| Age | Commit message (Collapse) | Author |
|---|
|
The main practical change here is that things that used to be `IRValue`s, like literals, are now being expressed as instructions in the global scope.
In order to validate that things are actually being handled correctly, this change introduces an explicit "validation" pass that can be run on the IR to check for different invariants (although it doesn't check many of the important ones right now). I've left the validation pass turned off by default, but with a command-line flag to enable it. We may want to make it be on by default in debug builds, just to keep us honest. The main invariant for the moment is that when on IR instruction is used as an operand to another, it had better come from the same IR module.
Some of the existing passes were violating this rule, in particular when it came to cloning of witness tables related to global generic parameter substitution. Those features can in theory be handled better now by allowing `specialize` instructions at other scopes, but I didn't want to over-complicate this change, so I make just enough fixes to ensure that these steps always clone witness tables they get from the "symbols" on an IR specialization context. In order for this to work when recursively specializing, I had to ensure that the logic for generic specialization had a notion of a "parent" specialization context that it would fall back to to perform cloning when necessary.
This change keeps the logic that was caching and re-using the instructions for literal values within a module, but adds some logic that isn't really being tested right now for picking the right parent instruction to insert a constant instruction into. This logic doesn't trigger right now because all of the cases we are using it on have zero operands (and so they always get "hoisted" to the global scope), but eventually for things like types we want to be able to support instructions with operands (e.g., `vector<float, 4>`) and handle the case where some of those operands come from different scopes (e.g., when nested inside a generic).
The final change here is mostly cosmetic: the `IRBuilder` is now more abstract about where insertion occurs: it tracks a single `IRParentInst` to insert into, and then an optional `IRInst` to insert before. In the common case, that parent is an `IRBlock`, but it could conceivably also be the global scope, or a witness table, etc. Use sites where we used to change those fields directly now use distinct methods `setInsertInto(parent)` and `setInsertBefore(inst)` which capture the two cases we care about. Accessors are also defined to extract the current block (if the current parent is a block), and the current "function" (global value with code, if the current parent is a global value with code, or a block inside one).
With this work in place, it should be possible for a follow-on change to start putting `specialize` instructions at the global scope and thus clean up some of the on-the-fly specialization work. This work should also help with some of the requirements around a distinct IR-level type system and more explicit generics.
|
|
* IR: "everything is an instruction"
This change tries to streamline the representation of the IR in the following ways:
* Every IR value is an instruction (there is no `IRValue` type any more)
* All IR values that can contain other values share a single base (`IRParentInstruction`)
* Dynamic casts to specific IR instruction types can be accomplished with a new `as<Type>(inst)` operation, that uses the IR opcode to implement casts.
The biggest change in terms of number of lines is getting rid of `IRValue`. The diff here could probably be smaller if I'd just done `typedef IRInst IRValue;`.
Along the way I also renamed the `getArg`/`getArgs`/`getArgCount` combination over to `getOperand`/`getOperands`/`getOperandCount` to avoid being confusing when we have something like a `call` instruction where the "arguments" of the call don't line up with the operands of the instruction.
I also tried to clean up the representation of lists of child instructions to try to make it easier to iterate over them with C++ range-based `for` loops. Developers still need to be careful about mutating the contents of a block while iterating over it in this fashion (e.g. if you remove the "current" element, the iteration will end prematurely).
Probably the thorniest change here is that parameters are now just represented as the first N instructions in a block, which means:
* We need to perform a linear search to find the end of the parameter list. This is probably not often a problem, because usually you would be iterating over the parameters anyway, and that will be linear in the number of parameters.
* Algorithms that iterate over a block either need to ignore parameters, treat parameters just like other instructions, or somehow cleave the list into the range of parameters, and the range of "ordinary" instructions (which involves the same linear search above).
* When inserting into a block, we need to be careful not to insert instructions at invalid locations (e.g., insert a temporary before the parameters, or insert a parameter in the middle of the code). I can't pretend that I've handled the details of that here. (This is no different than having to make the same adjustments for phi nodes in a typical SSA representation)
* One possible future-proof approach is to implement a pass that sorts the instructions in a block so that parameters always come first. That would let us implement passes without caring about this detail, and then clean up right before any pass that cares about the relative order of parameters and other instructions.
The current change is missing any work to make literals and other instructions that used to be `IRValue`s properly nest inside of their parent module. Right now these instructions are just left unparented, and may actually end up being shared between distinct modules. Fixing that will need a follow-up change. The biggest challenge there is that it introduces instructions at the global scope that aren't `IRGlobalValue`s.
This change doesn't try to take advantage of any of the new flexibility (e.g., by nesting `specialize` instructions inside of witness tables). The goal is to do exactly what we were doing before, just with a different representation.
* Warning fix
|
|
* Initial work on validating "constexpr"-ness in IR
The underlying issue here is that certain operations in the target shading languages constrain their operands to be compile-time constants. A notable example is the optional texel offset parameter to the `Texture2D.Sample` operation.
When calling these operations in GLSL, the user is required to pass a "constant expression," and any variables in that expression must therefore be marked with the `const` qualifier (and themselves be initialized with constant expressions). Any GLSL output we generate must of course respect these rules.
When calling these operations in HLSL, the user is not so constrained. Instead, they can pass an arbitrary expression, which may involve ordinary variables with no particular markup, and then the compiler is responsible for determining if the actual value after simplification works out to be a constant. In some cases, the requirement that a value be constant might actually trigger things like loop unrolling. Also, it is okay to use a function parameter to determine such a constant expression, as long as the argument turns out to be a constant at all call sites.
The way we have decided to tackle these challenges in Slang is that we we propagate a notion of `constexpr`-ness through the IR. This is currently being tackled in `ir-constexpr.cpp` with a combination of forward and backward iterative dataflow:
* When the operands to an instruction are all `constexpr`, and the opcode is one we believe can be constant-folded, then we infer that the instruction *can* be evaluated as `constexpr`
* When instruction is required to be `constexpr`, then we infer that all of its operands are also required to be `constexpr`.
If this process ever infers that a function parameter is required to be `constexpr`, then we might have to continue propagation at all the call sites to that function.
If after all the propagation is done, there are any cases where an instruction is *required* to be `constexpr`, but it *can't* be `constexpr` (we weren't able to infer `constexpr`-ness for its operands), then we issue an error.
This implementation encodes the idea of `constexpr`-ness in the IR as part of the type system, using a simplified notion of rates. This change adds a `RateQualifiedType` that can represent `@R T`, and then introduces a `ConstExprRate` that can be used for `R`. Many accessors for the type information on IR nodes were updated to distinguish when one wants the "full" type of an IR value (which might include rate information) vs. just the "data" type.
A `constexpr` qualifier was added in the front-end, and is being used to decorate the texel offset parameter for `Texture2D.Sample`. Lowering from AST to IR looks for this qalifier and infers when a function parameter must be typed as `@ConstExpr T` instead of just `T`.
There are lots of limitations and gotchas in the implementation so far:
* The `@ConstExpr` rate is the only one added in this change, but it seems clear that the conceptual `ThreadGroup` rate that was added to represent `groupshared` should probably get folded into the representation.
* I'm not 100% pleased with how many places in the IR I have to special-case for rate-qualified types. At the same type, pulling out rate as a distinct field on `IRValue` would probably require that we pay attention to rate everywhere.
* I've added a test case to show that we can issue errors when users fail to provide a constant expression for the texel offset, but the actual error message isn't great because it doesn't indicate *why* a constant expression was required. Realistically the "initial IR" should contain a few more decorations we can use to relate error conditions back to the original code (even if this is in a side-band structure).
* I've added a test case that is supposed to show that we can back-propagate `constexpr`-ness to local variables, and I've manually confirmed that it works for Vulkan/SPIR-V output, but the level of Vulkan support in `render_test` today means I can't enable the test for check-in.
* While I'm attempting to propagate `@ConstExpr` information from callees to callers, I haven't implemented any logic to specialize callee functions based on values at call sites.
* In a similar vein, there is no handling of control-flow dependence in the current code. If we infer that a phi (block parameter) needs to be `@ConstExpr`, then it isn't actually enough to require that the inputs to the phi (arguments from predecessor blocks) are all `@ConstExpr` because we also need any control-flow decisions that pick which incoming edge we take to be `@ConstExpr` as well.
* As a practical matter, implicit propagation of `@ConstExpr` from a function body to a function parameter should only be allowed for functions that are "local" to a module. Any function that might be accessed from outside of a module should really have had its `@ConstExpr` parameter marked manually, and our pass should validate that they follow their own rules. Right now we have no kind of visibility (`public` vs `private`) system, so I'm kind of ignoring this issue.
While that is a lot of gaps, this is also just enough code to get the Falcor MultiPassPostProcess example working, so I'm inclined to get it checked in.
* Fixup: missing expected output for test
* Fixup: disable test that relies on [unroll] for now
|
