diff --git a/src/codegen/target/arm/assembler.cpp b/src/codegen/target/arm/assembler.cpp
index 23b07ef201..b4046223e1 100644
--- a/src/codegen/target/arm/assembler.cpp
+++ b/src/codegen/target/arm/assembler.cpp
@@ -946,11 +946,20 @@ class MyAssembler : public Assembler {
           unsigned instruction = o->block->start + padding(o->block, o->offset)
                                  + o->offset;
 
+          int32_t* p = reinterpret_cast<int32_t*>(dst + instruction);
+
+#if AVIAN_TARGET_ARCH == AVIAN_ARCH_ARM64
+          int32_t v = entry - instruction;
+          expect(&con, v == (v & PoolOffsetMask));
+
+          const int32_t mask = (PoolOffsetMask >> 2) << 5;
+          *p = (((v >> 2) << 5) & mask) | ((~mask) & *p);
+#else
           int32_t v = (entry - 8) - instruction;
           expect(&con, v == (v & PoolOffsetMask));
 
-          int32_t* p = reinterpret_cast<int32_t*>(dst + instruction);
           *p = (v & PoolOffsetMask) | ((~PoolOffsetMask) & *p);
+#endif
 
           poolSize += TargetBytesPerWord;
         }
diff --git a/src/codegen/target/arm/fixup.cpp b/src/codegen/target/arm/fixup.cpp
index e1d41b6eb4..7f5c0ff277 100644
--- a/src/codegen/target/arm/fixup.cpp
+++ b/src/codegen/target/arm/fixup.cpp
@@ -92,14 +92,27 @@ bool bounded(int right, int left, int32_t v)
 
 void* updateOffset(vm::System* s, uint8_t* instruction, int64_t value)
 {
-  // ARM's PC is two words ahead, and branches drop the bottom 2 bits.
-  int32_t v = (reinterpret_cast<uint8_t*>(value) - (instruction + 8)) >> 2;
-
-  int32_t mask;
-  expect(s, bounded(0, 8, v));
-  mask = 0xFFFFFF;
-
   int32_t* p = reinterpret_cast<int32_t*>(instruction);
+
+#if AVIAN_TARGET_ARCH == AVIAN_ARCH_ARM64
+  int32_t v;
+  int32_t mask;
+  if ((*p >> 24) == 0x54) {
+    // conditional branch
+    v = ((reinterpret_cast<uint8_t*>(value) - instruction) >> 2) << 5;
+    mask = 0xFFFFE0;
+  } else {
+    // unconditional branch
+    v = (reinterpret_cast<uint8_t*>(value) - instruction) >> 2;
+    mask = 0x3FFFFFF;
+  }
+#else
+  int32_t v = (reinterpret_cast<uint8_t*>(value) - (instruction + 8)) >> 2;
+  const int32_t mask = 0xFFFFFF;
+#endif
+
+  expect(s, bounded(0, 8, v));
+
   *p = (v & mask) | ((~mask) & *p);
 
   return instruction + 4;
@@ -214,6 +227,101 @@ void appendPoolEvent(Context* con,
   b->poolEventTail = e;
 }
 
+bool needJump(MyBlock* b)
+{
+  return b->next or b->size != (b->size & PoolOffsetMask);
+}
+
+unsigned padding(MyBlock* b, unsigned offset)
+{
+  unsigned total = 0;
+  for (PoolEvent* e = b->poolEventHead; e; e = e->next) {
+    if (e->offset <= offset) {
+      if (needJump(b)) {
+        total += vm::TargetBytesPerWord;
+      }
+      for (PoolOffset* o = e->poolOffsetHead; o; o = o->next) {
+        total += vm::TargetBytesPerWord;
+      }
+    } else {
+      break;
+    }
+  }
+  return total;
+}
+
+void resolve(MyBlock* b)
+{
+  Context* con = b->context;
+
+  if (b->poolOffsetHead) {
+    if (con->poolOffsetTail) {
+      con->poolOffsetTail->next = b->poolOffsetHead;
+    } else {
+      con->poolOffsetHead = b->poolOffsetHead;
+    }
+    con->poolOffsetTail = b->poolOffsetTail;
+  }
+
+  if (con->poolOffsetHead) {
+    bool append;
+    if (b->next == 0 or b->next->poolEventHead) {
+      append = true;
+    } else {
+      int32_t v
+          = (b->start + b->size + b->next->size + vm::TargetBytesPerWord - 8)
+            - (con->poolOffsetHead->offset + con->poolOffsetHead->block->start);
+
+      append = (v != (v & PoolOffsetMask));
+
+      if (DebugPool) {
+        fprintf(stderr,
+                "current %p %d %d next %p %d %d\n",
+                b,
+                b->start,
+                b->size,
+                b->next,
+                b->start + b->size,
+                b->next->size);
+        fprintf(stderr,
+                "offset %p %d is of distance %d to next block; append? %d\n",
+                con->poolOffsetHead,
+                con->poolOffsetHead->offset,
+                v,
+                append);
+      }
+    }
+
+    if (append) {
+#ifndef NDEBUG
+      int32_t v
+          = (b->start + b->size - 8)
+            - (con->poolOffsetHead->offset + con->poolOffsetHead->block->start);
+
+      expect(con, v == (v & PoolOffsetMask));
+#endif  // not NDEBUG
+
+      appendPoolEvent(
+          con, b, b->size, con->poolOffsetHead, con->poolOffsetTail);
+
+      if (DebugPool) {
+        for (PoolOffset* o = con->poolOffsetHead; o; o = o->next) {
+          fprintf(stderr,
+                  "include %p %d in pool event %p at offset %d in block %p\n",
+                  o,
+                  o->offset,
+                  b->poolEventTail,
+                  b->size,
+                  b);
+        }
+      }
+
+      con->poolOffsetHead = 0;
+      con->poolOffsetTail = 0;
+    }
+  }
+}
+
 }  // namespace arm
 }  // namespace codegen
 }  // namespace avian
diff --git a/src/codegen/target/arm/fixup.h b/src/codegen/target/arm/fixup.h
index 5460295d95..2e9c0aca01 100644
--- a/src/codegen/target/arm/fixup.h
+++ b/src/codegen/target/arm/fixup.h
@@ -27,7 +27,11 @@ namespace arm {
 
 const bool DebugPool = false;
 
+#if AVIAN_TARGET_ARCH == AVIAN_ARCH_ARM64
+const int32_t PoolOffsetMask = 0x1FFFFF;
+#else
 const int32_t PoolOffsetMask = 0xFFF;
+#endif
 
 class Task {
  public:
diff --git a/src/codegen/target/arm/operations32.cpp b/src/codegen/target/arm/operations32.cpp
index 5a9f5e8a0e..07dd7f0175 100644
--- a/src/codegen/target/arm/operations32.cpp
+++ b/src/codegen/target/arm/operations32.cpp
@@ -181,101 +181,6 @@ void unsignedShiftRightC(Context* con,
   }
 }
 
-bool needJump(MyBlock* b)
-{
-  return b->next or b->size != (b->size & PoolOffsetMask);
-}
-
-unsigned padding(MyBlock* b, unsigned offset)
-{
-  unsigned total = 0;
-  for (PoolEvent* e = b->poolEventHead; e; e = e->next) {
-    if (e->offset <= offset) {
-      if (needJump(b)) {
-        total += vm::TargetBytesPerWord;
-      }
-      for (PoolOffset* o = e->poolOffsetHead; o; o = o->next) {
-        total += vm::TargetBytesPerWord;
-      }
-    } else {
-      break;
-    }
-  }
-  return total;
-}
-
-void resolve(MyBlock* b)
-{
-  Context* con = b->context;
-
-  if (b->poolOffsetHead) {
-    if (con->poolOffsetTail) {
-      con->poolOffsetTail->next = b->poolOffsetHead;
-    } else {
-      con->poolOffsetHead = b->poolOffsetHead;
-    }
-    con->poolOffsetTail = b->poolOffsetTail;
-  }
-
-  if (con->poolOffsetHead) {
-    bool append;
-    if (b->next == 0 or b->next->poolEventHead) {
-      append = true;
-    } else {
-      int32_t v
-          = (b->start + b->size + b->next->size + vm::TargetBytesPerWord - 8)
-            - (con->poolOffsetHead->offset + con->poolOffsetHead->block->start);
-
-      append = (v != (v & PoolOffsetMask));
-
-      if (DebugPool) {
-        fprintf(stderr,
-                "current %p %d %d next %p %d %d\n",
-                b,
-                b->start,
-                b->size,
-                b->next,
-                b->start + b->size,
-                b->next->size);
-        fprintf(stderr,
-                "offset %p %d is of distance %d to next block; append? %d\n",
-                con->poolOffsetHead,
-                con->poolOffsetHead->offset,
-                v,
-                append);
-      }
-    }
-
-    if (append) {
-#ifndef NDEBUG
-      int32_t v
-          = (b->start + b->size - 8)
-            - (con->poolOffsetHead->offset + con->poolOffsetHead->block->start);
-
-      expect(con, v == (v & PoolOffsetMask));
-#endif  // not NDEBUG
-
-      appendPoolEvent(
-          con, b, b->size, con->poolOffsetHead, con->poolOffsetTail);
-
-      if (DebugPool) {
-        for (PoolOffset* o = con->poolOffsetHead; o; o = o->next) {
-          fprintf(stderr,
-                  "include %p %d in pool event %p at offset %d in block %p\n",
-                  o,
-                  o->offset,
-                  b->poolEventTail,
-                  b->size,
-                  b);
-        }
-      }
-
-      con->poolOffsetHead = 0;
-      con->poolOffsetTail = 0;
-    }
-  }
-}
-
 void jumpR(Context* con, unsigned size UNUSED, lir::RegisterPair* target)
 {
   assertT(con, size == vm::TargetBytesPerWord);
diff --git a/src/codegen/target/arm/operations64.cpp b/src/codegen/target/arm/operations64.cpp
index c3058102df..32a31cf2f5 100644
--- a/src/codegen/target/arm/operations64.cpp
+++ b/src/codegen/target/arm/operations64.cpp
@@ -125,9 +125,16 @@ uint32_t orr(Register Rd, Register Rn, Register Rm, unsigned size)
   return (size == 8 ? 0xaa0003e0 : 0x2a0003e0) | (Rm.index() << 16) | (Rn.index() << 5) | Rd.index();
 }
 
+uint32_t addi(Register Rd, Register Rn, int value, int shift, unsigned size)
+{
+  return (size == 8 ? 0x91000000 : 0x11000000) | (shift ? 0x400000 : 0)
+    | (value << 10) | (Rn.index() << 5) | Rd.index();
+}
+
 uint32_t mov(Register Rd, Register Rn, unsigned size)
 {
-  return orr(Rd, Register(31), Rn, size);
+  return Rn.index() == 31 ? addi(Rd, Rn, 0, 0, size)
+    : orr(Rd, Register(31), Rn, size);
 }
 
 uint32_t movz(Register Rd, int value, unsigned shift, unsigned size)
@@ -150,7 +157,8 @@ uint32_t movk(Register Rd, int value, unsigned shift, unsigned size)
 
 uint32_t ldrPCRel(Register Rd, int offset, unsigned size)
 {
-  return (size == 8 ? 0x58000000 : 0x18000000) | (offset << 5) | Rd.index();
+  return (size == 8 ? 0x58000000 : 0x18000000) | ((offset >> 2) << 5)
+         | Rd.index();
 }
 
 uint32_t add(Register Rd, Register Rn, Register Rm, unsigned size)
@@ -186,12 +194,6 @@ uint32_t mul(Register Rd, Register Rn, Register Rm, unsigned size)
   return madd(Rd, Rn, Rm, Register(31), size);
 }
 
-uint32_t addi(Register Rd, Register Rn, int value, int shift, unsigned size)
-{
-  return (size == 8 ? 0x91000000 : 0x11000000) | (shift ? 0x400000 : 0)
-    | (value << 10) | (Rn.index() << 5) | Rd.index();
-}
-
 uint32_t subi(Register Rd, Register Rn, int value, int shift, unsigned size)
 {
   return (size == 8 ? 0xd1000000 : 0x51000000) | (shift ? 0x400000 : 0)
@@ -307,8 +309,8 @@ uint32_t strhi(Register Rs, Register Rn, int offset)
 
 uint32_t stri(Register Rs, Register Rn, int offset, unsigned size)
 {
-  return (size == 8 ? 0xb9000000 : 0xf9000000) | (offset << 10)
-         | (Rn.index() << 5) | Rs.index();
+  return (size == 8 ? 0xf9000000 : 0xb9000000)
+    | ((offset >> (size == 8 ? 3 : 2)) << 10) | (Rn.index() << 5) | Rs.index();
 }
 
 uint32_t ldrFd(Register Fd, Register Rn, Register Rm, unsigned size)
@@ -381,8 +383,8 @@ uint32_t ldrswi(Register Rd, Register Rn, int offset)
 
 uint32_t ldri(Register Rd, Register Rn, int offset, unsigned size)
 {
-  return (size == 8 ? 0xb9400000 : 0xf9400000) | (offset << 10)
-         | (Rn.index() << 5) | Rd.index();
+  return (size == 8 ? 0xf9400000 : 0xb9400000)
+    | ((offset >> (size == 8 ? 3 : 2)) << 10) | (Rn.index() << 5) | Rd.index();
 }
 
 uint32_t fcmp(Register Fn, Register Fm, unsigned size)
@@ -400,7 +402,7 @@ uint32_t neg(Register Rd, Register Rm, unsigned size)
 uint32_t cmp(Register Rn, Register Rm, unsigned size)
 {
   return (size == 8 ? 0xeb00001f : 0x6b00001f) | (Rm.index() << 16)
-         | (Rn.index() << 5);
+         | (Rn.index() == 31 ? 0x2063ff : (Rn.index() << 5));
 }
 
 uint32_t cmpi(Register Rn, int value, unsigned shift, unsigned size)
@@ -426,42 +428,42 @@ uint32_t blr(Register Rn)
 
 uint32_t beq(int offset)
 {
-  return 0x54000000 | (offset >> 2);
+  return 0x54000000 | ((offset >> 2) << 5);
 }
 
 uint32_t bne(int offset)
 {
-  return 0x54000001 | (offset >> 2);
+  return 0x54000001 | ((offset >> 2) << 5);
 }
 
 uint32_t blt(int offset)
 {
-  return 0x5400000b | (offset >> 2);
+  return 0x5400000b | ((offset >> 2) << 5);
 }
 
 uint32_t bgt(int offset)
 {
-  return 0x5400000c | (offset >> 2);
+  return 0x5400000c | ((offset >> 2) << 5);
 }
 
 uint32_t ble(int offset)
 {
-  return 0x5400000d | (offset >> 2);
+  return 0x5400000d | ((offset >> 2) << 5);
 }
 
 uint32_t bge(int offset)
 {
-  return 0x5400000a | (offset >> 2);
+  return 0x5400000a | ((offset >> 2) << 5);
 }
 
 uint32_t bhi(int offset)
 {
-  return 0x54000008 | (offset >> 2);
+  return 0x54000008 | ((offset >> 2) << 5);
 }
 
 uint32_t bpl(int offset)
 {
-  return 0x54000005 | (offset >> 2);
+  return 0x54000005 | ((offset >> 2) << 5);
 }
 
 uint32_t brk(int flag)
@@ -966,7 +968,7 @@ void store(Context* c,
     if (release) {
       c->client->releaseTemporary(normalized);
     }
-  } else if (abs(offset) == (abs(offset) & 0xFF)) {
+  } else if (abs(offset) == (abs(offset) & 0xFFF)) {
     if (isFpr(src)) {
       switch (size) {
       case 4:
@@ -988,7 +990,12 @@ void store(Context* c,
         break;
 
       case 4:
+        assertT(c, offset == (offset & (~3)));
+        append(c, stri(src->low, base, offset, size));
+        break;
+
       case 8:
+        assertT(c, offset == (offset & (~7)));
         append(c, stri(src->low, base, offset, size));
         break;
 
@@ -1020,8 +1027,21 @@ void moveRM(Context* c,
 {
   assertT(c, srcSize == dstSize);
 
-  store(
-      c, srcSize, src, dst->base, dst->offset, dst->index, dst->scale, true);
+  if (src->low.index() == 31) {
+    assertT(c, c->client == 0); // the compiler should never ask us to
+                                // store the SP; we'll only get here
+                                // when assembling a thunk
+
+    lir::RegisterPair tmp(Register(9)); // we're in a thunk, so we can
+                                        // clobber this
+
+    moveRR(c, srcSize, src, srcSize, &tmp);
+    store(
+          c, srcSize, &tmp, dst->base, dst->offset, dst->index, dst->scale, true);
+  } else {
+    store(
+          c, srcSize, src, dst->base, dst->offset, dst->index, dst->scale, true);
+  }
 }
 
 void load(Context* c,
@@ -1085,7 +1105,7 @@ void load(Context* c,
     if (release) {
       c->client->releaseTemporary(normalized);
     }
-  } else if (abs(offset) == (abs(offset) & 0xFF)) {
+  } else if (abs(offset) == (abs(offset) & 0xFFF)) {
     if (isFpr(dst)) {
       switch (srcSize) {
       case 4:
@@ -1119,6 +1139,7 @@ void load(Context* c,
         if (signExtend and srcSize == 4 and dstSize == 8) {
           append(c, ldrswi(dst->low, base, offset));
         } else {
+          assertT(c, offset == (offset & (srcSize == 8 ? (~7) : (~3))));
           append(c, ldri(dst->low, base, offset, srcSize));
         }
         break;
@@ -1238,7 +1259,8 @@ void moveAR(Context* c,
             unsigned dstSize,
             lir::RegisterPair* dst)
 {
-  assertT(c, srcSize == TargetBytesPerWord and dstSize == TargetBytesPerWord);
+  assertT(c, srcSize == vm::TargetBytesPerWord
+          and dstSize == vm::TargetBytesPerWord);
 
   lir::Constant constant(src->address);
   moveCR(c, srcSize, &constant, dstSize, dst);
@@ -1312,6 +1334,20 @@ void compareRM(Context* c,
   c->client->releaseTemporary(tmp.low);
 }
 
+void compareMR(Context* c,
+               unsigned aSize,
+               lir::Memory* a,
+               unsigned bSize,
+               lir::RegisterPair* b)
+{
+  assertT(c, aSize == bSize);
+
+  lir::RegisterPair tmp(c->client->acquireTemporary(GPR_MASK));
+  moveMR(c, aSize, a, aSize, &tmp);
+  compareRR(c, aSize, &tmp, bSize, b);
+  c->client->releaseTemporary(tmp.low);
+}
+
 int32_t branch(Context* c, lir::TernaryOperation op)
 {
   switch (op) {
@@ -1397,8 +1433,17 @@ void branchRM(Context* c,
   assertT(c, not isFloatBranch(op));
   assertT(c, size <= vm::TargetBytesPerWord);
 
-  compareRM(c, size, a, size, b);
-  branch(c, op, target);
+  if (a->low.index() == 31) {
+    // stack overflow checks need to compare to the stack pointer, but
+    // we can only encode that in the opposite operand order we're
+    // given, so we need to reverse everything:
+    assertT(c, op == lir::JumpIfGreaterOrEqual);
+    compareMR(c, size, b, size, a);
+    branch(c, lir::JumpIfLess, target);
+  } else {
+    compareRM(c, size, a, size, b);
+    branch(c, op, target);
+  }
 }
 
 void branchCM(Context* c,
@@ -1537,21 +1582,6 @@ void storeLoadBarrier(Context* c)
   memoryBarrier(c);
 }
 
-bool needJump(MyBlock*)
-{
-  return false;
-}
-
-unsigned padding(MyBlock*, unsigned)
-{
-  return 0;
-}
-
-void resolve(MyBlock*)
-{
-  // ignore
-}
-
 }  // namespace arm
 }  // namespace codegen
 }  // namespace avian
diff --git a/src/compile-arm64.S b/src/compile-arm64.S
index 65f76df6f3..744e6cd71e 100644
--- a/src/compile-arm64.S
+++ b/src/compile-arm64.S
@@ -16,11 +16,11 @@
 #define BYTES_PER_WORD 4
 
 #define LOCAL(x) .L##x
-   
+
 #ifdef __APPLE__
 #  define GLOBAL(x) _##x
 #else
-#  define GLOBAL(x) x   
+#  define GLOBAL(x) x
 #endif
 
 #define CONTINUATION_NEXT 4
@@ -29,7 +29,7 @@
 #define CONTINUATION_FRAME_POINTER_OFFSET 24
 #define CONTINUATION_LENGTH 28
 #define CONTINUATION_BODY 32
-   
+
 .globl GLOBAL(vmInvoke)
 .align 2
 GLOBAL(vmInvoke):
@@ -89,8 +89,7 @@ GLOBAL(vmInvoke_returnAddress):
   // MyProcess::getStackTrace in compile.cpp for details on how we get
   // a reliable stack trace from a thread that might be interrupted at
   // any point in its execution.
-  mov  x5, #0
-  str  x5, [x19, #TARGET_THREAD_STACK]
+  str  xzr, [x19, #TARGET_THREAD_STACK]
 
 .globl GLOBAL(vmInvoke_safeStack)
 .align 2
@@ -100,11 +99,10 @@ GLOBAL(vmInvoke_safeStack):
 #error todo
 #endif // AVIAN_CONTINUATIONS
 
-  mov   x5, #0
-  str   x5, [x19, #TARGET_THREAD_STACK]
+  str   xzr, [x19, #TARGET_THREAD_STACK]
 
   // restore return type
-  ldr   w5, [sp], #4
+  ldr   w5, [sp,#16]!
 
   // restore callee-saved register values
   ldp   x19, x20, [sp,#16]
@@ -112,7 +110,7 @@ GLOBAL(vmInvoke_safeStack):
   ldp   x23, x24, [sp,#48]
   ldp   x25, x26, [sp,#64]
   ldp   x27, x28, [sp,#80]
-  ldp   x29, x30, [sp],#96
+  ldp   x29, x30, [sp,#96]!
 
 LOCAL(vmInvoke_return):
   br    x30