Don't use compile-time constant to detect 32-bit registers:
* GCC defines __ILP32__ only in -mx32 but Clang also in -m32
* -mx32 uses 64-bit registers but doesn't define __LP64__

In file included from cm_device.cpp:31:
In file included from ./cm_device.h:37:
In file included from ./cm_array.h:31:
./cm_mem.h:200:17: error: invalid output size for constraint '=a'
                                             :"=a"(local_rax), "=r"(local_rbx),
                                                   ^
https://github.com/intel/cmrt/issues/18

--- src/cm_mem.h.orig	2016-09-07 23:51:38 UTC
+++ src/cm_mem.h
@@ -183,7 +182,7 @@ inline void GetCPUID(int CPUInfo[4], int InfoType)
 	__try {
 #endif
 
-		if (sizeof(void *) == 4) {
+#ifdef __i386__
 			unsigned int local_eax, local_ebx, local_ecx, local_edx;
 			__asm__ __volatile__("pushl %%ebx      \n\t" "cpuid  \n\t" "movl %%ebx, %1   \n\t" "popl %%ebx \n\t"	/* restore the old %ebx */
 					     :"=a"(local_eax), "=r"(local_ebx),
@@ -194,7 +193,7 @@ inline void GetCPUID(int CPUInfo[4], int InfoType)
 			CPUInfo[1] = local_ebx;
 			CPUInfo[2] = local_ecx;
 			CPUInfo[3] = local_edx;
-		} else {
+#else
 			uint64_t local_rax, local_rbx, local_rcx, local_rdx;
 			__asm__ __volatile__("push %%rbx      \n\t" "cpuid           \n\t" "mov %%rbx, %1   \n\t" "pop %%rbx       \n\t"	/* restore the old %ebx */
 					     :"=a"(local_rax), "=r"(local_rbx),
@@ -205,7 +204,7 @@ inline void GetCPUID(int CPUInfo[4], int InfoType)
 			CPUInfo[1] = local_rbx;
 			CPUInfo[2] = local_rcx;
 			CPUInfo[3] = local_rdx;
-		}
+#endif
 
 #ifndef NO_EXCEPTION_HANDLING
 	}