Generalize crash message for non-ok status. (#9552)

Author: ysiraichi

torch_xla/csrc/init_python_bindings.cpp

@@ -3377,17 +3377,9 @@ void InitXlaModuleBindings(py::module m) {
            [](const std::vector<at::Tensor>& tensors) -> py::bytes {
              absl::StatusOr<std::vector<absl_nonnull XLATensorPtr>>
                  xtensors_status = bridge::GetXlaTensors(tensors);
-             ABSL_CHECK(xtensors_status.ok())
-                 << "\n\n"
-                 << "Internal Error:\n"
-                 << "    _get_graph_hash(): error retrieving the XLA tensors "
-                    "from the given tensor arguments. "
-                 << "This is a bug! Please, open an issue in the PyTorch/XLA "
-                 << "GitHub repository: https://github.com/pytorch/xla"
-                 << "\n\n"
-                 << "Status Error:\n"
-                 << "    " << BuildStatusErrorMessage(xtensors_status.status())
-                 << "\n";
+             XLA_CHECK_OK(xtensors_status,
+                          "_get_graph_hash(): error retrieving the XLA tensors "
+                          "from the given tensor arguments.");
              std::vector<absl_nonnull XLATensorPtr> xtensors =
                  xtensors_status.value();
              torch::lazy::hash_t hash =

torch_xla/csrc/runtime/debug_macros.h

@@ -28,7 +28,6 @@
 // unnecessary or undesirable.
 #define XLA_ERROR() TF_ERROR_STREAM()
 #define XLA_CHECK(c) TF_CHECK(c)
-#define XLA_CHECK_OK(c) TF_CHECK_OK(c)
 #define XLA_CHECK_EQ(a, b) TF_CHECK_EQ(a, b)
 #define XLA_CHECK_NE(a, b) TF_CHECK_NE(a, b)
 #define XLA_CHECK_LE(a, b) TF_CHECK_LE(a, b)

torch_xla/csrc/status.cpp

@@ -124,4 +124,30 @@ void MaybeThrow(const absl::Status& status) {
 
 void GetValueOrThrow(const absl::Status& status) { MaybeThrow(status); }
 
+void OkOrDie(const absl::Status& status, const char* file, const int32_t line,
+             const char* function, std::string_view message) {
+  if (status.ok()) {
+    return;
+  }
+
+  std::ostringstream oss;
+  oss << "\n\n"
+      << "Internal Error:\n";
+
+  if (!message.empty()) {
+    oss << "    " << message << "\n";
+  }
+
+  oss << "    This is a bug! Please, open an issue in the PyTorch/XLA "
+      << "GitHub repository: https://github.com/pytorch/xla"
+      << "\n\n"
+      << "Status Error:\n"
+      << "    "
+      << BuildStatusErrorMessage(
+             status_internal::MaybeWithNewMessage(status, file, line, function))
+      << "\n";
+
+  ABSL_CHECK(status.ok()) << oss.str();
+}
+
 }  // namespace torch_xla

torch_xla/csrc/status.h

@@ -10,6 +10,8 @@
 #ifndef XLA_TORCH_XLA_CSRC_STATUS_H_
 #define XLA_TORCH_XLA_CSRC_STATUS_H_
 
+#include <sstream>
+
 #include "absl/status/statusor.h"
 
 namespace torch_xla {
@@ -125,6 +127,22 @@ constexpr char kStatusPropagationTraceKey[] =
                             lhs = std::move(XLA_STATUS_VAR_).value(), \
                             ##__VA_ARGS__)
 
+// Crashes if `status` is not an ok status.
+//
+// Example:
+//
+//     XLA_CHECK_OK(
+//         FnThatReturnStatus(),
+//         "New error message"
+//     );
+//
+// If `FnThatReturnStatus()` returns a non-ok status, this macro will
+// call `ABSL_CHECK()`, which will crash.
+//
+#define XLA_CHECK_OK(status, ...)                                       \
+  ::torch_xla::OkOrDie(::torch_xla::status_internal::GetStatus(status), \
+                       __FILE__, __LINE__, __FUNCTION__, ##__VA_ARGS__)
+
 namespace status_internal {
 
 // Adds source location information to the status propagation trace if
@@ -211,6 +229,14 @@ T GetValueOrThrow(absl::StatusOr<T>&& status) {
 // `GetValueOrThrow` overload for `Status`.
 void GetValueOrThrow(const absl::Status& status);
 
+// Checks that `status` is an ok status.
+//
+// Otherwise, it will create a new status instance with the given source
+// location information, and incorporate its message (alongside the
+// status propagation trace) to the crash report.
+void OkOrDie(const absl::Status& status, const char* file, const int32_t line,
+             const char* function, std::string_view message = "");
+
 }  // namespace torch_xla
 
 #endif  // XLA_TORCH_XLA_CSRC_STATUS_H_