match-sat-alu.pd: Allow wider ranges for signed saturating truncation
GCC now supports wider intermediate types for signed saturating truncation, improving optimization opportunities.
GCC’s match-sat-alu.pd pattern matcher now allows a wider unsigned range type for signed saturating truncation, provided it preserves all input values. This change enables more optimizations, especially in vector processing. An AArch64 example shows a reduction from a complex vector sequence to a simpler one with fewer iterations.
In Details
The signed_integer_sat_trunc pattern in match-sat-alu.pd has been broadened to allow a lossless unsigned range type that is at least as wide as the signed input. Previously, it required the range type to have the same precision as the input, limiting optimizations. The change also continues to require a narrowing result type and constructs offset/limit constants at the range precision using wide-int operations. This enhancement is beneficial for vectorization, as demonstrated by an AArch64 example showing a significant reduction in vector operations and iterations.
- signed saturating truncation
- A conversion to a narrower signed integer type where the result is clamped to the minimum or maximum representable value of the target type if the original value falls outside that range. This prevents overflow by capping the result.
- unsigned range type
- In the context of saturating operations, this refers to an intermediate unsigned data type used to hold values during calculations before they are potentially clamped and converted to the final signed result type. A 'lossless' range type means it can represent all possible values of the input without overflow.
- vector processing
- A type of computation where the same operation is applied to multiple data elements simultaneously, typically using specialized processor instructions (SIMD). This significantly speeds up tasks involving large datasets.
- pattern matcher
- A component within a compiler that identifies specific sequences of intermediate representation (IR) operations that can be replaced by more efficient instructions or simpler IR forms. This is crucial for optimization.