|
* Cleanup refactoring work around the IR builder
We have some long-term goals for the IR that require a more centralized and disciplined set of rules for how IR instructions get created/emitted. I had been working on trying to set things up so that all IR instruction creation goes through a single bottleneck point, but the non-trivial work in that branch was getting drowned out by the sheer volume of cleanup and refactoring changes. This change tries to pull together several of the more important cleanups.
The big pieces are:
* `IRBuilder` and `SharedIRBuilder` now protect their data members and rely on users to initialize them more directly via constructor of an `init()` method. This change affects a *bunch* of sites where `IRBuilder`s were created. I changed use sites to use the constructors whenever possible, and to use `init()` in cases where we had longer-lived builders that needed to be initialized multiple times.
* The insertion location for the `IRBuilder` now uses an encapsulated type called `IRInsertLoc`. This new type can replace what used to be just two `IRInst*` fields in the builder, and also covers some new functionality (if we ever want to take advantage of it). Very little client code cares about this change, but it is still a nice cleanup in terms of making things more explicit.
* The creation of an `IRModule` has been moded *out* of `IRBuilder`, because in practice we `IRBuilder` always wants to be associated with a pre-existing `IRModule` at creation time (via its `SharedIRBuilder`). There is now an `IRModule::create()` operation instead. This required changing the sequencing at many `IRModule` creation sites, since most had been contriving to make an `IRBuilder` first. There were also several cleanups because code had been carelessly using non-reference-counted pointers for `IRModule`s in ways that broke now that `IRModule::create()` always returns a `RefPtr`.
* The core operations to actually allocate memory for IR instructions were moved into `IRModule` (since they interact with the memory pool that the module owns). These *were* called `createEmptyInst()` but have been renamed into `_allocateInst()`. In principle these seem like they should only be needed to be called by the `IRBuilder`, but in practice they are also needed by the IR deserialization logic.
* A few core operations for emitting IR instructions that were associted with `IRBuilder` were moved to actually be methods on `IRBuilder`. First is `_findOrEmitConstant` which is the primary bottleneck for creating simple scalar constant values. Another is `_createInst` (formerly part of the templated `createInstImpl` along with `createInstWithSizeImpl`) which is the main bottleneck for allocation and initialization of any instruction other than a constant (well, the `IRModuleInst` is the other exception...). Finally, there is also `_maybeSetSourceLoc()`, which is obvious to scope inside the `IRBuilder` once it is protecting the source-location info.
Notes:
* The `minSizeInBytes` parameter to `_createInst()` might not actually be needed at all. At this point any `IRInst` subtypes that need data allocated for things other than their operands already get created manually via `_allocateInst` or `_findOrEmitConstant`, so I *think* we could remove that part. I will handle that in a subsequent cleanup if it turns out to be the case.
* There is one IR pass (`slang-ir-string-hash.cpp`) that is using manual `_allocateInst()` instead of going through an `IRBuilder`. It could be easily cleaned up to not do so (and I will probably make that change down the line), but for now I wanted to avoid doing anything that wasn't close to pure refactoring if I could.
* At this point in our design an `IRBuilder` is a very lightweight thing - it basically just owns the insertion location plus a source location to write into instructions. A lot of our code currently treats `IRBuilder`s like they are expensive and/or need to be re-used (which leads to them being used in more mutable/stateful ways). It is quite likely that as we clean up other aspects of the implementation of IR creation/emission we can make `IRBuilder` use feel more lightweight in ways that can streamline and simplify code.
* The next step for this work is to identify the different paths that eventually lead to `_createInst()` being called, and unify them at a single bottleneck operation that can own the decisions around when to create an instruction vs. when to re-use an existing one (rather than those decisions being baked into the various `IRBuilder` subroutines that create instructions of the various subtypes).
* fixup: gcc/clang C++ spec details
|
|
The Slang IR builder has a notion of "hoistable" instructions, which are basically those instructions that represent a pure side-effect-free operation on their operands, and which can and should be deduplicated. Most types are "hoistable" instructions.
In order to make deduplication of hoistable instructions work, we need to emit them at the right location. Consider if we had:
```hlsl
void myFunc<T>(...)
{
if(someCondition) {
vector<T, 4> a = ...;
...
} else {
vector<T, 4> b = ...;
}
}
```
The IR instruction that represents `vector<T,4>` can't be inserted at the global scope, because then the parameter `T` would not be visible to it. That instruction also shouldn't be inserted into the same block that declares `a`, because then the instruction itself wouldn't be visible at the point where `b` is declared.
The IR builder already has logic to pick the right parent instruction. In the example given, the IR instruction for `vector<T,4>` should be inserted into the body of the IR generic, but outside of the IR function that represents `myFunc`.
The problem this change fixes is that while the logic was picking the *parent* for a hoistable instruction correctly, it wasn't putting much care into pick the insertion *location*. The existing strategy amounted to:
* If the IR builder was set with an insertion location inside the chosen parent, then use that insertion location
* Otherwise, insert at the end of the chosen parent
Neither of those options is perfect. Either could lead to an instruction being inserted after one of its uses, and the second option could even lead to a type being inserted *after* the `return` instruction in a function/generic, which violates another structural invariant of our IR (that every block must end with a terminator, and terminators must only appear at the end of blocks).
This change updates the rules as follows:
* If the type of the instruction being created, or any of its operands are in the chosen parent, then insert immediately after whichever of those instructions is last in that parent.
* Otherwise, insert before the first non-decoration, non-parameter child of the chosen parent
The combined effect of these two rules is now that we insert any hoistable instruction as early as we can in its parent, without violating the structural validity rules.
(One small exception to these rules is that if the parent is the module then we don't worry about ordering and just insert at the end, since order-of-declaration isn't significant at module scope in our IR)
All of our existing tests work with this new behavior, although there could conceivably be future cases that lead to complicated breakage. For example, if a pass looks at the first "ordinary" instruction in a block and saves it to use as an insertion point for parameter, and then proceeds to manipulate code in the block before going back and inserting parameters at the chosen location, there is a chance that a hoistable instruction might have been inserted before the chosen insertion point, leading to a parameter being inserted after an ordinary instruction. In general, though, code that works like that would already be playing a dangerous game in that it is manipulating instructions in a block while assuming the first instruction will remain fixed.
This change is currently just a refactor, but the underlying issue surfaced as a bug when I made other changes in a feature branch.
|
