From ee18fb4aacc9b67b397a0b000dda6e350ad5b3f0 Mon Sep 17 00:00:00 2001
From: Fengwei Yin <fengwei.yin@intel.com>
Date: Wed, 28 Mar 2012 17:25:17 +0800
Subject: libstdc++: Fix x86 thread-safe one-time-construction implementation.

The root of the problem is that the existing implementation is based on the
ARM C++ ABI, which mandates a different guard variable layout than the
Itanium/x86 C++ one.

This patch modifies the implementation in a way that satisfies both ABIs (and
doesn't require changing the toolchains).

Change-Id: I885e9adc7f088b9c0a78355bd752f1e6aeec9f07
Signed-off-by: Fengwei Yin <fengwei.yin@intel.com>
Signed-off-by: Jack Ren <jack.ren@intel.com>
Signed-off-by: Bruce Beare <bruce.j.beare@intel.com>
---
 libstdc++/src/one_time_construction.cpp | 92 ++++++++++++++++++++++++++-------
 1 file changed, 74 insertions(+), 18 deletions(-)

(limited to 'libstdc++')

diff --git a/libstdc++/src/one_time_construction.cpp b/libstdc++/src/one_time_construction.cpp
index 2a44c79..00adce9 100644
--- a/libstdc++/src/one_time_construction.cpp
+++ b/libstdc++/src/one_time_construction.cpp
@@ -10,46 +10,102 @@
 
 #include <stddef.h>
 #include <sys/atomics.h>
+#include <endian.h>
 #include <bionic_futex.h>
 #include <bionic_atomic_inline.h>
 
-extern "C" int __cxa_guard_acquire(int volatile * gv)
+// ARM C++ ABI and Itanium/x86 C++ ABI has different definition for
+// one time construction:
+//
+//    ARM C++ ABI defines the LSB of guard variable should be tested
+//    by compiler-generated code before calling __cxa_guard_acquire et al.
+//
+//    The Itanium/x86 C++ ABI defines the low-order _byte_ should be
+//    tested instead.
+//
+//    Meanwhile, guard variable are 32bit aligned for ARM, and 64bit
+//    aligned for x86.
+//
+// Reference documentation:
+//
+//    section 3.2.3 of ARM IHI 0041C (for ARM)
+//    section 3.3.2 of the Itanium C++ ABI specification v1.83 (for x86).
+//
+// There is no C++ ABI available for other ARCH. But the gcc source
+// shows all other ARCH follow the definition of Itanium/x86 C++ ABI.
+
+
+#if defined(__arm__)
+// The ARM C++ ABI mandates that guard variable are
+// 32-bit aligned, 32-bit values. And only its LSB is tested by
+// the compiler-generated code before calling
+// __cxa_guard_acquire.
+//
+typedef union {
+    int volatile state;
+    int32_t aligner;
+} _guard_t;
+
+const static int ready = 0x1;
+const static int pending = 0x2;
+const static int waiting = 0x6;
+
+#else   // GCC sources indicates all none-arm follow the same ABI
+// The Itanium/x86 C++ ABI mandates that guard variables
+// are 64-bit aligned, 64-bit values. Also, the least-significant
+// byte is tested by the compiler-generated code before, we calling
+// __cxa_guard_acquire. We can access it through the first
+// 32-bit word in the union below.
+//
+typedef union {
+    int volatile state;
+    int64_t aligner;
+} _guard_t;
+
+const static int ready     = letoh32(0x1);
+const static int pending   = letoh32(0x100);
+const static int waiting   = letoh32(0x10000);
+#endif
+
+extern "C" int __cxa_guard_acquire(_guard_t* gv)
 {
-    // 0 -> 2, return 1
-    // 2 -> 6, wait and return 0
-    // 6 untouched, wait and return 0
-    // 1 untouched, return 0
+    // 0 -> pending, return 1
+    // pending -> waiting, wait and return 0
+    // waiting: untouched, wait and return 0
+    // ready: untouched, return 0
+
 retry:
-    if (__atomic_cmpxchg(0, 0x2, gv) == 0) {
+    if (__atomic_cmpxchg(0, pending, &gv->state) == 0) {
         ANDROID_MEMBAR_FULL();
         return 1;
     }
-    __atomic_cmpxchg(0x2, 0x6, gv); // Indicate there is a waiter
-    __futex_wait(gv, 0x6, NULL);
+    __atomic_cmpxchg(pending, waiting, &gv->state); // Indicate there is a waiter
+    __futex_wait(&gv->state, waiting, NULL);
 
-    if(*gv != 1) // __cxa_guard_abort was called, let every thread try since there is no return code for this condition
+    if (gv->state != ready) // __cxa_guard_abort was called, let every thread try since there is no return code for this condition
         goto retry;
 
     ANDROID_MEMBAR_FULL();
     return 0;
 }
 
-extern "C" void __cxa_guard_release(int volatile * gv)
+extern "C" void __cxa_guard_release(_guard_t* gv)
 {
-    // 2 -> 1
-    // 6 -> 1, and wake
+    // pending -> ready
+    // waiting -> ready, and wake
+
     ANDROID_MEMBAR_FULL();
-    if (__atomic_cmpxchg(0x2, 0x1, gv) == 0) {
+    if (__atomic_cmpxchg(pending, ready, &gv->state) == 0) {
         return;
     }
 
-    *gv = 0x1;
-    __futex_wake(gv, 0x7fffffff);
+    gv->state = ready;
+    __futex_wake(&gv->state, 0x7fffffff);
 }
 
-extern "C" void __cxa_guard_abort(int volatile * gv)
+extern "C" void __cxa_guard_abort(_guard_t* gv)
 {
     ANDROID_MEMBAR_FULL();
-    *gv = 0;
-    __futex_wake(gv, 0x7fffffff);
+    gv->state= 0;
+    __futex_wake(&gv->state, 0x7fffffff);
 }
-- 
cgit v1.1