A u64 converts to and from a float on x86 as it does under LLVM, across the top half of its range

TODO.org

@@ -1092,9 +1092,12 @@ build included, reaches =emit_globals_init= as =Emit.startup_plan='s
 straight into their symbol are =Tast.const_init= ones, which neither transfer
 nor read anything. Nothing to change.
 
-** TODO A u64 converted to f64 is signed on x86
+** DONE A u64 converted to f64 is signed on x86
-=(f64 (u64 18446744073709551615))= prints =1.84467e+19= under LLVM and =-1= under
+CLOSED: [2026-09-25]
-=--x86=, with a literal or a run-time =u64= alike.
+=--x86= converts =u64= to and from =f64= and =f32= the way LLVM's =uitofp= and
+=fptoui= do: the halve-and-double sequence one way, subtract 2^63 and set the top
+bit the other, ties rounding to even. =test/programs/u64-float.flan= runs on both
+backends. A cast that misses =u64='s range reports it as =[0 18446744073709551615]=.
 
 ** DONE An aggregate built in place never reads its own destination
 CLOSED: [2026-09-25]

lib/x86.ml

@@ -1333,6 +1333,7 @@ let option_lay f (t : Types.t) =
 let cc_e = 4 and cc_ne = 5
 let cc_b = 2 and cc_ae = 3 and cc_be = 6 and cc_a = 7
 let cc_l = 12 and cc_ge = 13 and cc_le = 14 and cc_g = 15
+let cc_s = 8
 
 let int_cc ~signed (p : Tast.prim) =
   match p, signed with
@@ -3419,9 +3420,29 @@ and cast f (a : Tast.expr) (target : Types.t) dst =
       cvtss2sd f.b ~dst:xmm0 ~src:xmm0;
     fstore f.b ~src:xmm0 ~mm:(lmem f dst ~scratch:r11) ~f64:(f64_of dst_t)
   | false, true ->
+    let f64 = f64_of dst_t in
     load_loc f ~reg:rax l src_t;
-    cvtsi2f f.b ~f64:(f64_of dst_t) ~dst:xmm0 ~src:rax;
+    if src_t = Types.Int Types.U64 then begin
-    fstore f.b ~src:xmm0 ~mm:(lmem f dst ~scratch:r11) ~f64:(f64_of dst_t)
+      (* [cvtsi2sd] reads the register as signed, so a u64 with the top bit
+         set is halved first and doubled after. The bit shifted out is ORed
+         back into the lowest one, which keeps the halved value from landing
+         exactly on a tie it was not on, so the one rounding the conversion
+         does is the rounding of the whole value. *)
+      let big = new_label f "u2fbig" and done_ = new_label f "u2fdone" in
+      test_rr f.b ~a:rax ~c:rax;
+      jcc_lbl f.b ~cc:cc_s big;
+      cvtsi2f f.b ~f64 ~dst:xmm0 ~src:rax;
+      jmp_lbl f.b done_;
+      lbl f.b big;
+      mov_rr f.b ~dst:rcx ~src:rax;
+      shr_imm f.b ~dst:rcx ~n:1;
+      and_imm f.b ~dst:rax 1;
+      or_rr f.b ~dst:rcx ~src:rax;
+      cvtsi2f f.b ~f64 ~dst:xmm0 ~src:rcx;
+      farith f.b ~op:0x58 ~f64 ~dst:xmm0 ~src:xmm0;
+      lbl f.b done_
+    end else cvtsi2f f.b ~f64 ~dst:xmm0 ~src:rax;
+    fstore f.b ~src:xmm0 ~mm:(lmem f dst ~scratch:r11) ~f64
   | true, false ->
     fload f.b ~dst:xmm0 ~mm:(lmem f l ~scratch:r11) ~f64:(f64_of src_t);
     (* [cvttsd2si] answers a fixed "integer indefinite" for a value out of
@@ -3431,7 +3452,24 @@ and cast f (a : Tast.expr) (target : Types.t) dst =
     (match src_t, dst_t with
      | Types.Float sk, Types.Int k -> check_cast f a.Tast.loc sk k
      | _ -> ());
-    cvttf2si f.b ~f64:(f64_of src_t) ~dst:rax ~src:xmm0;
+    let f64 = f64_of src_t in
+    if dst_t = Types.Int Types.U64 then begin
+      (* [cvttsd2si] answers only for the signed range, so a value at 2^63 or
+         above has 2^63 taken off before the conversion and the top bit set
+         after it. A NaN compares unordered and takes the first branch. *)
+      let big = new_label f "f2ubig" and done_ = new_label f "f2udone" in
+      fload f.b ~dst:1 ~mm:(Sym (float_const f (ldexp 1.0 63) ~f64, 0)) ~f64;
+      ucomis f.b ~f64 ~a:xmm0 ~c:1;
+      jcc_lbl f.b ~cc:cc_ae big;
+      cvttf2si f.b ~f64 ~dst:rax ~src:xmm0;
+      jmp_lbl f.b done_;
+      lbl f.b big;
+      farith f.b ~op:0x5c ~f64 ~dst:xmm0 ~src:1;
+      cvttf2si f.b ~f64 ~dst:rax ~src:xmm0;
+      imm_into f ~reg:rcx Int64.min_int;
+      xor_rr f.b ~dst:rax ~src:rcx;
+      lbl f.b done_
+    end else cvttf2si f.b ~f64 ~dst:rax ~src:xmm0;
     store_loc f ~reg:rax dst dst_t
 
 (* ── Call frame information ──────────────────────────────────────────── *)

runtime/flan_rt.c

@@ -1020,11 +1020,19 @@ static void flan_arith_fail(const uint8_t *loc, int64_t loclen, int32_t op,
             "to\n",
             (int)loclen, (const char *)loc);
     break;
+  /* An unsigned type's range starts at zero and a signed one's below it, so
+   * the lower bound says how to read the upper one: u64's is all ones. */
   default:
-    fprintf(stderr,
+    if (lhs == 0)
-            "%.*s: this value does not fit the integer type it is cast to, "
+      fprintf(stderr,
-            "which holds [%lld %lld]\n",
+              "%.*s: this value does not fit the integer type it is cast to, "
-            (int)loclen, (const char *)loc, (long long)lhs, (long long)rhs);
+              "which holds [0 %llu]\n",
+              (int)loclen, (const char *)loc, (unsigned long long)rhs);
+    else
+      fprintf(stderr,
+              "%.*s: this value does not fit the integer type it is cast to, "
+              "which holds [%lld %lld]\n",
+              (int)loclen, (const char *)loc, (long long)lhs, (long long)rhs);
     break;
   }
   rt_die();

test/programs/arith.flan

@@ -82,6 +82,9 @@
       (= n 10) (print (i32 (/ (f64 1.0) (f64 0.0))))
       (= n 11) (print (u8 (/ (f32 -1.0) (f32 0.0))))
       (= n 12) (print (i32 (/ (f32 0.0) (f32 0.0))))
+      ;; A u64 destination, whose upper bound is all ones and is printed as
+      ;; the unsigned number it is.
+      (= n 13) (print (u64 (* huge 1e-280)))
 
       :else (println "?"))
     0))

test/programs/u64-float.flan

@@ -0,0 +1,65 @@
+;;;; Conversions between the unsigned integers and the floats, across the top
+;;;; half of u64's range, where the value does not fit a signed register. Both
+;;;; backends print the same lines. Each case is written twice: once through a
+;;;; global, so the conversion happens at run time, and once on a literal.
+;;;;
+;;;; A float is printed by converting it back to an integer, because the
+;;;; printed float has six digits and every case here differs past the sixth.
+
+(defonce top   u64 18446744073709551615)   ; 2^64 - 1
+(defonce half1 u64 9223372036854775809)    ; 2^63 + 1
+(defonce tie0  u64 9223372036854776832)    ; 2^63 + 1024, a tie that rounds down to even
+(defonce tie1  u64 9223372036854778880)    ; 2^63 + 3072, a tie that rounds up to even
+(defonce above u64 9223372036854776833)    ; 2^63 + 1025, just past a tie
+(defonce small u64 12345)
+(defonce ftie0 u64 9223372586610589696)    ; 2^63 + 2^39, an f32 tie that rounds down
+(defonce fup   u64 9223372586610589697)    ; 2^63 + 2^39 + 1
+(defonce ftie1 u64 9223373686122217472)    ; 2^63 + 3*2^39, an f32 tie that rounds up
+(defonce u32top u32 4294967295)
+
+(defonce d63   f64 9223372036854775808.0)  ; 2^63
+(defonce dmax  f64 18446744073709549568.0) ; the largest f64 below 2^64
+(defonce dbelow f64 9223372036854774784.0) ; the largest f64 below 2^63
+(defonce d15   f64 1.5e19)
+(defonce dsmall f64 7.9)
+(defonce s63   f32 9223372036854775808.0)
+(defonce smax  f32 18446742974197923840.0) ; the largest f32 below 2^64
+(defonce d32   f64 4294967295.0)
+(defonce d3e9  f64 3e9)
+
+(defn main [] i32
+  ;; u64 to f64.
+  (println (f64 top))
+  (println (u64 (* (f64 top) 0.5)))
+  (println (u64 (f64 half1)))
+  (println (u64 (f64 tie0)))
+  (println (u64 (f64 tie1)))
+  (println (u64 (f64 above)))
+  (println (u64 (f64 small)))
+  (println (f64 (u64 18446744073709551615)))
+  (println (u64 (f64 (u64 9223372036854778880))))
+  ;; u64 to f32.
+  (println (u64 (* (f32 top) (f32 0.5))))
+  (println (u64 (f32 ftie0)))
+  (println (u64 (f32 fup)))
+  (println (u64 (f32 ftie1)))
+  (println (u64 (f32 small)))
+  (println (u64 (f32 (u64 9223372586610589697))))
+  ;; u32 to both.
+  (println (u64 (f64 u32top)))
+  (println (u64 (f32 u32top)))
+  ;; f64 and f32 to u64.
+  (println (u64 d63))
+  (println (u64 dmax))
+  (println (u64 dbelow))
+  (println (u64 d15))
+  (println (u64 dsmall))
+  (println (u64 s63))
+  (println (u64 smax))
+  (println (u64 18446744073709549568.0))
+  (println (u64 (f32 1.5e19)))
+  ;; f64 to u32.
+  (println (u32 d32))
+  (println (u32 d3e9))
+  (println (u32 4294967295.0))
+  0)

test/test_acceptance.ml

@@ -2639,6 +2639,8 @@ let () =
         "this value is infinite, which has no integer value to cast to";
       traps "an f32 NaN cast to an integer" "12"
         "this value is NaN, which has no integer value to cast to";
+      traps "a float too large for a u64" "13"
+        "which holds [0 18446744073709551615]";
       (try Sys.remove exe with Sys_error _ -> ())
     in
     arith ();
@@ -2772,6 +2774,29 @@ let () =
     outputs ~x86:true
       "an aggregate that reads its destination sees the old value, --x86"
       "programs/self-read.flan" self_read_out;
+    (* Conversions between u64 and the floats across the top half of u64's
+       range, where [cvtsi2sd] and [cvttsd2si] read the register as signed:
+       x86 answered -1 for (f64 (u64 18446744073709551615)). The expected
+       lines are what LLVM's [uitofp] and [fptoui] give, ties included. *)
+    let u64_float_out =
+      "1.84467e+19\n9223372036854775808\n9223372036854775808\n\
+       9223372036854775808\n9223372036854779904\n9223372036854777856\n\
+       12345\n1.84467e+19\n9223372036854779904\n9223372036854775808\n\
+       9223372036854775808\n9223373136366403584\n9223374235878031360\n\
+       12345\n9223373136366403584\n4294967295\n4294967296\n\
+       9223372036854775808\n18446744073709549568\n9223372036854774784\n\
+       15000000000000000000\n7\n9223372036854775808\n\
+       18446742974197923840\n18446744073709549568\n15000000520515485696\n\
+       4294967295\n3000000000\n4294967295\n"
+    in
+    outputs "u64 and the floats convert across the whole range"
+      "programs/u64-float.flan" u64_float_out;
+    outputs ~opt:"-O0"
+      "u64 and the floats convert across the whole range, -O0"
+      "programs/u64-float.flan" u64_float_out;
+    outputs ~x86:true
+      "u64 and the floats convert across the whole range, --x86"
+      "programs/u64-float.flan" u64_float_out;
 
     (* ── Packages: the link follows the program ────────────────────────
        A package's C and linker arguments used to come with the import,