Add ABI tagging to crates, adjust rustc output and runtime stack-frame setup so access to argv works.

12 files changed, 97 insertions, 42 deletions

diff --git a/src/boot/be/abi.ml b/src/boot/be/abi.ml
index 035d1f05..89e308bf 100644
--- a/src/boot/be/abi.ml
+++ b/src/boot/be/abi.ml
@@ -110,23 +110,33 @@ let indirect_args_elt_closure = 0;;
 (* Current worst case is by vec grow glue *)
 let worst_case_glue_call_args = 8;;
 
+(* 
+ * ABI tags used to inform the runtime which sort of frame to set up for new
+ * spawned functions. FIXME: There is almost certainly a better abstraction to
+ * use.
+ *)
+let abi_x86_rustboot_cdecl = 1;;
+let abi_x86_rustc_fastcall = 2;;
+
 type abi =
-  {
-    abi_word_sz: int64;
-    abi_word_bits: Il.bits;
-    abi_word_ty: Common.ty_mach;
+    {
+      abi_word_sz: int64;
+      abi_word_bits: Il.bits;
+      abi_word_ty: Common.ty_mach;
+
+      abi_tag: int;
 
-    abi_has_pcrel_data: bool;
-    abi_has_pcrel_code: bool;
+      abi_has_pcrel_data: bool;
+      abi_has_pcrel_code: bool;
 
-    abi_n_hardregs: int;
-    abi_str_of_hardreg: (int -> string);
+      abi_n_hardregs: int;
+      abi_str_of_hardreg: (int -> string);
 
-    abi_emit_target_specific: (Il.emitter -> Il.quad -> unit);
-    abi_constrain_vregs: (Il.quad -> (Il.vreg,Bits.t) Hashtbl.t -> unit);
+      abi_emit_target_specific: (Il.emitter -> Il.quad -> unit);
+      abi_constrain_vregs: (Il.quad -> (Il.vreg,Bits.t) Hashtbl.t -> unit);
 
-    abi_emit_fn_prologue: (Il.emitter
-                           -> Common.size        (* framesz *)
+      abi_emit_fn_prologue: (Il.emitter
+                             -> Common.size        (* framesz *)
                              -> Common.size      (* callsz  *)
                                -> Common.nabi
                                  -> Common.fixup (* grow_task *)
diff --git a/src/boot/be/x86.ml b/src/boot/be/x86.ml
index 30b49ed1..49b660be 100644
--- a/src/boot/be/x86.ml
+++ b/src/boot/be/x86.ml
@@ -1851,6 +1851,8 @@ let (abi:Abi.abi) =
     Abi.abi_word_bits = word_bits;
     Abi.abi_word_ty = word_ty;
 
+    Abi.abi_tag = Abi.abi_x86_rustboot_cdecl;
+
     Abi.abi_has_pcrel_data = false;
     Abi.abi_has_pcrel_code = true;
 
diff --git a/src/boot/me/trans.ml b/src/boot/me/trans.ml
index 4f717219..bbf49e83 100644
--- a/src/boot/me/trans.ml
+++ b/src/boot/me/trans.ml
@@ -2727,6 +2727,7 @@ let trans_visitor
                   [|
                     Il.Cell new_task;
                     exit_task_glue_fptr;
+                     (imm (Int64.of_int abi.Abi.abi_tag));
                     fptr_operand;
                     callsz
                   |];
@@ -2739,6 +2740,7 @@ let trans_visitor
                    [|
                      Il.Cell new_task;
                      exit_task_glue_fptr;
+                     (imm (Int64.of_int abi.Abi.abi_tag));
                      fptr_operand;
                      callsz
                    |];
@@ -6183,6 +6185,8 @@ let trans_visitor
             tab_sz cx.ctxt_required_rust_sym_num;
             tab_sz cx.ctxt_required_c_sym_num;
             tab_sz cx.ctxt_required_lib_num;
+
+            Asm.WORD (word_ty_mach, Asm.IMM (Int64.of_int abi.Abi.abi_tag));
           |]))
     in
 
diff --git a/src/comp/back/abi.rs b/src/comp/back/abi.rs
index 82a85a6c..f41f6e20 100644
--- a/src/comp/back/abi.rs
+++ b/src/comp/back/abi.rs
@@ -60,6 +60,9 @@ const int worst_case_glue_call_args = 7;
 
 const int n_upcall_glues = 7;
 
+const int abi_x86_rustboot_cdecl = 1;
+const int abi_x86_rustc_fastcall = 2;
+
 fn memcpy_glue_name() -> str {
     ret "rust_memcpy_glue";
 }
diff --git a/src/comp/back/x86.rs b/src/comp/back/x86.rs
index 907cb32e..10227df7 100644
--- a/src/comp/back/x86.rs
+++ b/src/comp/back/x86.rs
@@ -98,12 +98,12 @@ fn rust_activate_glue() -> vec[str] {
          *      | esi    |
          *      | ebx    | <-- current task->rust_sp == current esp
          *
-         * 
+         *
          *      This is a problem. If we return to "esp <- task->rust_sp" it
          *      will push esp back down by 5 words. This manifests as a rust
          *      stack that grows by 5 words on each yield/reactivate. Not
          *      good.
-         * 
+         *
          *      So what we do here is just adjust task->rust_sp up 5 words as
          *      well, to mirror the movement in esp we're about to
          *      perform. That way the "esp <- task->rust_sp" we 'ret' to below
diff --git a/src/comp/middle/trans.rs b/src/comp/middle/trans.rs
index 9da7e009..4a672f49 100644
--- a/src/comp/middle/trans.rs
+++ b/src/comp/middle/trans.rs
@@ -345,7 +345,8 @@ fn T_crate(type_names tn) -> TypeRef {
                           T_int(),      // size_t main_exit_task_glue_off
                           T_int(),      // int n_rust_syms
                           T_int(),      // int n_c_syms
-                          T_int()       // int n_libs
+                          T_int(),      // int n_libs
+                          T_int()       // uintptr_t abi_tag
                           ));
     tn.associate(s, t);
     ret t;
@@ -4545,7 +4546,7 @@ fn trans_exit_task_glue(@crate_ctxt cx) {
     let vec[ValueRef] V_args = vec();
 
     auto llfn = cx.glues.exit_task_glue;
-    let ValueRef lltaskptr = llvm.LLVMGetParam(llfn, 0u);
+    let ValueRef lltaskptr = llvm.LLVMGetParam(llfn, 3u);
     auto fcx = @rec(llfn=llfn,
                     lltaskptr=lltaskptr,
                     llenv=C_null(T_opaque_closure_ptr(cx.tn)),
@@ -4596,7 +4597,8 @@ fn create_crate_constant(@crate_ctxt cx) {
                      exit_task_glue_off,  // size_t main_exit_task_glue_off
                      C_null(T_int()),     // int n_rust_syms
                      C_null(T_int()),     // int n_c_syms
-                     C_null(T_int())      // int n_libs
+                     C_null(T_int()),     // int n_libs
+                     C_int(abi.abi_x86_rustc_fastcall) // uintptr_t abi_tag
                      ));
 
     llvm.LLVMSetInitializer(cx.crate_ptr, crate_val);
@@ -4810,7 +4812,10 @@ fn make_glues(ModuleRef llmod, type_names tn) -> @glue_fns {
               * this is the signature required to retrieve it.
               */
              exit_task_glue = decl_cdecl_fn(llmod, abi.exit_task_glue_name(),
-                                            T_fn(vec(T_taskptr(tn)),
+                                            T_fn(vec(T_int(),
+                                                     T_int(),
+                                                     T_int(),
+                                                     T_taskptr(tn)),
                                                  T_void())),
 
              upcall_glues =
diff --git a/src/rt/rust.cpp b/src/rt/rust.cpp
index 0ea167a4..46fcb22e 100644
--- a/src/rt/rust.cpp
+++ b/src/rt/rust.cpp
@@ -78,7 +78,7 @@ command_line_args : public dom_owned<command_line_args>
 
 extern "C" CDECL int
 rust_start(uintptr_t main_fn, rust_crate const *crate, int argc,
-    char **argv) {
+           char **argv) {
 
     rust_srv *srv = new rust_srv();
     rust_kernel *kernel = new rust_kernel(srv);
@@ -87,7 +87,8 @@ rust_start(uintptr_t main_fn, rust_crate const *crate, int argc,
     rust_dom *dom = handle->referent();
     command_line_args *args = new (dom) command_line_args(dom, argc, argv);
 
-    dom->log(rust_log::DOM, "startup: %d args", args->argc);
+    dom->log(rust_log::DOM, "startup: %d args in 0x%" PRIxPTR,
+             args->argc, (uintptr_t)args->args);
     for (int i = 0; i < args->argc; i++) {
         dom->log(rust_log::DOM,
             "startup: arg[%d] = '%s'", i, args->argv[i]);
@@ -99,7 +100,8 @@ rust_start(uintptr_t main_fn, rust_crate const *crate, int argc,
 
     uintptr_t main_args[4] = {0, 0, 0, (uintptr_t)args->args};
     dom->root_task->start(crate->get_exit_task_glue(),
-        main_fn, (uintptr_t)&main_args, sizeof(main_args));
+                          crate->abi_tag, main_fn,
+                          (uintptr_t)&main_args, sizeof(main_args));
     int ret = dom->start_main_loop();
     delete args;
     kernel->destroy_domain(dom);
diff --git a/src/rt/rust_internal.h b/src/rt/rust_internal.h
index 61716703..42b61801 100644
--- a/src/rt/rust_internal.h
+++ b/src/rt/rust_internal.h
@@ -88,6 +88,10 @@ static size_t const TIME_SLICE_IN_MS = 10;
 
 static intptr_t const CONST_REFCOUNT = 0x7badface;
 
+// ABI tags for rust_start, rust_task::start and friends.
+static uintptr_t const ABI_X86_RUSTBOOT_CDECL = 1;
+static uintptr_t const ABI_X86_RUSTC_FASTCALL = 2;
+
 // This accounts for logging buffers.
 
 static size_t const BUF_BYTES = 2048;
@@ -241,6 +245,8 @@ public:
     size_t n_c_syms;
     size_t n_libs;
 
+    uintptr_t abi_tag;
+
     // Crates are immutable, constructed by the compiler.
 
     uintptr_t get_image_base() const;
diff --git a/src/rt/rust_task.cpp b/src/rt/rust_task.cpp
index 3564b9e3..1afbfdd6 100644
--- a/src/rt/rust_task.cpp
+++ b/src/rt/rust_task.cpp
@@ -123,6 +123,7 @@ rust_task::~rust_task()
 
 void
 rust_task::start(uintptr_t exit_task_glue,
+                 uintptr_t spawnee_abi,
                  uintptr_t spawnee_fn,
                  uintptr_t args,
                  size_t callsz)
@@ -147,26 +148,29 @@ rust_task::start(uintptr_t exit_task_glue,
     // The exit_task_glue frame we synthesize above the frame we activate:
     *spp-- = (uintptr_t) 0;          // closure-or-obj
     *spp-- = (uintptr_t) this;       // task
-    *spp-- = (uintptr_t) 0;          // output
-    *spp-- = (uintptr_t) 0;          // retpc
+    *spp-- = (uintptr_t) 0x0;        // output
+    *spp-- = (uintptr_t) 0x0;        // retpc
 
     uintptr_t exit_task_frame_base;
 
-    for (size_t j = 0; j < n_callee_saves; ++j) {
+    if (spawnee_abi == ABI_X86_RUSTBOOT_CDECL) {
+        for (size_t j = 0; j < n_callee_saves; ++j) {
 
-        // We want 'frame_base' to point to the old fp in this (exit-task)
-        // frame, because we're going to inject this frame-pointer into the
-        // callee-save frame pointer value in the *next* (spawnee) frame. A
-        // cheap trick, but this means the spawnee frame will restore the
-        // proper frame pointer of the glue frame as it runs its epilogue.
-        if (j == callee_save_fp)
-            exit_task_frame_base = (uintptr_t)spp;
+            // We want 'frame_base' to point to the old fp in this (exit-task)
+            // frame, because we're going to inject this frame-pointer into
+            // the callee-save frame pointer value in the *next* (spawnee)
+            // frame. A cheap trick, but this means the spawnee frame will
+            // restore the proper frame pointer of the glue frame as it runs
+            // its epilogue.
+            if (j == callee_save_fp)
+                exit_task_frame_base = (uintptr_t)spp;
 
-        *spp-- = 0;
-    }
+            *spp-- = 0;
+        }
 
-    *spp-- = (uintptr_t) dom->root_crate;  // crate ptr
-    *spp-- = (uintptr_t) 0;                // frame_glue_fns
+        *spp-- = (uintptr_t) dom->root_crate;  // crate ptr
+        *spp-- = (uintptr_t) 0;                // frame_glue_fns
+    }
 
     // Copy args from spawner to spawnee.
     if (args)  {
@@ -174,12 +178,16 @@ rust_task::start(uintptr_t exit_task_glue,
         src += 1;                  // spawn-call output slot
         src += 1;                  // spawn-call task slot
         src += 1;                  // spawn-call closure-or-obj slot
+
+        // Undo previous sp-- so we're pointing at the last word pushed.
+        ++spp;
+
         // Memcpy all but the task, output and env pointers
         callsz -= (3 * sizeof(uintptr_t));
         spp = (uintptr_t*) (((uintptr_t)spp) - callsz);
         memcpy(spp, src, callsz);
 
-        // Move sp down to point to task cell.
+        // Move sp down to point to last implicit-arg cell (env).
         spp--;
     } else {
         // We're at root, starting up.
@@ -188,10 +196,18 @@ rust_task::start(uintptr_t exit_task_glue,
 
     // The *implicit* incoming args to the spawnee frame we're
     // activating:
+    *spp-- = (uintptr_t) 0x0;               // closure-or-obj
+
+    if (spawnee_abi == ABI_X86_RUSTBOOT_CDECL) {
+        // in CDECL mode we write the task + outptr to the spawnee stack.
+        *spp-- = (uintptr_t) this;            // task
+        *spp-- = (uintptr_t) 0;               // output addr
+    } else {
+        // in FASTCALL mode we don't, the outptr will be in ecx and the task
+        // in edx, and the activate_glue will make sure to set that up.
+        I(dom, spawnee_abi == ABI_X86_RUSTC_FASTCALL);
+    }
 
-    *spp-- = (uintptr_t) 0;               // closure-or-obj
-    *spp-- = (uintptr_t) this;            // task
-    *spp-- = (uintptr_t) 0;               // output addr
     *spp-- = (uintptr_t) exit_task_glue;  // retpc
 
     // The context the activate_glue needs to switch stack.
diff --git a/src/rt/rust_task.h b/src/rt/rust_task.h
index 9fbc67ac..5318ab71 100644
--- a/src/rt/rust_task.h
+++ b/src/rt/rust_task.h
@@ -56,6 +56,7 @@ rust_task : public maybe_proxy<rust_task>,
     ~rust_task();
 
     void start(uintptr_t exit_task_glue,
+               uintptr_t spawnee_abi,
                uintptr_t spawnee_fn,
                uintptr_t args,
                size_t callsz);
diff --git a/src/rt/rust_upcall.cpp b/src/rt/rust_upcall.cpp
index 46e53048..1dba1102 100644
--- a/src/rt/rust_upcall.cpp
+++ b/src/rt/rust_upcall.cpp
@@ -559,6 +559,7 @@ extern "C" CDECL rust_task *
 upcall_start_task(rust_task *spawner,
                   rust_task *task,
                   uintptr_t exit_task_glue,
+                  uintptr_t spawnee_abi,
                   uintptr_t spawnee_fn,
                   size_t callsz) {
     LOG_UPCALL_ENTRY(spawner);
@@ -570,7 +571,8 @@ upcall_start_task(rust_task *spawner,
              ", spawnee 0x%" PRIxPTR
              ", callsz %" PRIdPTR ")", task->name, task, exit_task_glue,
              spawnee_fn, callsz);
-    task->start(exit_task_glue, spawnee_fn, spawner->rust_sp, callsz);
+    task->start(exit_task_glue, spawnee_abi, spawnee_fn,
+                spawner->rust_sp, callsz);
     return task;
 }
 
@@ -623,6 +625,7 @@ extern "C" CDECL maybe_proxy<rust_task> *
 upcall_start_thread(rust_task *task,
                     rust_proxy<rust_task> *child_task_proxy,
                     uintptr_t exit_task_glue,
+                    uintptr_t spawnee_abi,
                     uintptr_t spawnee_fn,
                     size_t callsz) {
     LOG_UPCALL_ENTRY(task);
@@ -630,9 +633,11 @@ upcall_start_thread(rust_task *task,
     rust_handle<rust_task> *child_task_handle = child_task_proxy->handle();
     task->log(rust_log::UPCALL | rust_log::MEM | rust_log::TASK,
               "exit_task_glue: " PTR ", spawnee_fn " PTR
-              ", callsz %" PRIdPTR ")", exit_task_glue, spawnee_fn, callsz);
+              ", callsz %" PRIdPTR ")",
+              exit_task_glue, spawnee_fn, callsz);
     rust_task *child_task = child_task_handle->referent();
-    child_task->start(exit_task_glue, spawnee_fn, task->rust_sp, callsz);
+    child_task->start(exit_task_glue, spawnee_abi, spawnee_fn,
+                      task->rust_sp, callsz);
 #if defined(__WIN32__)
     HANDLE thread;
     thread = CreateThread(NULL, 0, rust_thread_start, child_task->dom, 0,
diff --git a/src/rt/test/rust_test_runtime.cpp b/src/rt/test/rust_test_runtime.cpp
index 1cde532e..e0e24156 100644
--- a/src/rt/test/rust_test_runtime.cpp
+++ b/src/rt/test/rust_test_runtime.cpp
@@ -54,6 +54,7 @@ rust_task_test::worker::run() {
         kernel->create_domain(crate, "test");
     rust_dom *domain = handle->referent();
     domain->root_task->start(crate->get_exit_task_glue(),
+                             ABI_X86_RUSTBOOT_CDECL,
                              (uintptr_t)&task_entry, (uintptr_t)NULL, 0);
     domain->start_main_loop();
     kernel->destroy_domain(domain);