AMDGPU/SI: Improve SILoadStoreOptimizer and run it before the scheduler

Summary:
The SILoadStoreOptimizer can now look ahead more then one instruction when
looking for instructions to merge, which greatly improves the number of
loads/stores that we are able to merge.

Moving the pass before scheduling avoids increasing register pressure after
the scheduler, so that the scheduler's register pressure estimates will be
more accurate.  It also gives more consistent results, since it is no longer
affected by minor scheduling changes.

Reviewers: arsenm

Subscribers: arsenm, kzhuravl, llvm-commits

Differential Revision: https://reviews.llvm.org/D23814

llvm-svn: 279991

llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

@@ -497,6 +497,7 @@ void GCNPassConfig::addMachineSSAOptimization() {
   // XXX - Can we get away without running DeadMachineInstructionElim again?
   addPass(&SIFoldOperandsID);
   addPass(&DeadMachineInstructionElimID);
+  addPass(&SILoadStoreOptimizerID);
 }
 
 void GCNPassConfig::addIRPasses() {
@@ -533,17 +534,6 @@ bool GCNPassConfig::addGlobalInstructionSelect() {
 #endif
 
 void GCNPassConfig::addPreRegAlloc() {
-  if (getOptLevel() > CodeGenOpt::None) {
-    // Don't do this with no optimizations since it throws away debug info by
-    // merging nonadjacent loads.
-
-    // This should be run after scheduling, but before register allocation. It
-    // also need extra copies to the address operand to be eliminated.
-
-    // FIXME: Move pre-RA and remove extra reg coalescer run.
-    insertPass(&MachineSchedulerID, &SILoadStoreOptimizerID);
-    insertPass(&MachineSchedulerID, &RegisterCoalescerID);
-  }
 
   addPass(createSIShrinkInstructionsPass());
   addPass(createSIWholeQuadModePass());