Add option for backends to run vectorization passes

Author: gbaraldi

src/interface.jl

@@ -222,6 +222,9 @@ end
 # Has the runtime available and does not require special handling
 uses_julia_runtime(@nospecialize(job::CompilerJob)) = false
 
+# Is it legal to run vectorization passes on this target
+can_vectorize(@nospecialize(job::CompilerJob)) = false
+
 # Should emit PTLS lookup that can be relocated
 dump_native(@nospecialize(job::CompilerJob)) = false
 

src/optim.jl

@@ -31,7 +31,7 @@ function buildNewPMPipeline!(mpm, @nospecialize(job::CompilerJob), opt_level)
     add!(mpm, NewPMFunctionPassManager()) do fpm
         buildLoopOptimizerPipeline(fpm, job, opt_level)
         buildScalarOptimizerPipeline(fpm, job, opt_level)
-        if uses_julia_runtime(job) && opt_level >= 2
+        if (uses_julia_runtime(job) || can_vectorize(job)) && opt_level >= 2
             # XXX: we disable vectorization, as this generally isn't useful for GPU targets
             #      and actually causes issues with some back-end compilers (like Metal).
             # TODO: Make this not dependent on `uses_julia_runtime` (likely CPU), but it's own control