Merge pull request #390 from haberman/musttail

Converted fasttable decoder to use __attribute__((musttail)) when it is available

tools/amalgamate.py

@@ -47,15 +47,14 @@ class Amalgamator:
     self.output_c.write(open("upb/port_def.inc").read())
 
     self.output_h.write("/* Amalgamated source file */\n")
-    self.output_h.write('#include <stdint.h>')
     self.output_h.write(open("upb/port_def.inc").read())
 
   def add_include_path(self, path):
       self.include_paths.append(path)
 
   def finish(self):
-    self.output_c.write(open("upb/port_undef.inc").read())
-    self.output_h.write(open("upb/port_undef.inc").read())
+    self._add_header("upb/port_undef.inc")
+    self.add_src("upb/port_undef.inc")
 
   def _process_file(self, infile_name, outfile):
     file = None
@@ -69,7 +68,17 @@ class Amalgamator:
     if not file:
         raise RuntimeError("Couldn't open file " + infile_name)
 
-    for line in file:
+    lines = file.readlines()
+
+    has_copyright = lines[1].startswith(" * Copyright")
+    if has_copyright:
+      while not lines[0].startswith(" */"):
+        lines.pop(0)
+      lines.pop(0)
+
+    lines.insert(0, "\n/** " + infile_name + " " + ("*" * 60) +"/");
+
+    for line in lines:
       if not self._process_include(line, outfile):
         outfile.write(line)
 

upb/decode_internal.h

@@ -142,10 +142,11 @@ bool decode_isdone(upb_decstate *d, const char **ptr) {
   }
 }
 
+#if UPB_FASTTABLE
 UPB_INLINE
 const char *fastdecode_tagdispatch(upb_decstate *d, const char *ptr,
                                     upb_msg *msg, intptr_t table,
-                                    uint64_t hasbits, uint32_t tag) {
+                                    uint64_t hasbits, uint64_t tag) {
   const upb_msglayout *table_p = decode_totablep(table);
   uint8_t mask = table;
   uint64_t data;
@@ -153,8 +154,10 @@ const char *fastdecode_tagdispatch(upb_decstate *d, const char *ptr,
   UPB_ASSUME((idx & 7) == 0);
   idx >>= 3;
   data = table_p->fasttable[idx].field_data ^ tag;
-  return table_p->fasttable[idx].field_parser(d, ptr, msg, table, hasbits, data);
+  UPB_MUSTTAIL return table_p->fasttable[idx].field_parser(d, ptr, msg, table,
+                                                           hasbits, data);
 }
+#endif
 
 UPB_INLINE uint32_t fastdecode_loadtag(const char* ptr) {
   uint16_t tag;

upb/port_def.inc

@@ -167,6 +167,26 @@
 
 /* Configure whether fasttable is switched on or not. *************************/
 
+#if defined(__clang__) && __has_attribute(musttail)
+#define UPB_MUSTTAIL __attribute__((musttail))
+#else
+#define UPB_MUSTTAIL
+#endif
+
+/* This check is not fully robust: it does not require that we have "musttail"
+ * support available. We need tail calls to avoid consuming arbitrary amounts
+ * of stack space.
+ *
+ * GCC/Clang can mostly be trusted to generate tail calls as long as
+ * optimization is enabled, but, debug builds will not generate tail calls
+ * unless "musttail" is available.
+ *
+ * We should probably either:
+ *   1. require that the compiler supports musttail.
+ *   2. add some fallback code for when musttail isn't available (ie. return
+ *      instead of tail calling). This is safe and portable, but this comes at
+ *      a CPU cost.
+ */
 #if (defined(__x86_64__) || defined(__aarch64__)) && defined(__GNUC__)
 #define UPB_FASTTABLE_SUPPORTED 1
 #else
@@ -178,7 +198,7 @@
  * for example for testing or benchmarking. */
 #if defined(UPB_ENABLE_FASTTABLE)
 #if !UPB_FASTTABLE_SUPPORTED
-#error fasttable is x86-64 + Clang/GCC only
+#error fasttable is x86-64/ARM64 only and requires GCC or Clang.
 #endif
 #define UPB_FASTTABLE 1
 /* Define UPB_TRY_ENABLE_FASTTABLE to use fasttable if possible.

upb/port_undef.inc

@@ -52,6 +52,7 @@
 #undef UPB_SETJMP
 #undef UPB_LONGJMP
 #undef UPB_PTRADD
+#undef UPB_MUSTTAIL
 #undef UPB_FASTTABLE_SUPPORTED
 #undef UPB_FASTTABLE
 #undef UPB_FASTTABLE_INIT