compiler: mark loaded pointers as !dereferenceable.

Also, lower the bound for test_pulse_rate_dds, since we generate better code for it now.
m-labs · Mar 31, 2016 · 10108e6 · 10108e6
1 parent 2c04979
commit 10108e6
Showing 2 changed files with 16 additions and 9 deletions.
diff --git a/artiq/compiler/transforms/llvm_ir_generator.py b/artiq/compiler/transforms/llvm_ir_generator.py
@@ -175,6 +175,7 @@ def __init__(self, engine, module_name, target, function_map, object_map, type_m
         self.llmodule = ll.Module(context=self.llcontext, name=module_name)
         self.llmodule.triple = target.triple
         self.llmodule.data_layout = target.data_layout
+        self.lldatalayout = llvm.create_target_data(self.llmodule.data_layout)
         self.function_flags = None
         self.llfunction = None
         self.llmap = {}
@@ -459,8 +460,6 @@ def emit_attribute_writeback(self):
             if llobject is not None:
                 llobjects[typ].append(llobject.bitcast(llptr))
 
-        lldatalayout = llvm.create_target_data(self.llmodule.data_layout)
-
         llrpcattrty = self.llcontext.get_identified_type("A")
         llrpcattrty.elements = [lli32, llptr, llptr]
 
@@ -512,9 +511,9 @@ def rpc_tag_error(typ):
             for attr in typ.attributes:
                 attrtyp   = typ.attributes[attr]
                 size      = self.llty_of_type(attrtyp). \
-                    get_abi_size(lldatalayout, context=self.llcontext)
+                    get_abi_size(self.lldatalayout, context=self.llcontext)
                 alignment = self.llty_of_type(attrtyp). \
-                    get_abi_alignment(lldatalayout, context=self.llcontext)
+                    get_abi_alignment(self.lldatalayout, context=self.llcontext)
 
                 if offset % alignment != 0:
                     offset += alignment - (offset % alignment)
@@ -688,13 +687,19 @@ def llptr_to_var(self, llenv, env_ty, var_name, var_type=None):
             llouterenv.metadata['nonnull'] = self.empty_metadata
             return self.llptr_to_var(llouterenv, env_ty.params["$outer"], var_name)
 
+    def mark_dereferenceable(self, load):
+        assert isinstance(load, ll.LoadInstr) and isinstance(load.type, ll.PointerType)
+        pointee_size = load.type.pointee.get_abi_size(self.lldatalayout, context=self.llcontext)
+        metadata = self.llmodule.add_metadata([ll.Constant(lli64, pointee_size)])
+        load.metadata['dereferenceable'] = metadata
+
     def process_GetLocal(self, insn):
         env = insn.environment()
         llptr = self.llptr_to_var(self.map(env), env.type, insn.var_name)
         llptr.name = "ptr.{}.{}".format(env.name, insn.var_name)
         llvalue = self.llbuilder.load(llptr, name="val.{}.{}".format(env.name, insn.var_name))
         if isinstance(llvalue.type, ll.PointerType):
-            llvalue.metadata['nonnull'] = self.empty_metadata
+            self.mark_dereferenceable(llvalue)
         return llvalue
 
     def process_SetLocal(self, insn):
@@ -818,7 +823,7 @@ def process_GetAttr(self, insn):
                 if types.is_instance(typ) and attr in typ.constant_attributes:
                     llvalue.metadata['invariant.load'] = self.empty_metadata
                 if isinstance(llvalue.type, ll.PointerType):
-                    llvalue.metadata['nonnull'] = self.empty_metadata
+                    self.mark_dereferenceable(llvalue)
                 return llvalue
 
     def process_SetAttr(self, insn):
@@ -849,7 +854,7 @@ def process_GetElem(self, insn):
                                    inbounds=True)
         llvalue = self.llbuilder.load(llelt)
         if isinstance(llvalue.type, ll.PointerType):
-            llvalue.metadata['nonnull'] = self.empty_metadata
+            self.mark_dereferenceable(llvalue)
         return llvalue
 
     def process_SetElem(self, insn):

diff --git a/artiq/test/coredevice/test_rtio.py b/artiq/test/coredevice/test_rtio.py
@@ -105,6 +105,8 @@ def run(self):
 
 
 class PulseRateDDS(EnvExperiment):
+    kernel_constant_attributes = {"core", "core_dds", "dds0", "dds1"}
+
     def build(self):
         self.setattr_device("core")
         self.setattr_device("core_dds")
@@ -113,7 +115,7 @@ def build(self):
 
     @kernel
     def run(self):
-        dt = seconds_to_mu(150*us)
+        dt = seconds_to_mu(50*us)
         while True:
             delay(10*ms)
             for i in range(1250):
@@ -262,7 +264,7 @@ def test_pulse_rate_dds(self):
         rate = self.dataset_mgr.get("pulse_rate")
         print(rate)
         self.assertGreater(rate, 100*us)
-        self.assertLess(rate, 400*us)
+        self.assertLess(rate, 150*us)
 
     def test_loopback_count(self):
         npulses = 2