Make patricia trees big-endian

backend/arm64/emit.mlp

@@ -247,6 +247,7 @@ let name_for_int_operation = function
   | Ilsl -> "lsl"
   | Ilsr -> "lsr"
   | Iasr -> "asr"
+  | Iclz { arg_is_non_zero = _ } -> "clz"
   | _ -> assert false
 
 (* Decompose an integer constant into four 16-bit shifted fragments.
@@ -1000,6 +1001,9 @@ let emit_instr i =
         `	smulh	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}, {emit_reg i.arg.(1)}\n`
     | Lop(Intop (Imulh { signed = false })) ->
         `	umulh	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}, {emit_reg i.arg.(1)}\n`
+    | Lop(Intop (Iclz _ as op)) ->
+        let instr = name_for_int_operation op in
+        `	{emit_string instr}	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}\n`
     | Lop(Intop op) ->
         let instr = name_for_int_operation op in
         `	{emit_string instr}	{emit_reg i.res.(0)}, {emit_reg i.arg.(0)}, {emit_reg i.arg.(1)}\n`

backend/arm64/proc.ml

@@ -484,7 +484,7 @@ let assemble_file infile outfile =
 let init () = ()
 
 let operation_supported = function
-  | Cclz _ | Cctz _ | Cpopcnt
+  | Cctz _ | Cpopcnt
   | Cprefetch _ | Catomic _
   (* CR mslater: (float32) arm64 *)
   | Cnegf Float32 | Cabsf Float32 | Caddf Float32
@@ -497,7 +497,7 @@ let operation_supported = function
                   Int_of_float Float32 | Float_of_int Float32 |
                   V128_of_scalar _ | Scalar_of_v128 _)
     -> false   (* Not implemented *)
-  | Cbswap _
+  | Cclz _ | Cbswap _
   | Capply _ | Cextcall _ | Cload _ | Calloc _ | Cstore _
   | Caddi | Csubi | Cmuli | Cmulhi _ | Cdivi | Cmodi
   | Cand | Cor | Cxor | Clsl | Clsr | Casr

middle_end/flambda2/algorithms/builtin_stubs.c

@@ -0,0 +1,47 @@
+#include "caml/mlvalues.h"
+#include "caml/fail.h"
+
+#if defined(_MSC_VER)
+#include <intrin.h>
+#endif
+
+// These are replaced with clz instructions by flambda2.
+//
+// Use weak symbols in order to allow compiler-libs and ocaml_intrinsics_kernel
+// to be shared dependencies of the same program.
+
+CAMLweakdef intnat caml_int_clz_tagged_to_untagged(value i) {
+    // Do not use Long_val(v1) conversion, instead preserving the tag.
+    // It guarantees that the input to builtin_clz / _BitScanReverse is
+    // non-zero, to guard against versions that are undefined for input 0. The
+    // tag does not change the number of leading zeros.
+#if defined(__GNUC__) || defined(__clang__)
+    #if SIZEOF_PTR == SIZEOF_INT
+    return __builtin_clz(i);
+    #elif SIZEOF_PTR == SIZEOF_LONG
+    return __builtin_clzl(i);
+    #elif SIZEOF_PTR == SIZEOF_LONGLONG
+    return __builtin_clzll(i);
+    #else
+    #error "No builtin clz function available"
+    #endif
+#elif defined(_MSC_VER)
+    unsigned long r = 0;
+    #ifdef SIZEOF_PTR == 8
+    _BitScanReverse64(&r, i);
+    r ^= 63;
+    #elif SIZEOF_PTR == 4
+    _BitScanReverse(&r, i);
+    r ^= 31;
+    #else
+    #error "No builtin bsr function available"
+    #endif
+    return r;
+#else
+    #error "Unsupported compiler"
+#endif
+}
+
+CAMLweakdef value caml_int_clz_tagged_to_tagged(value i) {
+    return Val_int(caml_int_clz_tagged_to_untagged(i));
+}

middle_end/flambda2/algorithms/dune

@@ -4,6 +4,19 @@
  (name flambda2_algorithms)
  (wrapped true)
  (instrumentation (backend bisect_ppx))
+ (foreign_stubs
+  (language c)
+  (names builtin_stubs)
+  (flags
+   ((:include %{project_root}/oc_cflags.sexp)
+    (:include %{project_root}/sharedlib_cflags.sexp)
+    (:include %{project_root}/oc_cppflags.sexp))))
  (ocamlopt_flags
   (:standard -O3 -open Int_replace_polymorphic_compare))
  (libraries ocamlcommon))
+
+(install
+ (files
+  (dllflambda2_algorithms_stubs.so as stublibs/dllflambda2_algorithms_stubs.so))
+ (section lib)
+ (package ocaml))