Workaround for QEMU behaviour. (#147)

* Fixes for ARM

* Workaround for QEMU behaviour.

src/aal/aal_arm.h

@@ -2,8 +2,14 @@
 
 #if defined(__aarch64__)
 #  define SNMALLOC_VA_BITS_64
+#  ifdef _MSC_VER
+#    include <arm64_neon.h>
+#  endif
 #else
 #  define SNMALLOC_VA_BITS_32
+#  ifdef _MSC_VER
+#    include <arm_neon.h>
+#  endif
 #endif
 
 #include <iostream>
@@ -20,13 +26,27 @@ namespace snmalloc
      */
     static constexpr uint64_t aal_features =
       IntegerPointers | NoCpuCycleCounters;
+
     /**
      * On pipelined processors, notify the core that we are in a spin loop and
      * that speculative execution past this point may not be a performance gain.
      */
     static inline void pause()
     {
+#ifdef _MSC_VER
+      __yield();
+#else
       __asm__ volatile("yield");
+#endif
+    }
+
+    static inline void prefetch(void* ptr)
+    {
+#ifdef _MSC_VER
+      __prefetch(ptr);
+#else
+      __asm__ inline("prfm pldl1keep, [%0]" : "=r"(ptr));
+#endif
     }
   };
 

src/pal/pal_linux.h

@@ -37,12 +37,16 @@ namespace snmalloc
     template<bool page_aligned = false>
     void zero(void* p, size_t size) noexcept
     {
+// QEMU does not seem to be giving the desired behaviour for MADV_DONTNEED.
+// switch back to memset only for QEMU.
+#  ifndef SNMALLOC_QEMU_WORKAROUND
       if (page_aligned || is_aligned_block<OS_PAGE_SIZE>(p, size))
       {
         SNMALLOC_ASSERT(is_aligned_block<OS_PAGE_SIZE>(p, size));
         madvise(p, size, MADV_DONTNEED);
       }
       else
+#  endif
       {
         ::memset(p, 0, size);
       }