Add support for native_compute on x86-32.

Author: silene

kernel/byterun/rocq_values.c

@@ -107,38 +107,75 @@ value rocq_tcode_array(value tcodes) {
   CAMLreturn(res);
 }
 
+/* The following architecture-dependent function returns a core pointer
+   (wrapped in an option) that can be used as field 0 of an OCaml closure.
+   But since the word before the start of the code contains a header that
+   causes it to be ignored by the garbage collector, the pointer can even be
+   put inside closures that do not have tag 247.
+
+   The code itself translates the call "foo x" (i.e., "%apply x foo") into
+   "accumulate foo.2 x". The first few instructions are those generated by
+   the OCaml compiler for the closure returned by the following partial
+   application:
+
+   let foo data = accumulate data (* in Nativevalues *)
+
+   The last few instructions invoke the code pointer stored in rocq_accumulate_addr.
+   This variable is initialized using field 2 of the closure passed as parameter
+   (i.e., "accumulate"), which is the pointer to the arity-2 direct call. */
+
 #if defined(__GNUC__) && defined(__amd64__)
 static code_t rocq_accumulate_addr asm("rocq_accumulate_addr") __attribute__((used));
 
 value rocq_proxy_accu(value clos) {
   value v;
   CAMLassert(Tag_val(clos) == Closure_tag && Arity_closinfo(Closinfo_val(clos)) == 2);
-  /* Field 2 of the closure contains the code pointer for the arity-2 direct call. */
   rocq_accumulate_addr = ((code_t *)clos)[2];
-  /* The following assembly block does not perform any meaningful computation;
-     it just returns a pointer to the inner code (notice the initial "jmp").
-     The inner code translates the call "foo x" (i.e., "%apply x foo") into
-     "accumulate foo.2 x". For both calls, the two arguments are stored in %rax
-     and %rbx, while register %rdi is caller-saved and hence usable. */
   asm("jmp 1f\n\t"
       ".align 8\n\t"
       ".quad 3067\n"
       "2:\n\t"
       "mov %%rax, %%rdi\n\t"
       "mov 16(%%rbx), %%rax\n\t"
       "mov %%rdi, %%rbx\n\t"
-      "mov rocq_accumulate_addr(%%rip), %%rdi\n\t"
-      "jmp *%%rdi\n"
+      "lea rocq_accumulate_addr(%%rip), %%rdi\n\t"
+      "jmp *(%%rdi)\n"
       "1:\n\t"
       "lea 2b(%%rip), %0\n\t"
       : "=r"(v));
-  /* v is a pointer that can be used as field 0 of an OCaml closure. But it is
-     also a pointer to a block that is ignored by the garbage collector (notice
-     the header 3067). So, v can be put inside closures that do not have tag 247. */
   value r = caml_alloc_small(1, 0);
   Field(r, 0) = v;
   return r;
 }
+
+#elif defined(__GNUC__) && defined(__i386__)
+static code_t rocq_accumulate_addr asm("rocq_accumulate_addr") __attribute__((used));
+
+value rocq_proxy_accu(value clos) {
+  value v;
+  CAMLassert(Tag_val(clos) == Closure_tag && Arity_closinfo(Closinfo_val(clos)) == 2);
+  rocq_accumulate_addr = ((code_t *)clos)[2];
+  asm("jmp 1f\n\t"
+      ".align 4\n\t"
+      ".long 3067\n"
+      "2:\n\t"
+      "mov %%eax, %%ecx\n\t"
+      "mov 8(%%ebx), %%eax\n\t"
+      "mov %%ecx, %%ebx\n\t"
+      "call __x86.get_pc_thunk.cx\n\t"
+      "add $_GLOBAL_OFFSET_TABLE_, %%ecx\n\t"
+      "mov rocq_accumulate_addr@GOT(%%ecx), %%ecx\n\t"
+      "jmp *(%%ecx)\n"
+      "1:\n\t"
+      "call __x86.get_pc_thunk.ax\n\t"
+      "add $_GLOBAL_OFFSET_TABLE_, %%eax\n\t"
+      "lea 2b@GOTOFF(%%eax), %0\n\t"
+      : "=r"(v) : : "eax");
+  value r = caml_alloc_small(1, 0);
+  Field(r, 0) = v;
+  return r;
+}
+
 #else
 value rocq_proxy_accu(value) {
   return 1;