tree-optimization/113078 - conditional subtraction reduction vectorization
commitcac9d2d2346bf06b29b34e12cf0a005c37eacdc9
authorRichard Biener <rguenther@suse.de>
Wed, 10 Jan 2024 13:13:25 +0000 (10 14:13 +0100)
committerRichard Biener <rguenther@suse.de>
Wed, 10 Jan 2024 14:21:37 +0000 (10 15:21 +0100)
tree2d107d2759fd1a4a4f3fa575b17e28099f8f3668
parentf8a70fb21767883f6d3c13e22be8d307dd98f342
tree-optimization/113078 - conditional subtraction reduction vectorization

When if-conversion was changed to use .COND_ADD/SUB for conditional
reduction it was forgotten to update reduction path handling to
canonicalize .COND_SUB to .COND_ADD for vectorizable_reduction
similar to what we do for MINUS_EXPR.  The following adds this
and testcases exercising this at runtime and looking for the
appropriate masked subtraction in the vectorized code on x86.

PR tree-optimization/113078
* tree-vect-loop.cc (check_reduction_path): Canonicalize
.COND_SUB to .COND_ADD.

* gcc.dg/vect/vect-reduc-cond-sub.c: New testcase.
* gcc.target/i386/vect-pr113078.c: Likewise.
gcc/testsuite/gcc.dg/vect/vect-reduc-cond-sub.c [new file with mode: 0644]
gcc/testsuite/gcc.target/i386/vect-pr113078.c [new file with mode: 0644]
gcc/tree-vect-loop.cc