[Pipeliner] Fix the bug in pragma that disables the pipeliner

[avl]

llvm/lib/CodeGen/MachinePipeliner.cpp

317

it looks like more appropriate place to set to default state would be in start of setPragmaPipelineOptions:

void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) {
  disabledByPragma = false;

Additionally, II_setByPragma could also be set to default state here. though this is for separate patch.

[avl]

llvm/lib/CodeGen/MachinePipeliner.cpp

295

now it becomes depending on order of metadata. It works for "distinct !{!8, !7, !9}" and does not work for distinct !{!8, !9, !7} . This default value should be set in the first line of setPragmaPipelineOptions (before other instructions):

void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) {
  disabledByPragma = false;

[sgundapa]

llvm/lib/CodeGen/MachinePipeliner.cpp

295

I did not realize this is inside the loop. I will move this before we start executing this loop.

[avl]

this lgtm for me, with nit in comments. Also, please wait some time for others opinions.

llvm/lib/CodeGen/MachinePipeliner.cpp

293

please delete this extra line.

[avl]

Already integrated by this commit - https://reviews.llvm.org/rGa04ab2ec08014d36337d686d43ecef7668622f1b

llvm/lib/CodeGen/MachinePipeliner.cpp

for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
     if (MD == nullptr)
       continue;
 
     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
 
     if (S == nullptr)
       continue;
 
+    // Reset the pragma for the next loop in iteration.
+    disabledByPragma = false;
+
     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.
 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;
   LI.FBB = nullptr;
   LI.BrCond.clear();
   if (TII->analyzeBranch(*L.getHeader(), LI.TBB, LI.FBB, LI.BrCond)) {
     LLVM_DEBUG(
         dbgs() << "Unable to analyzeBranch, can NOT pipeline current Loop\n");

llvm/test/CodeGen/Hexagon/swp-pragma-disable-bug.ll

+; RUN: llc -O2 -march=hexagon -enable-pipeliner  \
+; RUN: -debug-only=pipeliner < %s 2>&1 > /dev/null | FileCheck %s
+; REQUIRES: asserts
+;
+; Test that the pragma check that disables pipeliner does not disable pipelining
+; on both the loops.
+; CHECK: Can not pipeline loop
+; CHECK-NOT: Can not pipeline loop
+
+; Function Attrs: nofree norecurse nounwind
+define dso_local i32 @foo(i32* nocapture %a, i32* nocapture readonly %b, i32* nocapture %c) local_unnamed_addr #0 {
+entry:
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %i.023 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %b, i32 %i.023
+  %0 = load i32, i32* %arrayidx, align 4, !tbaa !2
+  %arrayidx1 = getelementptr inbounds i32, i32* %a, i32 %i.023
+  %1 = load i32, i32* %arrayidx1, align 4, !tbaa !2
+  %add = add nsw i32 %1, %0
+  store i32 %add, i32* %arrayidx1, align 4, !tbaa !2
+  %inc = add nuw nsw i32 %i.023, 1
+  %exitcond24 = icmp eq i32 %inc, 10
+  br i1 %exitcond24, label %for.body6, label %for.body, !llvm.loop !6
+
+for.cond.cleanup5:                                ; preds = %for.body6
+  ret i32 0
+
+for.body6:                                        ; preds = %for.body, %for.body6
+  %i2.022 = phi i32 [ %inc11, %for.body6 ], [ 0, %for.body ]
+  %arrayidx7 = getelementptr inbounds i32, i32* %a, i32 %i2.022
+  %2 = load i32, i32* %arrayidx7, align 4, !tbaa !2
+  %arrayidx8 = getelementptr inbounds i32, i32* %c, i32 %i2.022
+  %3 = load i32, i32* %arrayidx8, align 4, !tbaa !2
+  %add9 = add nsw i32 %3, %2
+  store i32 %add9, i32* %arrayidx8, align 4, !tbaa !2
+  %inc11 = add nuw nsw i32 %i2.022, 1
+  %exitcond = icmp eq i32 %inc11, 10
+  br i1 %exitcond, label %for.cond.cleanup5, label %for.body6, !llvm.loop !8
+}
+
+attributes #0 = { nofree norecurse nounwind "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="all" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="hexagonv60" "target-features"="+v60,-long-calls" "unsafe-fp-math"="false" "use-soft-float"="false" }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang version 11.0.0 (https://github.com/llvm/llvm-project.git 6e29846b29d2bcaa8a7a3d869f24f242bd93d272)"}
+!2 = !{!3, !3, i64 0}
+!3 = !{!"int", !4, i64 0}
+!4 = !{!"omnipotent char", !5, i64 0}
+!5 = !{!"Simple C/C++ TBAA"}
+!6 = distinct !{!6, !7}
+!7 = !{!"llvm.loop.unroll.disable"}
+!8 = distinct !{!8, !7, !9}
+!9 = !{!"llvm.loop.pipeline.disable", i1 true}