << offsetof(State, seg_caches) << "\n";
     }