Replace xlc __fetch_and_add() with inline asm.
authorNoah Misch <noah@leadboat.com>
Sat, 14 Sep 2019 02:34:06 +0000 (19:34 -0700)
committerNoah Misch <noah@leadboat.com>
Sat, 14 Sep 2019 02:34:06 +0000 (19:34 -0700)
PostgreSQL has been unusable when built with xlc 13 and newer, which are
incompatible with our use of __fetch_and_add().  Back-patch to 9.5,
which introduced pg_atomic_fetch_add_u32().

Reviewed by Tom Lane.

Discussion: https://postgr.es/m/20190831071157.GA3251746@rfd.leadboat.com

src/include/port/atomics/generic-xlc.h

index 5ddccff5532994ddacccd8345f9057314905e514..8b5c7329706a2a19345a763f0211d4f44fcc93e6 100644 (file)
@@ -73,11 +73,27 @@ pg_atomic_compare_exchange_u32_impl(volatile pg_atomic_uint32 *ptr,
 static inline uint32
 pg_atomic_fetch_add_u32_impl(volatile pg_atomic_uint32 *ptr, int32 add_)
 {
+   uint32 _t;
+   uint32 res;
+
    /*
-    * __fetch_and_add() emits a leading "sync" and trailing "isync", thereby
-    * providing sequential consistency.  This is undocumented.
+    * xlc has a no-longer-documented __fetch_and_add() intrinsic.  In xlc
+    * 12.01.0000.0000, it emits a leading "sync" and trailing "isync".  In
+    * xlc 13.01.0003.0004, it emits neither.  Hence, using the intrinsic
+    * would add redundant syncs on xlc 12.
     */
-   return __fetch_and_add((volatile int *)&ptr->value, add_);
+   __asm__ __volatile__(
+       "   sync                \n"
+       "   lwarx   %1,0,%4     \n"
+       "   add     %0,%1,%3    \n"
+       "   stwcx.  %0,0,%4     \n"
+       "   bne     $-12        \n"     /* branch to lwarx */
+       "   isync               \n"
+:      "=&r"(_t), "=&r"(res), "+m"(ptr->value)
+:      "r"(add_), "r"(&ptr->value)
+:      "memory", "cc");
+
+   return res;
 }
 
 #ifdef PG_HAVE_ATOMIC_U64_SUPPORT
@@ -103,7 +119,22 @@ pg_atomic_compare_exchange_u64_impl(volatile pg_atomic_uint64 *ptr,
 static inline uint64
 pg_atomic_fetch_add_u64_impl(volatile pg_atomic_uint64 *ptr, int64 add_)
 {
-   return __fetch_and_addlp((volatile long *)&ptr->value, add_);
+   uint64 _t;
+   uint64 res;
+
+   /* Like u32, but s/lwarx/ldarx/; s/stwcx/stdcx/ */
+   __asm__ __volatile__(
+       "   sync                \n"
+       "   ldarx   %1,0,%4     \n"
+       "   add     %0,%1,%3    \n"
+       "   stdcx.  %0,0,%4     \n"
+       "   bne     $-12        \n"     /* branch to ldarx */
+       "   isync               \n"
+:      "=&r"(_t), "=&r"(res), "+m"(ptr->value)
+:      "r"(add_), "r"(&ptr->value)
+:      "memory", "cc");
+
+   return res;
 }
 
 #endif /* PG_HAVE_ATOMIC_U64_SUPPORT */