perf: optimise from_word128_ptr by speculation (#3893)

...that the upper 64 bits of the cycle count are all zeros. Since 1T cycles is about one dollar, having 1 in the upper word will have a value more than 18M dollars. I guess that sending such amounts is rare.
However doing the bignum arithmetic to construct, shift and add the upper word is costly, see https://dfinity.github.io/canister-profiling/heartbeat/Motoko_heartbeat.svg

Of course the slow paths should be tested well :-) — Done by `test/run-drun/basic-cycles.mo`!

---------------
TODO
- [ ] We could explore an ultra-fast path too, when the whole amount is 0.

src/codegen/compile.ml

@@ -4439,17 +4439,26 @@ module Cycles = struct
 
   let from_word128_ptr env = Func.share_code1 env "from_word128_ptr" ("ptr", I32Type) [I32Type]
     (fun env get_ptr ->
-      get_ptr ^^
-      (G.i (Load {ty = I64Type; align = 0; offset = 0l; sz = None })) ^^
-      BigNum.from_word64 env ^^
-      get_ptr ^^
-      compile_add_const 8l ^^
-      (G.i (Load {ty = I64Type; align = 0; offset = 0l; sz = None })) ^^
-      BigNum.from_word64 env ^^
-      (* shift left 64 bits *)
-      compile_unboxed_const 64l ^^
-      BigNum.compile_lsh env ^^
-      BigNum.compile_add env)
+     let set_lower, get_lower = new_local env "lower" in
+     get_ptr ^^
+     G.i (Load {ty = I64Type; align = 0; offset = 0l; sz = None }) ^^
+     BigNum.from_word64 env ^^
+     set_lower ^^
+     get_ptr ^^
+     G.i (Load {ty = I64Type; align = 0; offset = 8l; sz = None }) ^^
+     G.i (Test (Wasm.Values.I64 I64Op.Eqz)) ^^
+     G.if1 I32Type
+       get_lower
+       begin
+         get_lower ^^
+         get_ptr ^^
+         G.i (Load {ty = I64Type; align = 0; offset = 8l; sz = None }) ^^
+         BigNum.from_word64 env ^^
+         (* shift left 64 bits *)
+         compile_unboxed_const 64l ^^
+         BigNum.compile_lsh env ^^
+         BigNum.compile_add env
+       end)
 
   (* takes a bignum from the stack, traps if ≥2^128, and leaves two 64bit words on the stack *)
   (* only used twice, so ok to not use share_code1; that would require I64Type support in FakeMultiVal *)