[NVPTX] Emit pragma nounroll for llvm.loop.unroll.count=1

[slydiman]

Summary

Emit pragma nounroll for llvm.loop.unroll.count=1 (#pragma unroll 1).

[tra]

LGTM with a nit.

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

432–436

This all can be folded directly into
if (mdconst::extract<ConstantInt>(UnrollCountMD->getOperand(1))->getZExtValue() == 1) return true;

llvm/lib/Target/NVPTX/NVPTXAsmPrinter.cpp

 
 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: consider "#pragma unroll 1" which is equivalent to "#pragma nounroll".
+// llvm.loop.unroll.disable or llvm.loop.unroll.count=1.
 bool NVPTXAsmPrinter::isLoopHeaderOfNoUnroll(
     const MachineBasicBlock &MBB) const {
   MachineLoopInfo &LI = getAnalysis<MachineLoopInfo>();
   // We insert .pragma "nounroll" only to the loop header.
   if (!LI.isLoopHeader(&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 (const MachineBasicBlock *PMBB : MBB.predecessors()) {
     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 (MDNode *LoopID =
               PBB->getTerminator()->getMetadata(LLVMContext::MD_loop)) {
         if (GetUnrollMetadata(LoopID, "llvm.loop.unroll.disable"))
           return true;
+        if (MDNode *UnrollCountMD =
+                GetUnrollMetadata(LoopID, "llvm.loop.unroll.count")) {
+          if (mdconst::extract<ConstantInt>(UnrollCountMD->getOperand(1))
+                  ->getZExtValue() == 1)
+            return true;
+        }
       }
     }
   }
   return false;
 }
 
 void NVPTXAsmPrinter::emitBasicBlockStart(const MachineBasicBlock &MBB) {
   AsmPrinter::emitBasicBlockStart(MBB);
   if (isLoopHeaderOfNoUnroll(MBB))

llvm/test/CodeGen/NVPTX/nounroll.ll

; CHECK: st.f32
   br i1 %exitcond, label %for.end, label %for.body, !llvm.loop !0
 ; CHECK-NOT: ld.f32
 ; CHECK-NOT: st.f32
 
 for.end:
   ret void
 }
 
+; Compiled from the following CUDA code:
+;
+;   #pragma unroll 1
+;   for (int i = 0; i < 2; ++i)
+;     output[i] = input[i];
+define void @unroll1(float* %input, float* %output) {
+; CHECK-LABEL: .visible .func unroll1(
+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, float* %input, i64 %idxprom
+  %0 = load float, float* %arrayidx, align 4
+; CHECK: ld.f32
+  %arrayidx2 = getelementptr inbounds float, 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 !2
+; CHECK-NOT: ld.f32
+; CHECK-NOT: st.f32
+
+for.end:
+  ret void
+}
+
 !0 = distinct !{!0, !1}
 !1 = !{!"llvm.loop.unroll.disable"}
+!2 = distinct !{!2, !3}
+!3 = !{!"llvm.loop.unroll.count", i32 1}