diff --git a/runtime/Makefile b/runtime/Makefile
index f8cf33e..859ea76 100644
--- a/runtime/Makefile
+++ b/runtime/Makefile
@@ -16,7 +16,7 @@ CFILES += ${RTDIR}/arch/${ARCH}/*.c
 #CFLAGS += -DDEBUG 
 CFLAGS += -D_GNU_SOURCE
 #CFLAGS += -DLOG_TO_FILE
-CFLAGS += -DUSE_HTTP_UVIO #-DUSE_HTTP_SYNC
+#CFLAGS += -DUSE_HTTP_UVIO #-DUSE_HTTP_SYNC
 #CFLAGS += -DUSE_SYSCALL
 #CFLAGS += -DPREEMPT_DISABLE
 CACHE_LINESIZE := $(shell getconf LEVEL1_DCACHE_LINESIZE)
diff --git a/runtime/src/arch/aarch64/env.c b/runtime/src/arch/aarch64/env.c
index 334288d..3949e82 100644
--- a/runtime/src/arch/aarch64/env.c
+++ b/runtime/src/arch/aarch64/env.c
@@ -10,10 +10,4 @@ __getcycles(void)
 	return virtual_timer_value;
 }
 
-INLINE unsigned long long
-env_getcycles(void)
-{
-        return __getcycles();
-}
-
 #endif
diff --git a/runtime/src/arch/x86_64/env.c b/runtime/src/arch/x86_64/env.c
index fec02f7..3ec3dcd 100644
--- a/runtime/src/arch/x86_64/env.c
+++ b/runtime/src/arch/x86_64/env.c
@@ -2,158 +2,6 @@
 
 #include <runtime.h>
 
-// Atomic functions, with definitions stolen from musl
-
-i32
-env_a_ctz_64(u64 x)
-{
-        __asm__("bsf %1,%0" : "=r"(x) : "r"(x));
-
-        return x;
-}
-
-// static inline int a_ctz_l(unsigned long x)
-//{
-//      __asm__( "bsf %1,%0" : "=r"(x) : "r"(x) );
-//      return x;
-//}
-
-INLINE void
-env_a_and_64(i32 p_off, u64 v)
-{
-        uint64_t *p = worker_thread_get_memory_ptr_void(p_off, sizeof(uint64_t));
-        *p &= v;
-
-        //      __asm__( "lock ; and %1, %0"
-        //                       : "=m"(*p) : "r"(v) : "memory" );
-}
-
-INLINE void
-env_a_or_64(i32 p_off, i64 v)
-{
-        assert(sizeof(i64) == sizeof(uint64_t));
-        uint64_t *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i64));
-        *p |= v;
-
-        //      __asm__( "lock ; or %1, %0"
-        //                       : "=m"(*p) : "r"(v) : "memory" );
-}
-
-// static inline void a_or_l(volatile void *p, long v)
-//{
-//      __asm__( "lock ; or %1, %0"
-//              : "=m"(*(long *)p) : "r"(v) : "memory" );
-//}
-//
-// static inline void *a_cas_p(volatile void *p, void *t, void *s)
-//{
-//      __asm__( "lock ; cmpxchg %3, %1"
-//              : "=a"(t), "=m"(*(long *)p) : "a"(t), "r"(s) : "memory" );
-//      return t;
-//}
-//
-i32
-env_a_cas(i32 p_off, i32 t, i32 s)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i32));
-
-        __asm__("lock ; cmpxchg %3, %1" : "=a"(t), "=m"(*p) : "a"(t), "r"(s) : "memory");
-
-        return t;
-}
-
-void
-env_a_or(i32 p_off, i32 v)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i32));
-        __asm__("lock ; or %1, %0" : "=m"(*p) : "r"(v) : "memory");
-}
-
-// static inline void a_and(volatile int *p, int v)
-//{
-//      __asm__( "lock ; and %1, %0"
-//              : "=m"(*p) : "r"(v) : "memory" );
-//}
-
-
-i32
-env_a_swap(i32 x_off, i32 v)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
-
-        __asm__("xchg %0, %1" : "=r"(v), "=m"(*x) : "0"(v) : "memory");
-
-        return v;
-}
-
-i32
-env_a_fetch_add(i32 x_off, i32 v)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
-
-        __asm__("lock ; xadd %0, %1" : "=r"(v), "=m"(*x) : "0"(v) : "memory");
-
-        return v;
-}
-
-void
-env_a_inc(i32 x_off)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
-
-        __asm__("lock ; incl %0" : "=m"(*x) : "m"(*x) : "memory");
-}
-
-void
-env_a_dec(i32 x_off)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
-
-        __asm__("lock ; decl %0" : "=m"(*x) : "m"(*x) : "memory");
-}
-
-
-void
-env_a_store(i32 p_off, i32 x)
-{
-        assert(sizeof(i32) == sizeof(volatile int));
-        volatile int *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i32));
-        __asm__ __volatile__("mov %1, %0 ; lock ; orl $0,(%%esp)" : "=m"(*p) : "r"(x) : "memory");
-}
-
-void
-env_do_spin(i32 i)
-{
-        __asm__ __volatile__("pause" : : : "memory");
-}
-
-void
-env_do_crash(i32 i)
-{
-        printf("crashing: %d\n", i);
-        assert(0);
-}
-
-int
-env_a_ctz_32(i32 x)
-{
-        __asm__("bsf %1,%0" : "=r"(x) : "r"(x));
-
-        return x;
-}
-
-void
-env_do_barrier(i32 x)
-{
-        __asm__ __volatile__("" : : : "memory");
-}
-
 unsigned long long int
 __getcycles(void)
 {
@@ -166,10 +14,4 @@ __getcycles(void)
         return cpu_time_in_cycles;
 }
 
-INLINE unsigned long long
-env_getcycles(void)
-{
-        return __getcycles();
-}
-
 #endif
diff --git a/runtime/src/env.c b/runtime/src/env.c
index 128b9e6..93c0a6e 100644
--- a/runtime/src/env.c
+++ b/runtime/src/env.c
@@ -17,13 +17,129 @@ env___syscall(i32 n, i32 a, i32 b, i32 c, i32 d, i32 e, i32 f)
 	return env_syscall_handler(n, a, b, c, d, e, f);
 }
 
-
 void
 env___unmapself(u32 base, u32 size)
 {
 	// Just do some no op
 }
 
+i32
+env_a_ctz_64(u64 x)
+{
+	return __builtin_ctzll(x);
+}
+
+INLINE void
+env_a_and_64(i32 p_off, u64 v)
+{
+	uint64_t *p = worker_thread_get_memory_ptr_void(p_off, sizeof(uint64_t));
+	*p &= v;
+}
+
+INLINE void
+env_a_or_64(i32 p_off, i64 v)
+{
+	assert(sizeof(i64) == sizeof(uint64_t));
+	uint64_t *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i64));
+	*p |= v;
+}
+
+i32
+env_a_cas(i32 p_off, i32 t, i32 s)
+{
+	assert(sizeof(i32) == sizeof(volatile int));
+	volatile int *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i32));
+
+	return __sync_val_compare_and_swap(p, t, s);
+}
+
+void
+env_a_or(i32 p_off, i32 v)
+{
+        assert(sizeof(i32) == sizeof(volatile int));
+        volatile int *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i32));
+        //__asm__("lock ; or %1, %0" : "=m"(*p) : "r"(v) : "memory");
+	*p |= v;
+}
+
+i32
+env_a_swap(i32 x_off, i32 v)
+{
+	assert(sizeof(i32) == sizeof(volatile int));
+	volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
+
+	//__asm__("xchg %0, %1" : "=r"(v), "=m"(*x) : "0"(v) : "memory");
+	int t = *x;
+	*x = v;
+	v = t;
+
+	return v;
+}
+
+i32
+env_a_fetch_add(i32 x_off, i32 v)
+{
+	assert(sizeof(i32) == sizeof(volatile int));
+	volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
+
+	//__asm__("lock ; xadd %0, %1" : "=r"(v), "=m"(*x) : "0"(v) : "memory");
+	return __sync_fetch_and_add(x, v);
+}
+
+void
+env_a_inc(i32 x_off)
+{
+	assert(sizeof(i32) == sizeof(volatile int));
+	volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
+	(*x)++;
+
+	//__asm__("lock ; incl %0" : "=m"(*x) : "m"(*x) : "memory");
+}
+
+void
+env_a_dec(i32 x_off)
+{
+	assert(sizeof(i32) == sizeof(volatile int));
+	volatile int *x = worker_thread_get_memory_ptr_void(x_off, sizeof(i32));
+	(*x)--;
+
+	//__asm__("lock ; decl %0" : "=m"(*x) : "m"(*x) : "memory");
+}
+
+void
+env_a_store(i32 p_off, i32 x)
+{
+	assert(sizeof(i32) == sizeof(volatile int));
+	volatile int *p = worker_thread_get_memory_ptr_void(p_off, sizeof(i32));
+	*p = x;
+	//__asm__ __volatile__("mov %1, %0 ; lock ; orl $0,(%%esp)" : "=m"(*p) : "r"(x) : "memory");
+}
+
+int
+env_a_ctz_32(i32 x)
+{
+	return __builtin_ctz(x);
+}
+
+void
+env_do_spin(i32 i)
+{
+	printf("nope! not happening: %d\n", i);
+	assert(0);
+}
+
+void
+env_do_crash(i32 i)
+{
+	printf("crashing: %d\n", i);
+	assert(0);
+}
+
+void
+env_do_barrier(i32 x)
+{
+	__sync_synchronize();
+}
 
 // Floating point routines
 INLINE double
@@ -38,3 +154,9 @@ env_cos(double d)
 	return cos(d);
 }
 
+INLINE unsigned long long
+env_getcycles(void)
+{
+	return __getcycles();
+}
+
diff --git a/runtime/tests/Makefile.inc b/runtime/tests/Makefile.inc
index 4c9e451..6a15ce2 100644
--- a/runtime/tests/Makefile.inc
+++ b/runtime/tests/Makefile.inc
@@ -20,7 +20,7 @@ RT_DIR=${BASE_DIR}/runtime/
 RT_MEM=${RT_DIR}/memory/
 RT_LIBC=${RT_DIR}/libc/libc_backing.c
 RT_RT=${RT_DIR}/runtime.c
-RT_ENV=${RT_DIR}/libc/env_${ARCH}.c
+RT_ENV=${RT_DIR}/libc/env.c ${RT_DIR}/libc/env_${ARCH}.c
 
 MEMC=${RT_MEM}/${MEMC_64}
 DUMMY=${BASE_DIR}/code_benches/dummy.c
diff --git a/silverfish b/silverfish
index 54d0987..e270b5e 160000
--- a/silverfish
+++ b/silverfish
@@ -1 +1 @@
-Subproject commit 54d098796036ee625688ec0f2796c07f5aecc89c
+Subproject commit e270b5edb24cf3a7d9cdc162eb4d6945cdabf1b5