Builtin functions have been handled until now as special cases in
gimple-range-fold.cc. This set of patches makes the changes required to
create a range_operator for those functions. This allows them to behave
like a normal unary/binary operation through out the ranger ecosystem.
In particular, it will enable us to make GORI aware of them as we can
now provide op1_range and op2_range routines, as well as registering an
relations as needed. None of these enhanced functions are provided yet,
this is strictly a conversion. This enables us to do this for any
operation with 1 or 2 operands.
There are 17 patches, some are bug fixes, some are infrastructure, a
couple are just missing functionality, but most are them are conversions
of the builtins. I did each builtin as a separate patch so if a
regression triggers, we can pinpoint it faster.
Of note:
Patch 2 : Modifies the range_op_handler class to store an integer
handler and a float handler rather than the old tree-code and type. By
looking up the handler immediately and storing the pointer, this opens
up the possibility of processing handlers which are not in a tree-code
table.
Patch 3 : Range-ops is suppose to be IL independent, designed to work in
RTL land as well. A little bit of gimple had crept in, and I needed a
layer that is gimple aware. This patch introduces a
gimple_range_op_handler which inherits from range_op_handler, and acts
as the connector between the gimple IL and range-ops. Some of that code
was in range-ops, and a lot more was located in the GORI file. All
those bits and pieces have been moved into the new class.
Patch 7 : This patch adjusts gimple_range_op_handler constructor to also
check if a builtin function call might have a range_operator object
available, and if so, return that. This initial conversion also adds
CFN_BUILT_IN_CONSTANT_P as the first builtin, removing it from the big
switch in gimple-range-fold.cc.
Patch 8-16 : Moves SIGNBIT, TOUPPER/LOWER, POPCOUNT, CLZ, CTZ, CLRSB,
UBSAN*, STRLEN, and GOACC to range-ops.
patch 17: Finally, moves CFA_BUILT_IN_PARITY to range-ops, and removes
the builtin-function code checks from range_of_call in gimple_range-fold.cc
These patches all bootstrap on x86_64-pc-linux-gnu with no regressions.
Performance wise, it all ends up as approximately a wash. (VRP a hair
slower, threading a hair faster)
Pushed.
Andrew
- [PATCH 00/17] Move builtin functions to ran... Andrew MacLeod via Gcc-patches
-