compiler: Cache expensive target data layout queries

On one typical experiment, this shaves about 3 seconds (~25%)
off the overall kernel compilation time.

GitHub: Closes #1370.
This commit is contained in:
David Nadlinger 2019-10-27 20:42:48 +00:00
parent 5d7f22ffa4
commit 611bcc4db4
1 changed files with 34 additions and 6 deletions

View File

@ -124,6 +124,35 @@ class DebugInfoEmitter:
"scope": scope "scope": scope
}) })
class ABILayoutInfo:
"""Caches DataLayout size/alignment lookup results.
llvmlite's Type.get_abi_{size, alignment}() are implemented in a very
inefficient way, in particular _get_ll_pointer_type() used to construct the
corresponding llvm::Type is. We thus cache the results, optionally directly
using the compiler type as a key.
(This is a separate class for use with @memoize.)
"""
def __init__(self, lldatalayout, llcontext, llty_of_type):
self.cache = {}
self.lldatalayout = lldatalayout
self.llcontext = llcontext
self.llty_of_type = llty_of_type
@memoize
def get_size_align(self, llty):
lowered = llty._get_ll_pointer_type(self.lldatalayout, self.llcontext)
return (self.lldatalayout.get_pointee_abi_size(lowered),
self.lldatalayout.get_pointee_abi_alignment(lowered))
@memoize
def get_size_align_for_type(self, typ):
return self.get_size_align(self.llty_of_type(typ))
class LLVMIRGenerator: class LLVMIRGenerator:
def __init__(self, engine, module_name, target, embedding_map): def __init__(self, engine, module_name, target, embedding_map):
self.engine = engine self.engine = engine
@ -134,6 +163,8 @@ class LLVMIRGenerator:
self.llmodule.triple = target.triple self.llmodule.triple = target.triple
self.llmodule.data_layout = target.data_layout self.llmodule.data_layout = target.data_layout
self.lldatalayout = llvm.create_target_data(self.llmodule.data_layout) self.lldatalayout = llvm.create_target_data(self.llmodule.data_layout)
self.abi_layout_info = ABILayoutInfo(self.lldatalayout, self.llcontext,
self.llty_of_type)
self.function_flags = None self.function_flags = None
self.llfunction = None self.llfunction = None
self.llmap = {} self.llmap = {}
@ -548,10 +579,7 @@ class LLVMIRGenerator:
llrpcattrs = [] llrpcattrs = []
for attr in typ.attributes: for attr in typ.attributes:
attrtyp = typ.attributes[attr] attrtyp = typ.attributes[attr]
size = self.llty_of_type(attrtyp). \ size, alignment = self.abi_layout_info.get_size_align_for_type(attrtyp)
get_abi_size(self.lldatalayout, context=self.llcontext)
alignment = self.llty_of_type(attrtyp). \
get_abi_alignment(self.lldatalayout, context=self.llcontext)
if offset % alignment != 0: if offset % alignment != 0:
offset += alignment - (offset % alignment) offset += alignment - (offset % alignment)
@ -735,7 +763,7 @@ class LLVMIRGenerator:
def mark_dereferenceable(self, load): def mark_dereferenceable(self, load):
assert isinstance(load, ll.LoadInstr) and isinstance(load.type, ll.PointerType) assert isinstance(load, ll.LoadInstr) and isinstance(load.type, ll.PointerType)
pointee_size = load.type.pointee.get_abi_size(self.lldatalayout, context=self.llcontext) pointee_size, _ = self.abi_layout_info.get_size_align(load.type.pointee)
metadata = self.llmodule.add_metadata([ll.Constant(lli64, pointee_size)]) metadata = self.llmodule.add_metadata([ll.Constant(lli64, pointee_size)])
load.set_metadata('dereferenceable', metadata) load.set_metadata('dereferenceable', metadata)