Improve ethread atomics

The ethread atomics API now also provide double word size atomics.
Double word size atomics are implemented using native atomic
instructions on x86 (when the cmpxchg8b instruction is available)
and on x86_64 (when the cmpxchg16b instruction is available). On
other hardware where 32-bit atomics or word size atomics are
available, an optimized fallback is used; otherwise, a spinlock,
or a mutex based fallback is used.

The ethread library now performs runtime tests for presence of
hardware features, such as for example SSE2 instructions, instead
of requiring this to be determined at compile time.

There are now functions implementing each atomic operation with the
following implied memory barrier semantics: none, read, write,
acquire, release, and full. Some of the operation-barrier
combinations aren't especially useful. But instead of filtering
useful ones out, and potentially miss a useful one, we implement
them all.

A much smaller set of functionality for native atomics are required
to be implemented than before. More or less only cmpxchg and a
membar macro are required to be implemented for each atomic size.
Other functions will automatically be constructed from these. It is,
of course, often wise to implement more that this if possible from a
performance perspective.

1 files changed, 278 insertions, 0 deletions

diff --git a/erts/include/internal/i386/ethr_dw_atomic.h b/erts/include/internal/i386/ethr_dw_atomic.h
new file mode 100644
index 0000000000..9fb89bbe43
--- /dev/null
+++ b/erts/include/internal/i386/ethr_dw_atomic.h
@@ -0,0 +1,278 @@
+/*
+ * %CopyrightBegin%
+ *
+ * Copyright Ericsson AB 2011. All Rights Reserved.
+ *
+ * The contents of this file are subject to the Erlang Public License,
+ * Version 1.1, (the "License"); you may not use this file except in
+ * compliance with the License. You should have received a copy of the
+ * Erlang Public License along with this software. If not, it can be
+ * retrieved online at http://www.erlang.org/.
+ *
+ * Software distributed under the License is distributed on an "AS IS"
+ * basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
+ * the License for the specific language governing rights and limitations
+ * under the License.
+ *
+ * %CopyrightEnd%
+ */
+
+/*
+ * Description: Native double word atomics for x86/x86_64
+ * Author: Rickard Green
+ */
+
+#ifndef ETHR_X86_DW_ATOMIC_H__
+#define ETHR_X86_DW_ATOMIC_H__
+
+#ifdef ETHR_GCC_HAVE_DW_CMPXCHG_ASM_SUPPORT
+
+#define ETHR_HAVE_NATIVE_DW_ATOMIC
+#define ETHR_NATIVE_DW_ATOMIC_IMPL "ethread"
+
+/*
+ * If ETHR_RTCHK_USE_NATIVE_DW_ATOMIC_IMPL__ is defined, it will be used
+ * at runtime in order to determine if native or fallback implementation
+ * should be used.
+ */
+#define ETHR_RTCHK_USE_NATIVE_DW_ATOMIC_IMPL__ \
+  ETHR_X86_RUNTIME_CONF_HAVE_DW_CMPXCHG__
+
+#if ETHR_SIZEOF_PTR == 4
+typedef volatile ethr_sint64_t * ethr_native_dw_ptr_t;
+#  define ETHR_DW_NATMC_ALIGN_MASK__ 0x7
+#  define ETHR_DW_CMPXCHG_SFX__ "8b"
+#  define ETHR_NATIVE_SU_DW_SINT_T ethr_sint64_t
+#else
+#ifdef ETHR_HAVE_INT128_T
+#  define ETHR_NATIVE_SU_DW_SINT_T ethr_sint128_t
+typedef volatile ethr_sint128_t * ethr_native_dw_ptr_t;
+#else
+typedef struct {
+    ethr_sint64_t sint64[2];
+} ethr_native_sint128_t__;
+typedef volatile ethr_native_sint128_t__ * ethr_native_dw_ptr_t;
+#endif
+#  define ETHR_DW_NATMC_ALIGN_MASK__ 0xf
+#  define ETHR_DW_CMPXCHG_SFX__ "16b"
+#endif
+
+/*
+ * We need 16 byte aligned memory in 64-bit mode, and 8 byte aligned
+ * memory in 32-bit mode. 16 byte aligned malloc in 64-bit mode is
+ * not common, and at least some glibc malloc implementations
+ * only 4 byte align in 32-bit mode.
+ *
+ * This code assumes 8 byte aligned memory in 64-bit mode, and 4 byte
+ * aligned memory in 32-bit mode. A malloc implementation that does
+ * not adhere to these alignment requirements is seriously broken,
+ * and we wont bother trying to work around it.
+ *
+ * Since memory alignment may be off by one word we need to align at
+ * runtime. We, therefore, need an extra word allocated.
+ */
+#define ETHR_DW_NATMC_MEM__(VAR) \
+   (&var->c[(int) ((ethr_uint_t) &(VAR)->c[0]) & ETHR_DW_NATMC_ALIGN_MASK__])
+typedef union {
+#ifdef ETHR_NATIVE_SU_DW_SINT_T
+    volatile ETHR_NATIVE_SU_DW_SINT_T dw_sint;
+#endif
+    volatile ethr_sint_t sint[3];
+    volatile char c[ETHR_SIZEOF_PTR*3];
+} ethr_native_dw_atomic_t;
+
+
+#if (defined(ETHR_TRY_INLINE_FUNCS) \
+     || defined(ETHR_ATOMIC_IMPL__) \
+     || defined(ETHR_X86_SSE2_ASM_C__)) \
+    && ETHR_SIZEOF_PTR == 4 \
+    && defined(ETHR_GCC_HAVE_SSE2_ASM_SUPPORT)
+ethr_sint64_t
+ethr_sse2_native_su_dw_atomic_read(ethr_native_dw_atomic_t *var);
+void
+ethr_sse2_native_su_dw_atomic_set(ethr_native_dw_atomic_t *var,
+				  ethr_sint64_t val);
+#endif
+
+#if (defined(ETHR_TRY_INLINE_FUNCS) \
+     || defined(ETHR_ATOMIC_IMPL__) \
+     || defined(ETHR_X86_SSE2_ASM_C__))
+#  ifdef ETHR_DEBUG
+#    define ETHR_DW_DBG_ALIGNED__(PTR) \
+       ETHR_ASSERT((((ethr_uint_t) (PTR)) & ETHR_DW_NATMC_ALIGN_MASK__) == 0);
+#  else
+#    define ETHR_DW_DBG_ALIGNED__(PTR)
+#  endif
+#endif
+
+#if defined(ETHR_TRY_INLINE_FUNCS) || defined(ETHR_ATOMIC_IMPL__)
+
+#define ETHR_HAVE_ETHR_NATIVE_DW_ATOMIC_ADDR
+static ETHR_INLINE ethr_sint_t *
+ethr_native_dw_atomic_addr(ethr_native_dw_atomic_t *var)
+{
+    return (ethr_sint_t *) ETHR_DW_NATMC_MEM__(var);
+}
+
+#if ETHR_SIZEOF_PTR == 4 && defined(__PIC__) && __PIC__
+/*
+ * When position independent code is used in 32-bit mode, the EBX register
+ * is used for storage of global offset table address, and we may not
+ * use it as input or output in an asm. We need to save and restore the
+ * EBX register explicitly (for some reason gcc doesn't provide this
+ * service to us).
+ */
+#  define ETHR_NO_CLOBBER_EBX__ 1
+#else
+#  define ETHR_NO_CLOBBER_EBX__ 0
+#endif
+
+#if ETHR_NO_CLOBBER_EBX__ && !defined(ETHR_CMPXCHG8B_REGISTER_SHORTAGE)
+/* When no optimization is on, we'll run into a register shortage */
+#  if defined(ETHR_DEBUG) || defined(DEBUG) || defined(VALGRIND) \
+      || defined(GCOV) || defined(PURIFY) || defined(PURECOV)
+#    define ETHR_CMPXCHG8B_REGISTER_SHORTAGE 1
+#  else
+#    define ETHR_CMPXCHG8B_REGISTER_SHORTAGE 0
+#  endif
+#endif
+
+
+#define ETHR_HAVE_ETHR_NATIVE_DW_ATOMIC_CMPXCHG_MB
+
+static ETHR_INLINE int
+ethr_native_dw_atomic_cmpxchg_mb(ethr_native_dw_atomic_t *var,
+				 ethr_sint_t *new,
+				 ethr_sint_t *xchg)
+{
+    ethr_native_dw_ptr_t p = (ethr_native_dw_ptr_t) ETHR_DW_NATMC_MEM__(var);
+    char xchgd;
+
+    ETHR_DW_DBG_ALIGNED__(p);
+
+    __asm__ __volatile__(
+#if ETHR_NO_CLOBBER_EBX__
+	"pushl %%ebx\n\t"
+#  if ETHR_CMPXCHG8B_REGISTER_SHORTAGE
+	"movl (%7), %%ebx\n\t"
+	"movl 4(%7), %%ecx\n\t"
+#  else
+	"movl %8, %%ebx\n\t"
+#  endif
+#endif
+	"lock; cmpxchg" ETHR_DW_CMPXCHG_SFX__ " %0\n\t"
+	"setz %3\n\t"
+#if ETHR_NO_CLOBBER_EBX__
+	"popl %%ebx\n\t"
+#endif
+	: "=m"(*p), "=d"(xchg[1]), "=a"(xchg[0]), "=c"(xchgd)
+	: "m"(*p), "1"(xchg[1]), "2"(xchg[0]),
+#if ETHR_NO_CLOBBER_EBX__
+#  if ETHR_CMPXCHG8B_REGISTER_SHORTAGE
+	  "3"(new)
+#  else
+	  "3"(new[1]),
+	  "r"(new[0])
+#  endif
+#else
+	  "3"(new[1]),
+	  "b"(new[0])
+#endif
+	: "cc", "memory");
+
+    return (int) xchgd;
+}
+
+#undef ETHR_NO_CLOBBER_EBX__
+
+#if ETHR_SIZEOF_PTR == 4 && defined(ETHR_GCC_HAVE_SSE2_ASM_SUPPORT)
+
+typedef union {
+    ethr_sint64_t sint64;
+    ethr_sint_t sint[2];
+} ethr_dw_atomic_no_sse2_convert_t;
+
+#define ETHR_HAVE_ETHR_NATIVE_SU_DW_ATOMIC_READ
+
+static ETHR_INLINE ethr_sint64_t
+ethr_native_su_dw_atomic_read(ethr_native_dw_atomic_t *var)
+{
+    if (ETHR_X86_RUNTIME_CONF_HAVE_SSE2__)
+	return ethr_sse2_native_su_dw_atomic_read(var);
+    else {
+	ethr_sint_t new[2];
+	ethr_dw_atomic_no_sse2_convert_t xchg;
+	new[0] = new[1] = xchg.sint[0] = xchg.sint[1] = 0x83838383;
+	(void) ethr_native_dw_atomic_cmpxchg_mb(var, new, xchg.sint);
+	return xchg.sint64;
+    }
+}
+
+#define ETHR_HAVE_ETHR_NATIVE_SU_DW_ATOMIC_SET
+
+static ETHR_INLINE void
+ethr_native_su_dw_atomic_set(ethr_native_dw_atomic_t *var,
+			     ethr_sint64_t val)
+{
+    if (ETHR_X86_RUNTIME_CONF_HAVE_SSE2__)
+	ethr_sse2_native_su_dw_atomic_set(var, val);
+    else {
+	ethr_sint_t xchg[2] = {0, 0};
+	ethr_dw_atomic_no_sse2_convert_t new;
+	new.sint64 = val;
+	while (!ethr_native_dw_atomic_cmpxchg_mb(var, new.sint, xchg));
+    }
+}
+
+#endif /* ETHR_SIZEOF_PTR == 4 */
+
+#endif /* ETHR_TRY_INLINE_FUNCS */
+
+#if defined(ETHR_X86_SSE2_ASM_C__) \
+    && ETHR_SIZEOF_PTR == 4 \
+    && defined(ETHR_GCC_HAVE_SSE2_ASM_SUPPORT)
+
+/*
+ * 8-byte aligned loads and stores of 64-bit values are atomic from
+ * pentium and forward. An ordinary volatile load or store in 32-bit
+ * mode generates two 32-bit operations (at least with gcc-4.1.2 using
+ * -msse2). In order to guarantee one 64-bit load/store operation
+ * from/to memory we load/store via an xmm register using movq.
+ *
+ * Load/store can be achieved using cmpxchg8b, however, using movq is
+ * much faster. Unfortunately we cannot do the same thing in 64-bit
+ * mode; instead, we have to do loads and stores via cmpxchg16b.
+ *
+ * We do not inline these, but instead compile these into a separate
+ * object file using -msse2. This since we don't want to use -msse2 for
+ * the whole system. If we detect sse2 support (pentium4 and forward)
+ * at runtime, we use them; otherwise, we fall back to using cmpxchg8b
+ * for loads and stores. This way the binary can be moved between
+ * processors with and without sse2 support.
+ */
+
+ethr_sint64_t
+ethr_sse2_native_su_dw_atomic_read(ethr_native_dw_atomic_t *var)
+{
+    ethr_native_dw_ptr_t p = (ethr_native_dw_ptr_t) ETHR_DW_NATMC_MEM__(var);
+    ethr_sint64_t val;
+    ETHR_DW_DBG_ALIGNED__(p);
+    __asm__ __volatile__("movq %1, %0\n\t" : "=x"(val) : "m"(*p) : "memory");
+    return val;
+}
+
+void
+ethr_sse2_native_su_dw_atomic_set(ethr_native_dw_atomic_t *var,
+				  ethr_sint64_t val)
+{
+    ethr_native_dw_ptr_t p = (ethr_native_dw_ptr_t) ETHR_DW_NATMC_MEM__(var);
+    ETHR_DW_DBG_ALIGNED__(p);
+    __asm__ __volatile__("movq %1, %0\n\t" : "=m"(*p) : "x"(val) : "memory");
+}
+
+#endif /* ETHR_X86_SSE2_ASM_C__ */
+
+#endif /* ETHR_GCC_HAVE_DW_CMPXCHG_ASM_SUPPORT */
+
+#endif /* ETHR_X86_DW_ATOMIC_H__ */
+