[NVPTX] Emit .pragma "nounroll" for loops marked with nounroll

[jingyue]

Summary

CUDA driver can unroll loops when jit-compiling PTX. To prevent CUDA
driver from unrolling a loop marked with llvm.loop.unroll.disable is not
unrolled by CUDA driver, we need to emit .pragma "nounroll" at the
header of that loop.

This patch also extracts getting unroll metadata from loop ID metadata
into a shared helper function.

[meheff]

LGTM

lib/Target/NVPTX/NVPTXAsmPrinter.cpp

426

FYI the loop unrolling pass should replace instances of "llvm.loop.unroll.count 1" from "#pragma unroll 1" with llvm.loop.unroll.disable.

lib/Transforms/Scalar/LoopUnrollPass.cpp

237

How about calling this function GetUnrollMetadataForLoop to avoid colliding with the new function you added?

[jingyue]

Rename the original GetUnrollMetadata to GetUnrollMetadataForLoop

[jingyue]

lib/Target/NVPTX/NVPTXAsmPrinter.cpp

426

Ack'ed.

lib/Transforms/Scalar/LoopUnrollPass.cpp

237

Done.

[jholewinski]

Looks good to me! Sorry for the delay, missed the original email notification.

include/llvm/CodeGen/AsmPrinter.h

public:
   /// Emit an alignment directive to the specified power of two boundary. For
   /// example, if you pass in 3 here, you will get an 8 byte alignment. If a
   /// global value is specified, and if that global has an explicit alignment
   /// requested, it will override the alignment request if required for
   /// correctness.
   ///
   void EmitAlignment(unsigned NumBits, const GlobalObject *GO = nullptr) const;
 
-  /// This method prints the label for the specified MachineBasicBlock, an
-  /// alignment (if present) and a comment describing it if appropriate.
-  void EmitBasicBlockStart(const MachineBasicBlock &MBB) const;
-
   /// Lower the specified LLVM Constant to an MCExpr.
   const MCExpr *lowerConstant(const Constant *CV);
 
   /// \brief Print a general LLVM constant to the .s file.
   void EmitGlobalConstant(const Constant *CV);
 
   //===------------------------------------------------------------------===//
   // Overridable Hooks
public:
   /// Targets can override this to emit stuff before the first basic block in
   /// the function.
   virtual void EmitFunctionBodyStart() {}
 
   /// Targets can override this to emit stuff after the last basic block in the
   /// function.
   virtual void EmitFunctionBodyEnd() {}
 
+  /// Targets can override this to emit stuff at the start of a basic block.
+  /// By default, this method prints the label for the specified
+  /// MachineBasicBlock, an alignment (if present) and a comment describing it
+  /// if appropriate.
+  virtual void EmitBasicBlockStart(const MachineBasicBlock &MBB) const;
+
   /// Targets can override this to emit stuff at the end of a basic block.
   virtual void EmitBasicBlockEnd(const MachineBasicBlock &MBB) {}
 
   /// Targets should implement this to emit instructions.
   virtual void EmitInstruction(const MachineInstr *) {
     llvm_unreachable("EmitInstruction not implemented");
   }
 

include/llvm/Transforms/Utils/UnrollLoop.h

 // actual pass or policy, but provides a single function to perform loop
 // unrolling.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_TRANSFORMS_UTILS_UNROLLLOOP_H
 #define LLVM_TRANSFORMS_UTILS_UNROLLLOOP_H
 
+#include "llvm/ADT/StringRef.h"
+
 namespace llvm {
 
 class AssumptionCache;
 class Loop;
 class LoopInfo;
 class LPPassManager;
+class MDNode;
 class Pass;
 
 bool UnrollLoop(Loop *L, unsigned Count, unsigned TripCount, bool AllowRuntime,
                 unsigned TripMultiple, LoopInfo *LI, Pass *PP,
                 LPPassManager *LPM, AssumptionCache *AC);
 
 bool UnrollRuntimeLoopProlog(Loop *L, unsigned Count, LoopInfo *LI,
                              LPPassManager* LPM);
+
+const MDNode *GetUnrollMetadata(const MDNode *LoopID, StringRef Name);
 }
 
 #endif

lib/Target/NVPTX/NVPTXAsmPrinter.h

class LLVM_LIBRARY_VISIBILITY NVPTXAsmPrinter : public AsmPrinter {
   void emitSrcInText(StringRef filename, unsigned line);
 
 private:
   const char *getPassName() const override { return "NVPTX Assembly Printer"; }
 
   const Function *F;
   std::string CurrentFnName;
 
+  void EmitBasicBlockStart(const MachineBasicBlock &MBB) const override;
   void EmitFunctionEntryLabel() override;
   void EmitFunctionBodyStart() override;
   void EmitFunctionBodyEnd() override;
   void emitImplicitDef(const MachineInstr *MI) const override;
 
   void EmitInstruction(const MachineInstr *) override;
   void lowerToMCInst(const MachineInstr *MI, MCInst &OutMI);
   bool lowerOperand(const MachineOperand &MO, MCOperand &MCOp);
private:
 
   static const char *getRegisterName(unsigned RegNo);
   void emitDemotedVars(const Function *, raw_ostream &);
 
   bool lowerImageHandleOperand(const MachineInstr *MI, unsigned OpNo,
                                MCOperand &MCOp);
   void lowerImageHandleSymbol(unsigned Index, MCOperand &MCOp);
 
+  bool isLoopHeaderOfNoUnroll(const MachineBasicBlock &MBB) const;
+
   LineReader *reader;
   LineReader *getReader(std::string);
 
   // Used to control the need to emit .generic() in the initializer of
   // module scope variables.
   // Although ptx supports the hybrid mode like the following,
   //    .global .u32 a;
   //    .global .u32 b;
NVPTXAsmPrinter(TargetMachine &TM, MCStreamer &Streamer)
     EmitGeneric = (nvptxSubtarget.getDrvInterface() == NVPTX::CUDA);
   }
 
   ~NVPTXAsmPrinter() {
     if (!reader)
       delete reader;
   }
 
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<MachineLoopInfo>();
+    AsmPrinter::getAnalysisUsage(AU);
+  }
+
   bool ignoreLoc(const MachineInstr &);
 
   std::string getVirtualRegisterName(unsigned) const;
 
   DebugLoc prevDebugLoc;
   void emitLineNumberAsDotLoc(const MachineInstr &);
 };
 } // end of namespace
 
 #endif

lib/Target/NVPTX/NVPTXAsmPrinter.cpp

 #include "NVPTXRegisterInfo.h"
 #include "NVPTXTargetMachine.h"
 #include "NVPTXUtilities.h"
 #include "cl_common_defines.h"
 #include "llvm/ADT/StringExtras.h"
 #include "llvm/Analysis/ConstantFolding.h"
 #include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineLoopInfo.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/IR/DebugInfo.h"
 #include "llvm/IR/DerivedTypes.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/GlobalVariable.h"
 #include "llvm/IR/Mangler.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/Operator.h"
 #include "llvm/MC/MCStreamer.h"
 #include "llvm/MC/MCSymbol.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/FormattedStream.h"
 #include "llvm/Support/Path.h"
 #include "llvm/Support/TargetRegistry.h"
 #include "llvm/Support/TimeValue.h"
 #include "llvm/Target/TargetLoweringObjectFile.h"
+#include "llvm/Transforms/Utils/UnrollLoop.h"
 #include <sstream>
 using namespace llvm;
 
 #define DEPOTNAME "__local_depot"
 
 static cl::opt<bool>
 EmitLineNumbers("nvptx-emit-line-numbers", cl::Hidden,
                 cl::desc("NVPTX Specific: Emit Line numbers even without -G"),

 }
 
 void NVPTXAsmPrinter::printReturnValStr(const MachineFunction &MF,
                                         raw_ostream &O) {
   const Function *F = MF.getFunction();
   printReturnValStr(F, O);
 }
 
+// Return true if MBB is the header of a loop marked with
+// llvm.loop.unroll.disable.
+// TODO(jingyue): consider "#pragma unroll 1" which is equivalent to "#pragma
+// nounroll".
+bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
+    const MachineBasicBlock &MBB) const {
+  MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
+  // TODO(jingyue): isLoopHeader() should take "const MachineBasicBlock *".
+  // We insert .pragma "nounroll" only to the loop header.
+  if (!LI.isLoopHeader(const_cast<MachineBasicBlock *>(&MBB)))
+    return false;
+
+  // llvm.loop.unroll.disable is marked on the back edges of a loop. Therefore,
+  // we iterate through each back edge of the loop with header MBB, and check
+  // whether its metadata contains llvm.loop.unroll.disable.
+  for (auto I = MBB.pred_begin(); I != MBB.pred_end(); ++I) {
+    const MachineBasicBlock *PMBB = *I;
+    if (LI.getLoopFor(PMBB) != LI.getLoopFor(&MBB)) {
+      // Edges from other loops to MBB are not back edges.
+      continue;
+    }
+    if (const BasicBlock *PBB = PMBB->getBasicBlock()) {
+      if (const MDNode *LoopID =
+              PBB->getTerminator()->getMetadata("llvm.loop")) {
+        if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
+          return true;
+      }
+    }
+  }
+  return false;
+}
+
+void NVPTXAsmPrinter::EmitBasicBlockStart(const MachineBasicBlock &MBB) const {
+  AsmPrinter::EmitBasicBlockStart(MBB);
+  if (isLoopHeaderOfNoUnroll(MBB))
+    OutStreamer.EmitRawText(StringRef("\t.pragma \"nounroll\";\n"));
+}
+
 void NVPTXAsmPrinter::EmitFunctionEntryLabel() {
   SmallString<128> Str;
   raw_svector_ostream O(Str);
 
   if (!GlobalsEmitted) {
     emitGlobals(*MF->getFunction()->getParent());
     GlobalsEmitted = true;
   }

lib/Transforms/Scalar/LoopUnrollPass.cpp

static unsigned ApproximateLoopSize(const Loop *L, unsigned &NumCalls,
   LoopSize = std::max(LoopSize, 3u);
 
   return LoopSize;
 }
 
 // Returns the loop hint metadata node with the given name (for example,
 // "llvm.loop.unroll.count").  If no such metadata node exists, then nullptr is
 // returned.
 static const MDNode *GetUnrollMetadata(const Loop *L, StringRef Name) {
   MDNode *LoopID = L->getLoopID();
   if (!LoopID)
     return nullptr;
-
-  // First operand should refer to the loop id itself.
-  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
-  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
-
-  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
-    const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
-    if (!MD)
-      continue;
-
-    const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
-    if (!S)
-      continue;
-
-    if (Name.equals(S->getString()))
-      return MD;
-  }
-  return nullptr;
+  return GetUnrollMetadata(LoopID, Name);
 }
 
 // Returns true if the loop has an unroll(full) pragma.
 static bool HasUnrollFullPragma(const Loop *L) {
   return GetUnrollMetadata(L, "llvm.loop.unroll.full");
 }
 
 // Returns true if the loop has an unroll(disable) pragma.

lib/Transforms/Utils/LoopUnroll.cpp

if (OuterL) {
           OuterL = OuterL->getParentLoop();
 
       formLCSSARecursively(*OuterL, *DT, LI, SE);
     }
   }
 
   return true;
 }
+
+/// Given an llvm.loop loop id metadata node, returns the loop hint metadata
+/// node with the given name (for example, "llvm.loop.unroll.count"). If no
+/// such metadata node exists, then nullptr is returned.
+const MDNode *llvm::GetUnrollMetadata(const MDNode *LoopID, StringRef Name) {
+  // First operand should refer to the loop id itself.
+  assert(LoopID->getNumOperands() > 0 && "requires at least one operand");
+  assert(LoopID->getOperand(0) == LoopID && "invalid loop id");
+
+  for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
+    const MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
+    if (!MD)
+      continue;
+
+    const MDString *S = dyn_cast<MDString>(MD->getOperand(0));
+    if (!S)
+      continue;
+
+    if (Name.equals(S->getString()))
+      return MD;
+  }
+  return nullptr;
+}

test/CodeGen/NVPTX/nounroll.ll

+; RUN: llc < %s -march=nvptx64 -mcpu=sm_20 | FileCheck %s
+
+target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
+target triple = "nvptx64-unknown-unknown"
+
+; Compiled from the following CUDA code:
+;
+;   #pragma nounroll
+;   for (int i = 0; i < 2; ++i)
+;     output[i] = input[i];
+define void @nounroll(float* %input, float* %output) {
+; CHECK-LABEL: .visible .func nounroll(
+entry:
+  br label %for.body
+
+for.body:
+; CHECK: .pragma "nounroll"
+  %i.06 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %idxprom = sext i32 %i.06 to i64
+  %arrayidx = getelementptr inbounds float* %input, i64 %idxprom
+  %0 = load float* %arrayidx, align 4
+; CHECK: ld.f32
+  %arrayidx2 = getelementptr inbounds float* %output, i64 %idxprom
+  store float %0, float* %arrayidx2, align 4
+; CHECK: st.f32
+  %inc = add nuw nsw i32 %i.06, 1
+  %exitcond = icmp eq i32 %inc, 2
+  br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
+; CHECK-NOT: ld.f32
+; CHECK-NOT: st.f32
+
+for.end:
+  ret void
+}
+
+!0 = distinct !{!0, !1}
+!1 = !{!"llvm.loop.unroll.disable"}