AMDGPU: Re-organize the outer loop of SILoadStoreOptimizer

[nhaehnle]

Summary

The entire algorithm operates per basic-block, so for cache locality
it should be better to re-optimize a basic-block immediately rather than
in a separate loop.

I don't have performance measurements.

Change-Id: I85106570bd623c4ff277faaa50ee43258e1ddcc5

[rampitec]

LGTM

llvm/trunk/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

 //===- SILoadStoreOptimizer.cpp -------------------------------------------===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // This pass tries to fuse DS instructions with close by immediate offsets.
 // This will fuse operations such as
 //  ds_read_b32 v0, v2 offset:16
 //  ds_read_b32 v1, v2 offset:32
 // ==>
 //   ds_read2_b32 v[0:1], v2, offset0:4 offset1:8
 //
-// The same is done for certain SMEM opcodes, e.g.:
+// The same is done for certain SMEM and VMEM opcodes, e.g.:
 //  s_buffer_load_dword s4, s[0:3], 4
 //  s_buffer_load_dword s5, s[0:3], 8
 // ==>
 //  s_buffer_load_dwordx2 s[4:5], s[0:3], 4
 //
 //
 // Future improvements:
 //
bool SILoadStoreOptimizer::runOnMachineFunction(MachineFunction &MF) {
   MRI = &MF.getRegInfo();
   AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
 
   assert(MRI->isSSA() && "Must be run on SSA");
 
   DEBUG(dbgs() << "Running SILoadStoreOptimizer\n");
 
   bool Modified = false;
-  CreatedX2 = 0;
 
-  for (MachineBasicBlock &MBB : MF)
+  for (MachineBasicBlock &MBB : MF) {
+    CreatedX2 = 0;
     Modified |= optimizeBlock(MBB);
 
     // Run again to convert x2 to x4.
-  if (CreatedX2 >= 1) {
-    for (MachineBasicBlock &MBB : MF)
+    if (CreatedX2 >= 1)
       Modified |= optimizeBlock(MBB);
   }
 
   return Modified;
 }