gcc: Prevent vectorizing conversions from _BitInt with incorrect extension
Fixes vectorization of _BitInt conversions in GCC by adding a check for proper sign/zero extension.
This commit improves GCC’s vectorization capabilities by adding a crucial check to prevent the vectorization of conversions from _BitInt types. Specifically, it ensures that _BitInt values are appropriately sign-extended or zero-extended to match the target mode before vectorization occurs, avoiding potential incorrect results. A new test case, vect-bitint-1.c, is included to verify this correction.
In Details
The vectorization pass in GCC, specifically within tree-vect-stmts.cc, needed an update to correctly handle conversions involving the _BitInt type. The vectorizable_conversion function now includes a check to ensure that _BitInt values are sign- or zero-extended to the required mode precision. Without this, converting non-mode-precision _BitInt values could lead to incorrect vectorized operations. This impacts the tree-vect pass and its interaction with arbitrary-precision integer types.
- _BitInt
- A GCC extension for arbitrary-precision integers, allowing developers to specify integer types with a width not conforming to standard C integer types.
- vectorization
- An optimization technique where a single instruction is used to perform the same operation on multiple data points simultaneously, leveraging SIMD (Single Instruction, Multiple Data) capabilities of modern processors.
- sign extension
- A method of converting a signed integer to a larger integer type by replicating the most significant bit (the sign bit) to fill the new higher-order bits. This preserves the number's sign and value.
- zero extension
- A method of converting an unsigned integer (or a signed integer treated as unsigned) to a larger integer type by filling the new higher-order bits with zeros. This preserves the value for unsigned numbers.