Simplification includes dropping dependencies onto
`FunctionMetadataCachePass` and `LoadModelWrapperPass`, as well
as removing "StructName" logic outright.
Fix the needed attributes to allow navigation from the use of an
artificial return struct for raw functions to their definition in
`types-and-globals.h`.
This commit sets a name for types we use in OpaqueExtractValues.
These are StructTypes that can only be returned from isolated functions
with RawFunctionType prototype on the model, or from helpers that do not
represent isolated functions.
The name is required because when we will converto to MLIR LLVM Dialect,
there are checks in place that forbid unnamed types.
The code re-uses code that is used to print C code, in order to make
sure that the struct names we use for this are the same as those we emit
in C, to ease debugging.
This commit also cleans up the unittest to make it more concise, while
still testing the same things.
Now extractvalue instruction are replaced by dedicated
OpaqueExtractValue custom opcode, that prevents LLVM from doing strange
things with extractvalues during optimizations (such as e.g. sinking).
This is important since extractvalue instructions and struct-typed
values in general in our LLVM IR are not real first-class citizens, but
only a byproduct of the binary lifting process, and they actually
represent bundles of registers that are returned from isolated
functions.
ModelGEP calls in LLVM IR now return an integer that has the same
size of the field being addressed by the ModelGEP.
AddressOf calls, instead, can accept any integer size as argument but
always return a pointer-sized integer.
Calls to `OpaqueExtractValue()` are meant to replace `extractvalue`s
found in the LLVM IR. Since the type of an `OpaqueExtractValue` is
identified by both the return type (extracted value) and the first
argument's type (aggregate operand of the `extractvalue` instruction),
we need to consider both when building the associated FunctionPool.
Previously, we were identifying each `OpaqueExtractValue` variant using
only the returned value, which was wrong. In fact, if we have two
`extractvalue` instructions that extract a value of the same type
(e.g. i32) from two different aggregate types (e.g. structA and
structB), we have to define two different `OpaqueExtractValue`: one
that returns an i32 and has a parameter of type structA, and one that
returns an i32 and has a parameter of structB. If we use only the
return type, we are not able to distinguish the two.
Change company name to "rev.ng Labs Srl" in all license headers
to reflect changed company name and legal status
Add missing license headers to files that didn't have one
* RemoveExtractValues transforms every `extractvalue` instruction
into an opaque call. This prevents the optimization pipeline from
moving around and optimizing `extractvalue`s, since they have a
particular meaning in our IR
* RestoreExtractValues transforms such opaque calls back to regular
`extractvalue`s at the end of the pipeline
Alessandro Di Federico