Failure to optimize pattern to x * 8

[GabrielRavier]

int r3(int n, int q)
{
	return n - q * -7;
}

This can be optimised to return (n + q * 8) - q; on architectures where n + q * 8 can be done in a single instruction (which includes x86). This transformation is done by GCC, but not by LLVM.

See also https://godbolt.org/z/4anx8846f for a Godbolt comparison

[llvmbot]

@llvm/issue-subscribers-backend-x86

[RKSimon]

Shouldn't be too hard to add this to X86DAGToDAGISel::matchAddress - did you have other targets in mind?

[Kmeakin]

Also missing on AArch64: https://godbolt.org/z/nT4a7Wor4

[llvmbot]

@llvm/issue-subscribers-backend-aarch64