This patch adds a new NVPTX back-end to LLVM which supports code generation for NVIDIA PTX 3.0. This back-end will (eventually) replace the current PTX back-end, while maintaining compatibility with it.

The new target machines are:

nvptx (old ptx32) => 32-bit PTX
nvptx64 (old ptx64) => 64-bit PTX

The sources are based on the internal NVIDIA NVPTX back-end, and
contain more functionality than the current PTX back-end currently
provides.

NV_CONTRIB

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@156196 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/NVPTX/NVPTXSplitBBatBar.cpp

@@ -0,0 +1,77 @@
+//===- NVPTXSplitBBatBar.cpp - Split BB at Barrier  --*- C++ -*--===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+// Split basic blocks so that a basic block that contains a barrier instruction
+// only contains the barrier instruction.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Function.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/IntrinsicInst.h"
+#include "llvm/Support/InstIterator.h"
+#include "NVPTXUtilities.h"
+#include "NVPTXSplitBBatBar.h"
+
+using namespace llvm;
+
+namespace llvm {
+FunctionPass *createSplitBBatBarPass();
+}
+
+char NVPTXSplitBBatBar::ID = 0;
+
+bool NVPTXSplitBBatBar::runOnFunction(Function &F) {
+
+  SmallVector<Instruction *, 4> SplitPoints;
+  bool changed = false;
+
+  // Collect all the split points in SplitPoints
+  for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
+    BasicBlock::iterator IB = BI->begin();
+    BasicBlock::iterator II = IB;
+    BasicBlock::iterator IE = BI->end();
+
+    // Skit the first intruction. No splitting is needed at this
+    // point even if this is a bar.
+    while (II != IE) {
+      if (IntrinsicInst *inst = dyn_cast<IntrinsicInst>(II)) {
+        Intrinsic::ID id = inst->getIntrinsicID();
+        // If this is a barrier, split at this instruction
+        // and the next instruction.
+        if (llvm::isBarrierIntrinsic(id)) {
+          if (II != IB)
+            SplitPoints.push_back(II);
+          II++;
+          if ((II != IE) && (!II->isTerminator())) {
+            SplitPoints.push_back(II);
+            II++;
+          }
+          continue;
+        }
+      }
+      II++;
+    }
+  }
+
+  for (unsigned i = 0; i != SplitPoints.size(); i++) {
+    changed = true;
+    Instruction *inst = SplitPoints[i];
+    inst->getParent()->splitBasicBlock(inst, "bar_split");
+  }
+
+  return changed;
+}
+
+// This interface will most likely not be necessary, because this pass will
+// not be invoked by the driver, but will be used as a prerequisite to
+// another pass.
+FunctionPass *llvm::createSplitBBatBarPass() {
+  return new NVPTXSplitBBatBar();
+}