[Pipeliner] Add two pragmas to control software pipelining optimization

#pragma clang loop pipeline(disable)
  
    Disable SWP optimization for the next loop.
    “disable” is the only possible value.
  
#pragma clang loop pipeline_initiation_interval(number)
  
    Set value of initiation interval for SWP
    optimization to specified number value for
    the next loop. Number is the positive value
    greater than 0.
  
These pragmas could be used for debugging or reducing
compile time purposes. It is possible to disable SWP for
concrete loops to save compilation time or to find bugs
by not doing SWP to certain loops. It is possible to set
value of initiation interval to concrete number to save
compilation time by not doing extra pipeliner passes or
to check created schedule for specific initiation interval.

That is llvm part of the fix

Clang part of fix: https://reviews.llvm.org/D55710

Patch by Alexey Lapshin!

Differential Revision: https://reviews.llvm.org/D56403

llvm-svn: 351923

1 files changed, 76 insertions, 7 deletions

diff --git a/llvm/lib/CodeGen/MachinePipeliner.cpp b/llvm/lib/CodeGen/MachinePipeliner.cpp
index 6e62cb5136c..90a54e423ec 100644
--- a/llvm/lib/CodeGen/MachinePipeliner.cpp
+++ b/llvm/lib/CodeGen/MachinePipeliner.cpp
@@ -210,8 +210,11 @@ bool MachinePipeliner::scheduleLoop(MachineLoop &L) {
   }
 #endif
 
-  if (!canPipelineLoop(L))
+  setPragmaPipelineOptions(L);
+  if (!canPipelineLoop(L)) {
+    LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n");
     return Changed;
+  }
 
   ++NumTrytoPipeline;
 
@@ -220,6 +223,50 @@ bool MachinePipeliner::scheduleLoop(MachineLoop &L) {
   return Changed;
 }
 
+void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) {
+  MachineBasicBlock *LBLK = L.getTopBlock();
+
+  if (LBLK == nullptr)
+    return;
+
+  const BasicBlock *BBLK = LBLK->getBasicBlock();
+  if (BBLK == nullptr)
+    return;
+
+  const Instruction *TI = BBLK->getTerminator();
+  if (TI == nullptr)
+    return;
+
+  MDNode *LoopID = TI->getMetadata(LLVMContext::MD_loop);
+  if (LoopID == nullptr)
+    return;
+
+  assert(LoopID->getNumOperands() > 0 && "requires atleast one operand");
+  assert(LoopID->getOperand(0) == LoopID && "invalid loop");
+
+  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
+    MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
+
+    if (MD == nullptr)
+      continue;
+
+    MDString *S = dyn_cast<MDString>(MD->getOperand(0));
+
+    if (S == nullptr)
+      continue;
+
+    if (S->getString() == "llvm.loop.pipeline.initiationinterval") {
+      assert(MD->getNumOperands() == 2 &&
+             "Pipeline initiation interval hint metadata should have two operands.");
+      II_setByPragma =
+          mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
+      assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive.");
+    } else if (S->getString() == "llvm.loop.pipeline.disable") {
+      disabledByPragma = true;
+    }
+  }
+}
+
 /// Return true if the loop can be software pipelined.  The algorithm is
 /// restricted to loops with a single basic block.  Make sure that the
 /// branch in the loop can be analyzed.
@@ -227,6 +274,9 @@ bool MachinePipeliner::canPipelineLoop(MachineLoop &L) {
   if (L.getNumBlocks() != 1)
     return false;
 
+  if (disabledByPragma)
+    return false;
+
   // Check if the branch can't be understood because we can't do pipelining
   // if that's the case.
   LI.TBB = nullptr;
@@ -285,7 +335,8 @@ void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) {
 bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) {
   assert(L.getBlocks().size() == 1 && "SMS works on single blocks only.");
 
-  SwingSchedulerDAG SMS(*this, L, getAnalysis<LiveIntervals>(), RegClassInfo);
+  SwingSchedulerDAG SMS(*this, L, getAnalysis<LiveIntervals>(), RegClassInfo,
+                        II_setByPragma);
 
   MachineBasicBlock *MBB = L.getHeader();
   // The kernel should not include any terminator instructions.  These
@@ -308,6 +359,20 @@ bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) {
   return SMS.hasNewSchedule();
 }
 
+void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) {
+  if (II_setByPragma > 0)
+    MII = II_setByPragma;
+  else
+    MII = std::max(ResMII, RecMII);
+}
+
+void SwingSchedulerDAG::setMAX_II() {
+  if (II_setByPragma > 0)
+    MAX_II = II_setByPragma;
+  else
+    MAX_II = MII + 10;
+}
+
 /// We override the schedule function in ScheduleDAGInstrs to implement the
 /// scheduling part of the Swing Modulo Scheduling algorithm.
 void SwingSchedulerDAG::schedule() {
@@ -334,9 +399,11 @@ void SwingSchedulerDAG::schedule() {
   if (SwpIgnoreRecMII)
     RecMII = 0;
 
-  MII = std::max(ResMII, RecMII);
-  LLVM_DEBUG(dbgs() << "MII = " << MII << " (rec=" << RecMII
-                    << ", res=" << ResMII << ")\n");
+  setMII(ResMII, RecMII);
+  setMAX_II();
+
+  LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II
+                    << " (rec=" << RecMII << ", res=" << ResMII << ")\n");
 
   // Can't schedule a loop without a valid MII.
   if (MII == 0)
@@ -1744,8 +1811,9 @@ bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) {
     return false;
 
   bool scheduleFound = false;
+  unsigned II = 0;
   // Keep increasing II until a valid schedule is found.
-  for (unsigned II = MII; II < MII + 10 && !scheduleFound; ++II) {
+  for (II = MII; II <= MAX_II && !scheduleFound; ++II) {
     Schedule.reset();
     Schedule.setInitiationInterval(II);
     LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n");
@@ -1817,7 +1885,8 @@ bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) {
       scheduleFound = Schedule.isValidSchedule(this);
   }
 
-  LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound << "\n");
+  LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound << " (II=" << II
+                    << ")\n");
 
   if (scheduleFound)
     Schedule.finalizeSchedule(this);