Vectorization pattern for constant multiplication now correctly sets vector types.
A bug preventing correct vector type assignment for constant multiplication patterns in vectorization has been resolved.
This commit addresses a bug in GCC’s vectorization pass where, during the pattern recognition of integer multiplication by a constant, the appropriate unsigned vector type was not being assigned to the resulting pattern statement. This omission could lead to incorrect vector types, particularly affecting code dealing with undefined behavior checks for vector reduction and potentially causing wrong code in non-reduction cases. The fix ensures that the correct vector type is explicitly assigned, improving the accuracy of vectorized code.
In Details
This patch targets vect_synth_mult_by_constant in tree-vect-patterns.cc, resolving an issue where the vector type for a synthesized constant multiplication pattern, especially when an unsigned conversion was required, was not correctly updated. This oversight meant the pattern statement retained the signed result's vector type, confusing subsequent analyses, such as UB checks for reduction vectorization, and potentially leading to wrong code in general. The fix ensures the emitted sequence statements for the pattern are assigned the correct vector type, maintaining type consistency and co…
For Context
Compilers try to make your code faster by using "vectorization," which allows a single instruction to operate on multiple pieces of data at once. This is especially useful for loops that perform the same operation many times. This commit fixes a bug in GCC's vectorization process related to how it handles multiplying integers by a constant value. The problem was that when the compiler tried to create a vectorized version of such a multiplication, and a conversion to an 'unsigned' integer type was needed, it sometimes assigned the wrong data type (a 'signed' vector type instead of an 'unsigned' one) to the vectorized instruction. This could lead to incorrect calculations or errors, particularly in specialized scenarios like reducing a vector to a single result. The fix ensures that the compiler assigns the correct data type, making vectorization more reliable and preventing potential errors in your programs.