Moving ManagedMemoryRewritePass when hybrid option is selected

Compiling with -polly-target=hybrid was causing Polly to occur two times
in the pipeline. The reason was how the ManagedMemoryRewritePass was
registered in the pass manager. ManagedMemoryRewritePass being a
ModulePass was forcing all previous passes to get recomputed. This
commit avoids Polly to appear two times in the pipeline registering the
ManagedMemoryRewritePass later in the pass manager.

Patch by Lorenzo Chelini <l.chelini@icloud.com>

Differential Revision: https://reviews.llvm.org/D59263

llvm-svn: 356965

polly/lib/Support/RegisterPasses.cpp

@@ -349,12 +349,9 @@ void registerPollyPasses(llvm::legacy::PassManagerBase &PM) {
     PM.add(polly::createPruneUnprofitablePass());
 
 #ifdef GPU_CODEGEN
-  if (Target == TARGET_HYBRID) {
+  if (Target == TARGET_HYBRID)
     PM.add(
         polly::createPPCGCodeGenerationPass(GPUArchChoice, GPURuntimeChoice));
-    PM.add(polly::createManagedMemoryRewritePassPass(GPUArchChoice,
-                                                     GPURuntimeChoice));
-  }
 #endif
   if (Target == TARGET_CPU || Target == TARGET_HYBRID)
     switch (Optimizer) {
@@ -388,6 +385,12 @@ void registerPollyPasses(llvm::legacy::PassManagerBase &PM) {
   }
 #endif
 
+#ifdef GPU_CODEGEN
+  if (Target == TARGET_HYBRID)
+    PM.add(polly::createManagedMemoryRewritePassPass(GPUArchChoice,
+                                                     GPURuntimeChoice));
+#endif
+
   // FIXME: This dummy ModulePass keeps some programs from miscompiling,
   // probably some not correctly preserved analyses. It acts as a barrier to
   // force all analysis results to be recomputed.