Fix uninitialised vars in memory-pool setup in offload runtime

This patch fixes a bug in the AMDGPU offload plugin during initialisation of
memory pools.

The function preAllocateDeviceMemoryPool is called (only) from a location where
memory pool metadata structures have already been enumerated and initialised
(in initImpl, by calling initMemoryPools).  But it then calls
retrieveMemoryPools *again*, but does not go on to initialise the newly-created
duplicate metadata properly before using it, meaning some fields (Segment,
GlobalFlags) are left with uninitialised data.  That can lead to unpredictable
behaviour.

The fix is just to remove the "retrieval", i.e. allocation, of duplicate pool
metadata.

Change-Id: I7cd7e46c0f2c655a8007299595c553f33451ccb1

offload/plugins-nextgen/amdgpu/src/rtl.cpp

@@ -4096,9 +4096,6 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
   /// Allocate and zero initialize a small memory pool from the coarse grained
   /// device memory of each device.
   Error preAllocateDeviceMemoryPool() {
-    Error Err = retrieveAllMemoryPools();
-    if (Err)
-      return Plugin::error("Unable to retieve all memmory pools");
 
     void *DevPtr;
     for (AMDGPUMemoryPoolTy *MemoryPool : AllMemoryPools) {
@@ -4109,7 +4106,7 @@ struct AMDGPUDeviceTy : public GenericDeviceTy, AMDGenericDeviceTy {
         DevPtr = nullptr;
         size_t PreAllocSize = hsa_utils::PER_DEVICE_PREALLOC_SIZE;
 
-        Err = MemoryPool->allocate(PreAllocSize, &DevPtr);
+        Error Err = MemoryPool->allocate(PreAllocSize, &DevPtr);
         if (Err)
           return Plugin::error("Device memory pool preallocation failed");