[llvm-bugs] [Bug 50942] New: [LoopVectorizer] Vectorization of running reduction/predication

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=50942

            Bug ID: 50942
           Summary: [LoopVectorizer] Vectorization of running
                    reduction/predication
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: lebedev.ri at gmail.com
                CC: llvm-bugs at lists.llvm.org

Consider https://godbolt.org/z/Wo86sfav1

void test(int pred, int* data, int width) {
    for(int i = 0; i != width; ++i) {
        int& out = data[i];
        pred += out;
        out = pred;
    }
}


So the first element was incremented by `pred`,
and each next element is incremented by the value of all preceding elements.
This is a common pattern in image processing.

Currently we don't recognize the PHI, and don't vectorize,
even though it's somewhat simple:
https://godbolt.org/z/sEob7fE6h 

That snippet as-is doesn't seem better vectorized:
https://godbolt.org/z/aMaqYz7f1
(better RThroughput, same cycle count, but more uOps/IPC)

However if we unroll x8 https://godbolt.org/z/Mb7vcGrzs
(and i do believe both loops unrolled still compute the same elt count)
i think we see a win: https://godbolt.org/z/EW77Max86
A ~third less cycles.

This makes sense because most of the computations there don't touch `pred`,
so they can be executed out-of-order, even though we won't be able to
finish and store until the previous group has finished processing.

The story will be somewhat different with two-element predictor,
different data types, etc.

Is this recipe  something that could fit into LV?

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20210630/fa90f227/attachment-0001.html>