[compiler] Implement 64-bit atomics.

* The mux device capabilities provide a way to indicate which widths are
  supported for atomic operations. Currently, atomic operations of 8,
  16, or 32 bits are always presumed supported. Atomic operations of 64
  bits are presumed supported on targets with 64-bit pointers.
* The OpenCL extensions cl_khr_int64_base_atomics and
  cl_khr_int64_extended_atomics are implemented on targets that support
  64-bit atomic operations. Note that despite the name, in accordance
  with the OpenCL specification, these extensions also cover atomic
  operations on type 'double' when that type is supported.

doc/specifications/mux-runtime-spec.rst

@@ -186,6 +186,7 @@ in the system, but it can still be useful as a target for compilation.
      uint32_t allocation_capabilities;
      uint32_t source_capabilities;
      uint32_t address_capabilities;
+     uint32_t atomic_capabilities;
      uint32_t cache_capabilities;
      uint32_t half_capabilities;
      uint32_t float_capabilities;
@@ -243,6 +244,8 @@ in the system, but it can still be useful as a target for compilation.
    ``mux_source_capabilities_e``.
 -  ``address_capabilities`` - a bitfield of
    ``mux_address_capabilities_e``.
+-  ``atomic_capabilities`` - a bitfield of
+   ``mux_atomic_capabilities_e``.
 -  ``cache_capabilities`` - a bitfield of ``mux_cache_capabilities_e``.
 -  ``half_capabilities`` - half floating point support, a bitfield of
    ``mux_floating_point_capabilities_e``.

modules/compiler/builtins/scripts/generate_c11_atomic_builtins.py

@@ -23,8 +23,11 @@
 
 type_pairs = {
     'atomic_int': 'int',
+    'atomic_long': 'long',
     'atomic_uint': 'uint',
+    'atomic_ulong': 'ulong',
     'atomic_float': 'float',
+    'atomic_double': 'double',
 }
 
 address_spaces = ['__local', '__global']

modules/compiler/builtins/scripts/generate_header_30.sh

@@ -377,53 +377,63 @@ function all_atomic()
   echo ""
   for k in __local __global ""
   do
-    for i in int uint float
+    for i in int long uint ulong float double
     do
+      double_support_begin $i
       echo "void __CL_BUILTIN_ATTRIBUTES atomic_init(volatile $k atomic_$i *obj, $i value);"
+      double_support_end $i
     done
   done
   echo "void __CL_BUILTIN_ATTRIBUTES atomic_work_item_fence(cl_mem_fence_flags flags, memory_order order, memory_scope scope);"
 
   for k in __local __global ""
   do
-    for i in int uint float
+    for i in int long uint ulong float double
     do
+      double_support_begin $i
       echo "void __CL_BUILTIN_ATTRIBUTES atomic_store_explicit(volatile $k atomic_$i *object, $i desired, memory_order order, memory_scope scope);"
+      double_support_end $i
     done
   done
 
   for k in __local __global ""
   do
-    for i in int uint float
+    for i in int long uint ulong float double
     do
+      double_support_begin $i
       echo "$i __CL_BUILTIN_ATTRIBUTES atomic_load_explicit(volatile $k atomic_$i *object, memory_order order, memory_scope scope);"
+      double_support_end $i
     done
   done
 
   for k in __local __global ""
   do
-    for i in int uint float
+    for i in int long uint ulong float double
     do
+      double_support_begin $i
       echo "$i __CL_BUILTIN_ATTRIBUTES atomic_exchange_explicit(volatile $k atomic_$i *object, $i desired, memory_order order, memory_scope scope);"
+      double_support_end $i
     done
   done
 
   for name in atomic_compare_exchange_strong_explicit atomic_compare_exchange_weak_explicit
   do
     for k1 in __local __global ""
     do
-      for i in int uint float
+      for i in int long uint ulong float double
       do
+        double_support_begin $i
         for k2 in __local __global __private ""
         do
           echo "bool __CL_BUILTIN_ATTRIBUTES $name(volatile $k1 atomic_$i *object, $k2 $i *expected, $i desired, memory_order success, memory_order failure);"
           echo "bool __CL_BUILTIN_ATTRIBUTES $name(volatile $k1 atomic_$i *object, $k2 $i *expected, $i desired, memory_order success, memory_order failure, memory_scope scope);"
         done
+        double_support_end $i
       done
     done
   done
 
-  for t in int uint
+  for t in int long uint ulong
   do
     for op in add sub or xor and min max
     do

modules/compiler/source/base/source/module.cpp

@@ -975,11 +975,13 @@ void BaseModule::addDefaultOpenCLPreprocessorOpts(
   }
 
   // Disable `cl_khr_int64_base_atomics` and `cl_khr_int64_extended_atomics`
-  // until we support them. (CA-518)
-  addOpenCLOpt("-cl_khr_int64_base_atomics", opencl_opts);
-  addMacroUndef("cl_khr_int64_base_atomics", macro_defs);
-  addOpenCLOpt("-cl_khr_int64_extended_atomics", opencl_opts);
-  addMacroUndef("cl_khr_int64_extended_atomics", macro_defs);
+  // unless supported by the device.
+  if (!(device_info->atomic_capabilities & mux_atomic_capabilities_64bit)) {
+    addOpenCLOpt("-cl_khr_int64_base_atomics", opencl_opts);
+    addMacroUndef("cl_khr_int64_base_atomics", macro_defs);
+    addOpenCLOpt("-cl_khr_int64_extended_atomics", opencl_opts);
+    addMacroUndef("cl_khr_int64_extended_atomics", macro_defs);
+  }
 
   if (options.standard == Standard::OpenCLC30) {
     // work-group collective functions are an optional feature in OpenCL 3.0.