llvm-project

f57abf51 - [SPIRV] Promote scalar arguments to vector for `OpExtInst` in `generateExtInst` instead of `SPIRVRegularizer` (#170155)

[SPIRV] Promote scalar arguments to vector for `OpExtInst` in `generateExtInst` instead of `SPIRVRegularizer` (#170155) This patch consist of 2 parts: * A first part that removes the scalar to vector promotion for built-ins in the `SPIRVRegularizer`; * and a second part that implements the promotion for built-ins from scalar to vector in `generateExtInst`. The implementation in `SPIRVRegularizer` had several issues: * It rolled its own built-in pattern matching that was extremely permissive * the compiler would crash if the built-in had a definition * the compiler would crash if the built-in had no arguments * The compiler would crash if there were more than 2 function definitions in the module. * It'd be better if this was implemented as a module pass; where we iterate over the users of the function, instead of scanning the whole module for callers. This patch does the scalar to vector promotion just before the `OpExtInst` is generated. Without relying on the IR transformation. One change in the generated code from the previous implementation is that this version uses a single `OpCompositeConstruct` operation to convert the scalar into a vector. The old implementation inserted an element at the 0 position in an `undef` vector (using `OpCompositeInsert`); then copied that element for every vector element using `OpVectorShuffle`. This patch also adds a test (`OpExtInst_vector_promotion_bug.ll`) that highlights an issue in the builtin pattern matching that we're using: our pattern matching doesn't consider the number of arguments, only the demangled name, first and last arguments (`min(int,int,int)` matches the same builtin as `min(int, int)`).