[pypy-commit] pypy libgccjit-backend: Get test_ovf_operations and test_ovf_operations_reversed to pass

[dmalcolm]

Author: David Malcolm <dmalcolm at redhat.com>
Branch: libgccjit-backend
Changeset: r75092:dfab06318973
Date: 2014-12-23 15:41 -0500
http://bitbucket.org/pypy/pypy/changeset/dfab06318973/

Log:	Get test_ovf_operations and test_ovf_operations_reversed to pass

diff --git a/rpython/jit/backend/libgccjit/assembler.py b/rpython/jit/backend/libgccjit/assembler.py
--- a/rpython/jit/backend/libgccjit/assembler.py
+++ b/rpython/jit/backend/libgccjit/assembler.py
@@ -14,6 +14,11 @@
 from rpython.rtyper.lltypesystem.rffi import *
 from rpython.rtyper.lltypesystem import lltype, rffi, rstr, llmemory
 
+# Global state
+num_anon_loops = 0
+num_guard_failure_fns = 0
+num_bridges = 0
+
 class Params:
     def __init__(self, assembler):
         self.paramlist = []
@@ -77,8 +82,9 @@
     """
     def __init__(self, assembler):
         self.failure_params = Params(assembler)
-        self.serial = assembler.num_guard_failure_fns
-        assembler.num_guard_failure_fns += 1
+        global num_guard_failure_fns
+        self.serial = num_guard_failure_fns
+        num_guard_failure_fns += 1
         self.t_fn_ptr_type = assembler.ctxt.new_function_ptr_type (
             assembler.t_jit_frame_ptr,
             self.failure_params.paramtypes,
@@ -155,10 +161,6 @@
         self.propagate_exception_path = 0
         #self.teardown()
 
-        self.num_anon_loops = 0
-        self.num_guard_failure_fns = 0
-        self.num_bridges = 0
-
         self.sizeof_signed = rffi.sizeof(lltype.Signed)
         self.label_for_descr = {}
         self.block_for_label_descr = {}
@@ -251,8 +253,9 @@
         self.lvalue_for_box = {}
 
         if not loopname:
-            loopname = 'anonloop_%i' % self.num_anon_loops
-            self.num_anon_loops += 1
+            global num_anon_loops
+            loopname = 'anonloop_%i' % num_anon_loops
+            num_anon_loops += 1
         print("  loopname: %r" % loopname)
         self.make_function(loopname, inputargs, operations)
 
@@ -295,8 +298,9 @@
                                                            operations)
         self.lvalue_for_box = {}
 
-        name = "bridge_%i" % self.num_bridges
-        self.num_bridges += 1
+        global num_bridges
+        name = "bridge_%i" % num_bridges
+        num_bridges += 1
 
         self.make_function(name, inputargs, operations)
 
@@ -399,6 +403,7 @@
                                    self.loop_params.paramlist,
                                    r_int(0)))
 
+        self.lvalue_ovf = self.fn.new_local(self.t_bool, "ovf")
         self.b_current = self.fn.new_block("initial")
 
         # Add an initial comment summarizing the loop or bridge:
@@ -725,6 +730,14 @@
                              self.t_bool,
                              boolval_is_nonnull, boolval_vtable_equality))
 
+    def emit_guard_no_overflow(self, resop):
+        self._impl_guard(resop, r_int(0),
+                         self.lvalue_ovf.as_rvalue())
+
+    def emit_guard_overflow(self, resop):
+        self._impl_guard(resop, r_int(1),
+                         self.lvalue_ovf.as_rvalue())
+
     def _impl_write_output_args(self, params, args):
         # Write outputs back:
         for idx, arg in enumerate(args):
@@ -1031,3 +1044,29 @@
             # ... = ARG1,
             self.ctxt.new_cast(self.expr_to_rvalue(resop._arg1),
                                t_field))
+
+    # "INT_*_OVF" operations:
+    def _impl_int_ovf(self, resop, builtin_name):
+        """
+        We implement ovf int binops using GCC builtins,
+        generating a call of the form:
+          ovf = __builtin_FOO_overflow(arg0, arg1, &result);
+        This is then optimized by gcc's builtins.c
+        fold_builtin_arith_overflow.
+        """
+        rval0 = self.expr_to_rvalue(resop._arg0)
+        rval1 = self.expr_to_rvalue(resop._arg1)
+        lvalres = self.expr_to_lvalue(resop.result)
+        builtin_fn = self.ctxt.get_builtin_function(builtin_name)
+        # "ovf = __builtin_FOO_overflow(arg0, arg1, &result);"
+        call = self.ctxt.new_call(builtin_fn,
+                                  [rval0,
+                                   rval1,
+                                   lvalres.get_address()])
+        self.b_current.add_assignment(self.lvalue_ovf, call)
+    def emit_int_add_ovf(self, resop):
+        self._impl_int_ovf(resop, '__builtin_add_overflow')
+    def emit_int_sub_ovf(self, resop):
+        self._impl_int_ovf(resop, '__builtin_sub_overflow')
+    def emit_int_mul_ovf(self, resop):
+        self._impl_int_ovf(resop, '__builtin_mul_overflow')
diff --git a/rpython/jit/backend/libgccjit/rffi_bindings.py b/rpython/jit/backend/libgccjit/rffi_bindings.py
--- a/rpython/jit/backend/libgccjit/rffi_bindings.py
+++ b/rpython/jit/backend/libgccjit/rffi_bindings.py
@@ -230,6 +230,10 @@
                                               self.PARAM_P_P,
                                               INT]),
 
+            (self.GCC_JIT_FUNCTION_P,
+             'gcc_jit_context_get_builtin_function', [self.GCC_JIT_CONTEXT_P,
+                                                      CCHARP]),
+
             (self.GCC_JIT_OBJECT_P,
              'gcc_jit_function_as_object', [self.GCC_JIT_FUNCTION_P]),
 
@@ -751,6 +755,13 @@
 
         return Function(self.lib, self, fn)
 
+    def get_builtin_function(self, name):
+        name_charp = str2charp(name)
+        fn = self.lib.gcc_jit_context_get_builtin_function(self.inner_ctxt,
+                                                           name_charp)
+        free_charp(name_charp)
+        return Function(self.lib, self, fn)
+
     def new_global(self, kind, type_, name):
         name_charp = str2charp(name)
         lvalue = self.lib.gcc_jit_context_new_global(self.inner_ctxt,