[turbofan] Run memory optimizer last.

Run memory optimizer after all the other graph phases. This is
a step towards enabling allocation folding for arrays with
>16 elements because constant additions will be properly
constant-folded.

Bug: v8:8984
Change-Id: Ia3c78a3bd32264f4f83d3e20bd78abf240d9292b
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1539496Reviewed-by: Benedikt Meurer <bmeurer@chromium.org>
Commit-Queue: Jaroslav Sevcik <jarin@chromium.org>
Cr-Commit-Position: refs/heads/master@{#60465}

src/compiler/pipeline.cc

@@ -2107,16 +2107,13 @@ bool PipelineImpl::OptimizeGraph(Linkage* linkage) {
     RunPrintAndVerify(ControlFlowOptimizationPhase::phase_name(), true);
   }
 
+  Run<LateOptimizationPhase>();
+  RunPrintAndVerify(LateOptimizationPhase::phase_name(), true);
+
   // Optimize memory access and allocation operations.
   Run<MemoryOptimizationPhase>();
-  // TODO(jarin, rossberg): Remove UNTYPED once machine typing works.
   RunPrintAndVerify(MemoryOptimizationPhase::phase_name(), true);
 
-  // Lower changes that have been inserted before.
-  Run<LateOptimizationPhase>();
-  // TODO(jarin, rossberg): Remove UNTYPED once machine typing works.
-  RunPrintAndVerify(LateOptimizationPhase::phase_name(), true);
-
   data->source_positions()->RemoveDecorator();
   if (data->info()->trace_turbo_json_enabled()) {
     data->node_origins()->RemoveDecorator();