Support stack walk on Windows ARM64 in base profiler

Bug: 893460
Change-Id: I9def3ef624b5121c7e902efc01a18754f448b8ce
Reviewed-on: https://chromium-review.googlesource.com/c/1332698Reviewed-by: Mike Wittman <wittman@chromium.org>
Reviewed-by: Daniel Cheng <dcheng@chromium.org>
Reviewed-by: Bruce Dawson <brucedawson@chromium.org>
Commit-Queue: Tom Tan <Tom.Tan@microsoft.com>
Cr-Commit-Position: refs/heads/master@{#608119}

base/BUILD.gn

@@ -2161,7 +2161,7 @@ if (is_win) {
 }
 
 if (is_win || is_mac) {
-  if (current_cpu == "x64") {
+  if (current_cpu == "x64" || (current_cpu == "arm64" && is_win)) {
     # Must be a shared library so that it can be unloaded during testing.
     shared_library("base_profiler_test_support_library") {
       sources = [

base/allocator/allocator_shim_unittest.cc

@@ -24,6 +24,7 @@
 
 #if defined(OS_WIN)
 #include <windows.h>
+#include <malloc.h>
 #elif defined(OS_MACOSX)
 #include <malloc/malloc.h>
 #include "base/allocator/allocator_interception_mac.h"

base/profiler/native_stack_sampler_win.cc

@@ -24,6 +24,7 @@
 #include "base/stl_util.h"
 #include "base/time/time.h"
 #include "base/trace_event/trace_event.h"
+#include "build/build_config.h"
 
 namespace base {
 
@@ -157,10 +158,19 @@ void RewritePointersToStackMemory(uintptr_t top,
                                   uintptr_t bottom,
                                   CONTEXT* context,
                                   void* stack_copy) {
-#if defined(_WIN64)
+#if defined(ARCH_CPU_64_BITS)
   DWORD64 CONTEXT::*const nonvolatile_registers[] = {
+#if defined(ARCH_CPU_X86_64)
       &CONTEXT::R12, &CONTEXT::R13, &CONTEXT::R14, &CONTEXT::R15, &CONTEXT::Rdi,
-      &CONTEXT::Rsi, &CONTEXT::Rbx, &CONTEXT::Rbp, &CONTEXT::Rsp};
+      &CONTEXT::Rsi, &CONTEXT::Rbx, &CONTEXT::Rbp, &CONTEXT::Rsp
+#elif defined(ARCH_CPU_ARM64)
+      &CONTEXT::X19, &CONTEXT::X20, &CONTEXT::X21, &CONTEXT::X22, &CONTEXT::X23,
+      &CONTEXT::X24, &CONTEXT::X25, &CONTEXT::X26, &CONTEXT::X27, &CONTEXT::X28,
+      &CONTEXT::Fp, &CONTEXT::Lr
+#else
+#error Unsupported Windows 64-bit Arch
+#endif
+  };
 
   // Rewrite pointers in the context.
   for (size_t i = 0; i < size(nonvolatile_registers); ++i) {
@@ -213,9 +223,9 @@ NativeStackSamplerError RecordStack(CONTEXT* context,
   stack->reserve(128);
 
   Win32StackFrameUnwinder frame_unwinder;
-  while (context->Rip) {
+  while (ContextPC(context)) {
     const void* instruction_pointer =
-        reinterpret_cast<const void*>(context->Rip);
+        reinterpret_cast<const void*>(ContextPC(context));
     ScopedModuleHandle module;
     if (!frame_unwinder.TryUnwind(context, &module))
       return NATIVE_STACK_SAMPLER_TRY_UNWIND_FAILED;
@@ -355,8 +365,10 @@ NativeStackSamplerError SuspendThreadAndRecordStack(
       if (!::GetThreadContext(thread_handle, &thread_context))
         return NATIVE_STACK_SAMPLER_GET_THREAD_CONTEXT_FAILED;
 
-#if defined(_WIN64)
+#if defined(ARCH_CPU_X86_64)
       bottom = thread_context.Rsp;
+#elif defined(ARCH_CPU_ARM64)
+      bottom = thread_context.Sp;
 #else
       bottom = thread_context.Esp;
 #endif

base/profiler/win32_stack_frame_unwinder.cc

@@ -9,6 +9,7 @@
 #include <utility>
 
 #include "base/macros.h"
+#include "build/build_config.h"
 
 namespace base {
 
@@ -121,7 +122,7 @@ bool Win32StackFrameUnwinder::TryUnwind(CONTEXT* context,
   // TODO(chengx): update base::ModuleCache to return a ScopedModuleHandle and
   // use it for this module lookup.
   ScopedModuleHandle frame_module =
-      unwind_functions_->GetModuleForProgramCounter(context->Rip);
+      unwind_functions_->GetModuleForProgramCounter(ContextPC(context));
   if (!frame_module.IsValid()) {
     // There's no loaded module containing the instruction pointer. This can be
     // due to executing code that is not in a module. In particular,
@@ -142,20 +143,31 @@ bool Win32StackFrameUnwinder::TryUnwind(CONTEXT* context,
   ULONG64 image_base;
   // Try to look up unwind metadata for the current function.
   PRUNTIME_FUNCTION runtime_function =
-      unwind_functions_->LookupFunctionEntry(context->Rip, &image_base);
+      unwind_functions_->LookupFunctionEntry(ContextPC(context), &image_base);
 
   if (runtime_function) {
-    unwind_functions_->VirtualUnwind(image_base, context->Rip, runtime_function,
-                                     context);
+    unwind_functions_->VirtualUnwind(image_base, ContextPC(context),
+                                     runtime_function, context);
     at_top_frame_ = false;
   } else {
     if (at_top_frame_) {
       at_top_frame_ = false;
 
       // This is a leaf function (i.e. a function that neither calls a function,
-      // nor allocates any stack space itself) so the return address is at RSP.
+      // nor allocates any stack space itself).
+#if defined(ARCH_CPU_X86_64)
+      // For X64, return address is at RSP.
       context->Rip = *reinterpret_cast<DWORD64*>(context->Rsp);
       context->Rsp += 8;
+#elif defined(ARCH_CPU_ARM64)
+      // For leaf function on Windows ARM64, return address is at LR(X30).
+      // Add CONTEXT_UNWOUND_TO_CALL flag to avoid unwind ambiguity for tailcall
+      // on ARM64, because padding after tailcall is not guaranteed.
+      context->Pc = context->Lr;
+      context->ContextFlags |= CONTEXT_UNWOUND_TO_CALL;
+#else
+#error Unsupported Windows 64-bit Arch
+#endif
     } else {
       // In theory we shouldn't get here, as it means we've encountered a
       // function without unwind information below the top of the stack, which

base/profiler/win32_stack_frame_unwinder.h

@@ -12,6 +12,7 @@
 #include "base/base_export.h"
 #include "base/macros.h"
 #include "base/win/scoped_handle.h"
+#include "build/build_config.h"
 
 namespace base {
 
@@ -25,6 +26,18 @@ struct RUNTIME_FUNCTION {
 using PRUNTIME_FUNCTION = RUNTIME_FUNCTION*;
 #endif  // !defined(_WIN64)
 
+#if defined(ARCH_CPU_64_BITS)
+inline ULONG64 ContextPC(CONTEXT* context) {
+#if defined(ARCH_CPU_X86_64)
+  return context->Rip;
+#elif defined(ARCH_CPU_ARM64)
+  return context->Pc;
+#else
+#error Unsupported Windows 64-bit Arch
+#endif
+}
+#endif
+
 // Traits class to adapt GenericScopedHandle for HMODULES.
 class ModuleHandleTraits : public win::HandleTraits {
  public: