fixup tramp
[nit.git] / src / compiler / separate_compiler.nit
index 0ad35b9..9010227 100644 (file)
@@ -29,8 +29,15 @@ redef class ToolContext
        var opt_no_union_attribute = new OptionBool("Put primitive attibutes in a box instead of an union", "--no-union-attribute")
        # --no-shortcut-equate
        var opt_no_shortcut_equate = new OptionBool("Always call == in a polymorphic way", "--no-shortcut-equal")
+
        # --colors-are-symbols
-       var opt_colors_are_symbols = new OptionBool("Store colors as symbols (faster)", "--colors-are-symbols")
+       var opt_colors_are_symbols = new OptionBool("Store colors as symbols (link-boost)", "--colors-are-symbols")
+       # --trampoline-call
+       var opt_trampoline_call = new OptionBool("Use an indirection when calling", "--trampoline-call")
+       # --substitute-monomorph
+       var opt_substitute_monomorph = new OptionBool("Replace monomorph trampoline with direct call (link-boost)", "--substitute-monomorph")
+       # --link-boost
+       var opt_link_boost = new OptionBool("Enable all link-boost optimizations", "--link-boost")
 
        # --inline-coloring-numbers
        var opt_inline_coloring_numbers = new OptionBool("Inline colors and ids (semi-global)", "--inline-coloring-numbers")
@@ -53,7 +60,8 @@ redef class ToolContext
                self.option_context.add_option(self.opt_separate)
                self.option_context.add_option(self.opt_no_inline_intern)
                self.option_context.add_option(self.opt_no_union_attribute)
-               self.option_context.add_option(self.opt_no_shortcut_equate, opt_colors_are_symbols)
+               self.option_context.add_option(self.opt_no_shortcut_equate)
+               self.option_context.add_option(opt_colors_are_symbols, opt_trampoline_call, opt_substitute_monomorph, opt_link_boost)
                self.option_context.add_option(self.opt_inline_coloring_numbers, opt_inline_some_methods, opt_direct_call_monomorph, opt_skip_dead_methods, opt_semi_global)
                self.option_context.add_option(self.opt_colo_dead_methods)
                self.option_context.add_option(self.opt_tables_metrics)
@@ -70,6 +78,13 @@ redef class ToolContext
                        tc.opt_direct_call_monomorph.value = true
                        tc.opt_skip_dead_methods.value = true
                end
+               if tc.opt_link_boost.value then
+                       tc.opt_colors_are_symbols.value = true
+                       tc.opt_substitute_monomorph.value = true
+               end
+               if tc.opt_substitute_monomorph.value then
+                       tc.opt_trampoline_call.value = true
+               end
        end
 
        var separate_compiler_phase = new SeparateCompilerPhase(self, null)
@@ -560,7 +575,30 @@ class SeparateCompiler
                                var r = pd.separate_runtime_function
                                r.compile_to_c(self)
                                var r2 = pd.virtual_runtime_function
-                               r2.compile_to_c(self)
+                               if r2 != r then r2.compile_to_c(self)
+
+                               # Generate trampolines
+                               if modelbuilder.toolcontext.opt_trampoline_call.value then
+                                       r2.compile_trampolines(self)
+
+                                       # Replace monomorphic call to a trampoline by a direct call to the virtual implementation
+                                       if modelbuilder.toolcontext.opt_substitute_monomorph.value then do
+                                               var m = pd.mproperty
+                                               if rta == null then
+                                                       # Without RTA, monomorphic means alone (uniq name)
+                                                       if m.mpropdefs.length != 1 then break label
+                                               else
+                                                       # With RTA, monomorphic means only live methoddef
+                                                       if not rta.live_methoddefs.has(pd) then break label
+                                                       for md in m.mpropdefs do
+                                                               if md != pd and rta.live_methoddefs.has(md) then break label
+                                                       end
+                                               end
+                                               # Here the trick, GNU ld can substitute symbols with specific values.
+                                               var n2 = "CALL_" + m.const_color
+                                               linker_script.add("{n2} = {r2.c_name};")
+                                       end label
+                               end
                        end
                end
                self.mainmodule = old_module
@@ -1104,7 +1142,7 @@ class SeparateCompilerVisitor
                        return res
                end
 
-               return table_send(mmethod, arguments, mmethod.const_color)
+               return table_send(mmethod, arguments, mmethod)
        end
 
        # Handle common special cases before doing the effective method invocation
@@ -1170,7 +1208,7 @@ class SeparateCompilerVisitor
                return res
        end
 
-       private fun table_send(mmethod: MMethod, arguments: Array[RuntimeVariable], const_color: String): nullable RuntimeVariable
+       private fun table_send(mmethod: MMethod, arguments: Array[RuntimeVariable], mentity: MEntity): nullable RuntimeVariable
        do
                compiler.modelbuilder.nb_invok_by_tables += 1
                if compiler.modelbuilder.toolcontext.opt_invocation_metrics.value then add("count_invoke_by_tables++;")
@@ -1180,8 +1218,10 @@ class SeparateCompilerVisitor
 
                var res0 = before_send(mmethod, arguments)
 
+               var runtime_function = mmethod.intro.virtual_runtime_function
+               var msignature = runtime_function.called_signature
+
                var res: nullable RuntimeVariable
-               var msignature = mmethod.intro.msignature.resolve_for(mmethod.intro.mclassdef.bound_mtype, mmethod.intro.mclassdef.bound_mtype, mmethod.intro.mclassdef.mmodule, true)
                var ret = msignature.return_mtype
                if ret == null then
                        res = null
@@ -1189,10 +1229,8 @@ class SeparateCompilerVisitor
                        res = self.new_var(ret)
                end
 
-               var s = new FlatBuffer
                var ss = new FlatBuffer
 
-               s.append("val*")
                ss.append("{recv}")
                for i in [0..msignature.arity[ do
                        var a = arguments[i+1]
@@ -1200,16 +1238,20 @@ class SeparateCompilerVisitor
                        if i == msignature.vararg_rank then
                                t = arguments[i+1].mcasttype
                        end
-                       s.append(", {t.ctype}")
                        a = self.autobox(a, t)
                        ss.append(", {a}")
                end
 
-
-               var r
-               if ret == null then r = "void" else r = ret.ctype
-               self.require_declaration(const_color)
-               var call = "(({r} (*)({s}))({arguments.first}->class->vft[{const_color}]))({ss}) /* {mmethod} on {arguments.first.inspect}*/"
+               var const_color = mentity.const_color
+               var call
+               if not compiler.modelbuilder.toolcontext.opt_trampoline_call.value then
+                       self.require_declaration(const_color)
+                       call = "(({runtime_function.c_funptrtype})({arguments.first}->class->vft[{const_color}]))({ss}) /* {mmethod} on {arguments.first.inspect}*/"
+               else
+                       var callsym = "CALL_" + const_color
+                       self.require_declaration(callsym)
+                       call = "{callsym}({ss}) /* {mmethod} on {arguments.first.inspect}*/"
+               end
 
                if res != null then
                        self.add("{res} = {call};")
@@ -1285,7 +1327,7 @@ class SeparateCompilerVisitor
                        self.compiler.mainmodule = main
                        return res
                end
-               return table_send(m.mproperty, arguments, m.const_color)
+               return table_send(m.mproperty, arguments, m)
        end
 
        redef fun vararg_instance(mpropdef, recv, varargs, elttype)
@@ -1802,7 +1844,20 @@ redef class MMethodDef
                        # Because the function is virtual, the signature must match the one of the original class
                        var intromclassdef = mproperty.intro.mclassdef
                        var recv = intromclassdef.bound_mtype
+
+                       res = separate_runtime_function
+                       if res.called_recv == recv then
+                               self.virtual_runtime_function_cache = res
+                               return res
+                       end
+
                        var msignature = mproperty.intro.msignature.resolve_for(recv, recv, intromclassdef.mmodule, true)
+
+                       if recv.ctype == res.called_recv.ctype and msignature.c_equiv(res.called_signature) then
+                               self.virtual_runtime_function_cache = res
+                               return res
+                       end
+
                        res = new SeparateRuntimeFunction(self, recv, msignature, "VIRTUAL_{c_name}")
                        self.virtual_runtime_function_cache = res
                        res.is_thunk = true
@@ -1812,6 +1867,23 @@ redef class MMethodDef
        private var virtual_runtime_function_cache: nullable SeparateRuntimeFunction
 end
 
+redef class MSignature
+       # Does the C-version of `self` the same than the C-version of `other`?
+       fun c_equiv(other: MSignature): Bool
+       do
+               if self == other then return true
+               if arity != other.arity then return false
+               for i in [0..arity[ do
+                       if mparameters[i].mtype.ctype != other.mparameters[i].mtype.ctype then return false
+               end
+               if return_mtype != other.return_mtype then
+                       if return_mtype == null or other.return_mtype == null then return false
+                       if return_mtype.ctype != other.return_mtype.ctype then return false
+               end
+               return true
+       end
+end
+
 # The C function associated to a methoddef separately compiled
 class SeparateRuntimeFunction
        super AbstractRuntimeFunction
@@ -1830,6 +1902,37 @@ class SeparateRuntimeFunction
 
        redef fun to_s do return self.mmethoddef.to_s
 
+       # The C return type (something or `void`)
+       var c_ret: String is lazy do
+               var ret = called_signature.return_mtype
+               if ret != null then
+                       return ret.ctype
+               else
+                       return "void"
+               end
+       end
+
+       # The C signature (only the parmeter part)
+       var c_sig: String is lazy do
+               var sig = new FlatBuffer
+               sig.append("({called_recv.ctype} self")
+               for i in [0..called_signature.arity[ do
+                       var mtype = called_signature.mparameters[i].mtype
+                       if i == called_signature.vararg_rank then
+                               mtype = mmethoddef.mclassdef.mmodule.get_primitive_class("Array").get_mtype([mtype])
+                       end
+                       sig.append(", {mtype.ctype} p{i}")
+               end
+               sig.append(")")
+               return sig.to_s
+       end
+
+       # The C type for the function pointer.
+       var c_funptrtype: String is lazy do return "{c_ret}(*){c_sig}"
+
+       # The arguments, as generated by `compile_to_c`
+       private var arguments: Array[RuntimeVariable] is noinit
+
        redef fun compile_to_c(compiler)
        do
                var mmethoddef = self.mmethoddef
@@ -1842,17 +1945,14 @@ class SeparateRuntimeFunction
                v.frame = frame
 
                var msignature = called_signature
+               var ret = called_signature.return_mtype
 
                var sig = new FlatBuffer
                var comment = new FlatBuffer
-               var ret = msignature.return_mtype
-               if ret != null then
-                       sig.append("{ret.ctype} ")
-               else
-                       sig.append("void ")
-               end
+               sig.append(c_ret)
+               sig.append(" ")
                sig.append(self.c_name)
-               sig.append("({selfvar.mtype.ctype} {selfvar}")
+               sig.append(c_sig)
                comment.append("({selfvar}: {selfvar.mtype}")
                arguments.add(selfvar)
                for i in [0..msignature.arity[ do
@@ -1861,16 +1961,15 @@ class SeparateRuntimeFunction
                                mtype = v.get_class("Array").get_mtype([mtype])
                        end
                        comment.append(", {mtype}")
-                       sig.append(", {mtype.ctype} p{i}")
                        var argvar = new RuntimeVariable("p{i}", mtype, mtype)
                        arguments.add(argvar)
                end
-               sig.append(")")
                comment.append(")")
                if ret != null then
                        comment.append(": {ret}")
                end
                compiler.provide_declaration(self.c_name, "{sig};")
+               self.arguments = arguments.to_a
 
                v.add_decl("/* method {self} for {comment} */")
                v.add_decl("{sig} \{")
@@ -1896,6 +1995,50 @@ class SeparateRuntimeFunction
                v.add("\}")
                compiler.names[self.c_name] = "{mmethoddef.full_name} ({mmethoddef.location.file.filename}:{mmethoddef.location.line_start})"
        end
+
+       # Compile the trampolines used to implement late-binding.
+       #
+       # See `opt_trampoline_call`.
+       fun compile_trampolines(compiler: SeparateCompiler)
+       do
+               var recv = self.mmethoddef.mclassdef.bound_mtype
+               var selfvar = arguments.first
+               var ret = called_signature.return_mtype
+
+               if mmethoddef.is_intro and recv.ctype == "val*" then
+                       var m = mmethoddef.mproperty
+                       var n2 = "CALL_" + m.const_color
+                       compiler.provide_declaration(n2, "{c_ret} {n2}{c_sig};")
+                       var v2 = compiler.new_visitor
+                       v2.add "{c_ret} {n2}{c_sig} \{"
+                       v2.require_declaration(m.const_color)
+                       var call = "(({c_funptrtype})({selfvar}->class->vft[{m.const_color}]))({arguments.join(", ")});"
+                       if ret != null then
+                               v2.add "return {call}"
+                       else
+                               v2.add call
+                       end
+
+                       v2.add "\}"
+
+               end
+               if mmethoddef.has_supercall and recv.ctype == "val*" then
+                       var m = mmethoddef
+                       var n2 = "CALL_" + m.const_color
+                       compiler.provide_declaration(n2, "{c_ret} {n2}{c_sig};")
+                       var v2 = compiler.new_visitor
+                       v2.add "{c_ret} {n2}{c_sig} \{"
+                       v2.require_declaration(m.const_color)
+                       var call = "(({c_funptrtype})({selfvar}->class->vft[{m.const_color}]))({arguments.join(", ")});"
+                       if ret != null then
+                               v2.add "return {call}"
+                       else
+                               v2.add call
+                       end
+
+                       v2.add "\}"
+               end
+       end
 end
 
 redef class MEntity