[llvm] [AArch64] Fix resource length computation for STP. (PR #81749)

[Florian Hahn via llvm-commits]

fhahn wrote:

> Thanks for the explanation, I hadn't known where store-pair suppression was useful when looking at it in the past. The new logic looks OK to me.

Thanks. Another case where it is likely detrimental is when we create an LDP for 2 loads, but then not form STPs for a subsequent pair of stores to the same address, but that could be adjusted separately. This patch should drastically reduce the number of suppressed STPs, especially on CPUs with a larger number of vector units.

https://github.com/llvm/llvm-project/pull/81749