[LoopRotate] Use llvm.experimental.noalias.scope.decl for duplicating noalias metadata as needed

[jeroen.dobbelaere]

Summary

Similar to D92887, LoopRotation also needs duplicate the noalias scopes when rotating a @llvm.experimental.noalias.scope.decl across a block boundary.
This is based on the version from the Full Restrict paches (D68511).

The problem it fixes also showed up in Transforms/Coroutines/ex5.ll after D93040 (when enabling strict checking with -verify-noalias-scope-decl-dom).

[nikic]

This is tricker than the unroll case, but also looks sound to me.

llvm/lib/Transforms/Utils/LoopRotationUtils.cpp

481

// if D is part of OrigHeader

482

// if D and U are part of OrigHeader

498

llvm:: probably not needed.

505

Is there any reason to insert a new one before NAD and then move NAD, rather then just inserting the new one at NewHeaderInsertionPoint?

llvm/test/Transforms/LoopRotate/noalias.ll

45

In line with the usual convention, this should be at the start of the @test_02 function.

[jeroen.dobbelaere]

• Update according to comments

[jeroen.dobbelaere]

• insert clone directly at the right location

[jeroen.dobbelaere]

llvm/lib/Transforms/Utils/LoopRotationUtils.cpp

505

hmm. Probably not.. let me change this.

[jeroen.dobbelaere]

llvm/lib/Transforms/Utils/LoopRotationUtils.cpp

415

I'll remove the unnecessary braces in a next version.

[nikic]

LGTM

llvm/lib/Transforms/Utils/LoopRotationUtils.cpp

473

nit: "local restrict declarations" -> "noalias scope declarations"

478

nit: "provenance.noalias" -> "!noalias" or something like that.

514

I think the 1 suffix in these names doesn't make much sense, as there's no 2 etc...

llvm/test/Transforms/LoopRotate/noalias.ll

174

You should be able to drop the TBAA metadata here. It's not used in the test and not relevant for it.

[jeroen.dobbelaere]

Adapted to comments

[jeroen.dobbelaere]

In D92887#2514156, @asbirlea wrote:

My concern with this patch is in the iteration on all instructions in the blocks in cloneAndAdaptNoAliasScopes and identifyNoAliasScopesToClone.
Perhaps for LoopUnroll is not an issue, if unroll cost model is taken into account before this (i.e. unroll doesn't happen for large BBs). But when doing loop rotate, this may explode for IRs with large BBs, so I'd say some measurements would be helpful here.
For loop rotate, which instructions are cloned from the header to the preheader is known at the callsite, so as to not iterate over the entire original preheader, but I'm not sure this is sufficient.

One problem I see is that it is possible that instructions with references to the scopes have been moved before a llvm.experimental.noalias.scope.decl. On the other hand, as far as I see, it should not be 'wrong' to only duplicate the scopes for the instructions that have been moved: at worst, analysis will fall back to 'MayAlias' when comparing a modified to a non-modified version.

I'll try to reduce the instruction range for the OrigPreHeader.

[jeroen.dobbelaere]

Only adapt instructions in the OrigPreHeader starting from the first new llvm.experimental.noalias.scope.decl. This avoids going over potential very large OrigPreHeader blocks, as mentioned by @asbirlea

[jeroen.dobbelaere]

In D94306#2517595, @jeroen.dobbelaere wrote:

Only adapt instructions in the OrigPreHeader starting from the first new llvm.experimental.noalias.scope.decl. This avoids going over potential very large OrigPreHeader blocks, as mentioned by @asbirlea

@nikic, could you take a look at the new changes ? Thanks !

[nikic]

Still LGTM.

[jeroen.dobbelaere]

Make use of NoAliasScopeDeclInst

llvm/include/llvm/Transforms/Utils/Cloning.h

void adaptNoAliasScopes(
     LLVMContext &Context);
 
 /// Clone the specified noalias decl scopes. Then adapt all instructions in the
 /// NewBlocks basicblocks to the cloned versions.
 /// 'Ext' will be added to the duplicate scope names.
 void cloneAndAdaptNoAliasScopes(ArrayRef<MetadataAsValue *> NoAliasDeclScopes,
                                 ArrayRef<BasicBlock *> NewBlocks,
                                 LLVMContext &Context, StringRef Ext);
+
+/// Clone the specified noalias decl scopes. Then adapt all instructions in the
+/// [IStart, IEnd] (IEnd included !) range to the cloned versions. 'Ext' will be
+/// added to the duplicate scope names.
+void cloneAndAdaptNoAliasScopes(ArrayRef<MetadataAsValue *> NoAliasDeclScopes,
+                                Instruction *IStart, Instruction *IEnd,
+                                LLVMContext &Context, StringRef Ext);
 } // end namespace llvm
 
 #endif // LLVM_TRANSFORMS_UTILS_CLONING_H

llvm/lib/Transforms/Utils/CloneFunction.cpp

void llvm::cloneAndAdaptNoAliasScopes(
   cloneNoAliasScopes(NoAliasDeclScopes, ClonedScopes, ClonedMVScopes, Ext,
                      Context);
   // Identify instructions using metadata that needs adaptation
   for (BasicBlock *NewBlock : NewBlocks)
     for (Instruction &I : *NewBlock)
       adaptNoAliasScopes(&I, ClonedScopes, ClonedMVScopes, Context);
 }
 
+void llvm::cloneAndAdaptNoAliasScopes(
+    ArrayRef<MetadataAsValue *> NoAliasDeclScopes, Instruction *IStart,
+    Instruction *IEnd, LLVMContext &Context, StringRef Ext) {
+  if (NoAliasDeclScopes.empty())
+    return;
+
+  DenseMap<MDNode *, MDNode *> ClonedScopes;
+  DenseMap<MetadataAsValue *, MetadataAsValue *> ClonedMVScopes;
+  LLVM_DEBUG(dbgs() << "cloneAndAdaptNoAliasScopes: cloning "
+                    << NoAliasDeclScopes.size() << " node(s)\n");
+
+  cloneNoAliasScopes(NoAliasDeclScopes, ClonedScopes, ClonedMVScopes, Ext,
+                     Context);
+  // Identify instructions using metadata that needs adaptation
+  assert(IStart->getParent() == IEnd->getParent() && "different basic block ?");
+  auto ItStart = IStart->getIterator();
+  auto ItEnd = IEnd->getIterator();
+  ++ItEnd; // IEnd is included, increment ItEnd to get the end of the range
+  for (auto &I : llvm::make_range(ItStart, ItEnd))
+    adaptNoAliasScopes(&I, ClonedScopes, ClonedMVScopes, Context);
+}
+
 void llvm::identifyNoAliasScopesToClone(
     ArrayRef<BasicBlock *> BBs,
     SmallVectorImpl<MetadataAsValue *> &NoAliasDeclScopes) {
   for (BasicBlock *BB : BBs)
     for (Instruction &I : *BB)
       if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
         NoAliasDeclScopes.push_back(cast<MetadataAsValue>(
             Decl->getOperand(Intrinsic::NoAliasScopeDeclScopeArg)));
 }

llvm/lib/Transforms/Utils/LoopRotationUtils.cpp

 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Module.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Cloning.h"
 #include "llvm/Transforms/Utils/Local.h"
 #include "llvm/Transforms/Utils/LoopUtils.h"
 #include "llvm/Transforms/Utils/SSAUpdater.h"
 #include "llvm/Transforms/Utils/ValueMapper.h"
 using namespace llvm;
 
 #define DEBUG_TYPE "loop-rotate"
 
do {
     for (auto I = std::next(OrigPreheader->rbegin()), E = OrigPreheader->rend();
          I != E; ++I) {
       if (auto *DII = dyn_cast<DbgVariableIntrinsic>(&*I))
         DbgIntrinsics.insert(makeHash(DII));
       else
         break;
     }
 
+    // Remember the local noalias scope declarations in the header. After the
+    // rotation, they must be duplicated and the scope must be cloned. This
+    // avoids unwanted interaction across iterations.
+    SmallVector<Instruction *, 6> NoAliasDeclInstructions;
+    for (Instruction &I : *OrigHeader)
+      if (auto *Decl = dyn_cast<NoAliasScopeDeclInst>(&I))
+        NoAliasDeclInstructions.push_back(Decl);
+
     while (I != E) {
       Instruction *Inst = &*I++;
 
       // If the instruction's operands are invariant and it doesn't read or write
       // memory, then it is safe to hoist.  Doing this doesn't change the order of
       // execution in the preheader, but does prevent the instruction from
       // executing in each iteration of the loop.  This means it is safe to hoist
       // something that might trap, but isn't safe to hoist something that reads
       // memory (without proving that the loop doesn't write).
       if (L->hasLoopInvariantOperands(Inst) && !Inst->mayReadFromMemory() &&
           !Inst->mayWriteToMemory() && !Inst->isTerminator() &&
           !isa<DbgInfoIntrinsic>(Inst) && !isa<AllocaInst>(Inst)) {
while (I != E) {
             AC->registerAssumption(II);
         // MemorySSA cares whether the cloned instruction was inserted or not, and
         // not whether it can be remapped to a simplified value.
         if (MSSAU)
           InsertNewValueIntoMap(ValueMapMSSA, Inst, C);
       }
     }
 
+    if (!NoAliasDeclInstructions.empty()) {
+      // There are noalias scope declarations:
+      // (general):
+      // Original:    OrigPre              { OrigHeader NewHeader ... Latch }
+      // after:      (OrigPre+OrigHeader') { NewHeader ... Latch OrigHeader }
+      //
+      // with D: llvm.experimental.noalias.scope.decl,
+      //      U: !noalias or !alias.scope depending on D
+      //       ... { D U1 U2 }   can transform into:
+      // (0) : ... { D U1 U2 }        // no relevant rotation for this part
+      // (1) : ... D' { U1 U2 D }     // D is part of OrigHeader
+      // (2) : ... D' U1' { U2 D U1 } // D, U1 are part of OrigHeader
+      //
+      // We now want to transform:
+      // (1) -> : ... D' { D U1 U2 D'' }
+      // (2) -> : ... D' U1' { D U2 D'' U1'' }
+      // D: original llvm.experimental.noalias.scope.decl
+      // D', U1': duplicate with replaced scopes
+      // D'', U1'': different duplicate with replaced scopes
+      // This ensures a safe fallback to 'may_alias' introduced by the rotate,
+      // as U1'' and U1' scopes will not be compatible wrt to the local restrict
+
+      // Clone the llvm.experimental.noalias.decl again for the NewHeader.
+      Instruction *NewHeaderInsertionPoint = &(*NewHeader->getFirstNonPHI());
+      for (Instruction *NAD : NoAliasDeclInstructions) {
+        LLVM_DEBUG(dbgs() << "  Cloning llvm.experimental.noalias.scope.decl:"
+                          << *NAD << "\n");
+        Instruction *NewNAD = NAD->clone();
+        NewNAD->insertBefore(NewHeaderInsertionPoint);
+      }
+
+      // Scopes must now be duplicated, once for OrigHeader and once for
+      // OrigPreHeader'.
+      {
+        auto &Context = NewHeader->getContext();
+
+        SmallVector<MetadataAsValue *, 8> NoAliasDeclScopes;
+        for (Instruction *NAD : NoAliasDeclInstructions)
+          NoAliasDeclScopes.push_back(cast<MetadataAsValue>(
+              NAD->getOperand(Intrinsic::NoAliasScopeDeclScopeArg)));
+
+        LLVM_DEBUG(dbgs() << "  Updating OrigHeader scopes\n");
+        cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, {OrigHeader}, Context,
+                                   "h.rot");
+        LLVM_DEBUG(OrigHeader->dump());
+
+        // Keep the compile time impact low by only adapting the inserted block
+        // of instructions in the OrigPreHeader. This might result in slightly
+        // more aliasing between these instructions and those that were already
+        // present, but it will be much faster when the original PreHeader is
+        // large.
+        LLVM_DEBUG(dbgs() << "  Updating part of OrigPreheader scopes\n");
+        auto *FirstDecl =
+            cast<Instruction>(ValueMap[*NoAliasDeclInstructions.begin()]);
+        auto *LastInst = &OrigPreheader->back();
+        cloneAndAdaptNoAliasScopes(NoAliasDeclScopes, FirstDecl, LastInst,
+                                   Context, "pre.rot");
+        LLVM_DEBUG(OrigPreheader->dump());
+
+        LLVM_DEBUG(dbgs() << "  Updated NewHeader:\n");
+        LLVM_DEBUG(NewHeader->dump());
+      }
+    }
+
     // Along with all the other instructions, we just cloned OrigHeader's
     // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's
     // successors by duplicating their incoming values for OrigHeader.
     for (BasicBlock *SuccBB : successors(OrigHeader))
       for (BasicBlock::iterator BI = SuccBB->begin();
            PHINode *PN = dyn_cast<PHINode>(BI); ++BI)
         PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader);
 

llvm/test/Transforms/LoopRotate/noalias.ll

+; RUN: opt -S -loop-rotate < %s | FileCheck %s
+; RUN: opt -S -loop-rotate -enable-mssa-loop-dependency=true -verify-memoryssa < %s | FileCheck %s
+; RUN: opt -S -passes='require<targetir>,require<assumptions>,loop(loop-rotate)' < %s | FileCheck %s
+; RUN: opt -S -passes='require<targetir>,require<assumptions>,loop(loop-rotate)' -enable-mssa-loop-dependency=true -verify-memoryssa  < %s | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+declare void @g(i32*)
+
+define void @test_02(i32* nocapture %_pA) nounwind ssp {
+; CHECK-LABEL: @test_02(
+; CHECK: entry:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !2)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !2
+; CHECK: for.body:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !5)
+; CHECK:   store i32 0, i32* %arrayidx, align 16, !noalias !5
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !7)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !7
+; CHECK: for.end:
+
+entry:
+  %array = alloca [20 x i32], align 16
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  tail call void @llvm.experimental.noalias.scope.decl(metadata !2)
+  store i32 42, i32* %_pA, align 16, !alias.scope !2
+  %cmp = icmp slt i32 %i.0, 100
+  %arrayidx = getelementptr inbounds [20 x i32], [20 x i32]* %array, i64 0, i64 0
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  store i32 0, i32* %arrayidx, align 16, !noalias !2
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  %arrayidx.lcssa = phi i32* [ %arrayidx, %for.cond ]
+  call void @g(i32* %arrayidx.lcssa) nounwind
+  ret void
+}
+
+define void @test_03(i32* nocapture %_pA) nounwind ssp {
+; CHECK-LABEL: @test_03(
+; CHECK: entry:
+; CHECK: for.body:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !5)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !5
+; CHECK:   store i32 0, i32* %arrayidx, align 16, !noalias !5
+; CHECK: for.end:
+
+entry:
+  %array = alloca [20 x i32], align 16
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  %cmp = icmp slt i32 %i.0, 100
+  %arrayidx = getelementptr inbounds [20 x i32], [20 x i32]* %array, i64 0, i64 0
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  tail call void @llvm.experimental.noalias.scope.decl(metadata !2)
+  store i32 42, i32* %_pA, align 16, !alias.scope !2
+  store i32 0, i32* %arrayidx, align 16, !noalias !2
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  %arrayidx.lcssa = phi i32* [ %arrayidx, %for.cond ]
+  call void @g(i32* %arrayidx.lcssa) nounwind
+  ret void
+}
+
+define void @test_04(i32* nocapture %_pA) nounwind ssp {
+; CHECK-LABEL: @test_04(
+; CHECK: entry:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !9)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !9
+; CHECK: for.body:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !5)
+; CHECK:   store i32 0, i32* %arrayidx, align 16, !noalias !5
+; CHECK:   store i32 43, i32* %_pA, align 16, !alias.scope !5
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !11)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !11
+; CHECK: for.end:
+entry:
+  %array = alloca [20 x i32], align 16
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  tail call void @llvm.experimental.noalias.scope.decl(metadata !2)
+  store i32 42, i32* %_pA, align 16, !alias.scope !2
+  %cmp = icmp slt i32 %i.0, 100
+  %arrayidx = getelementptr inbounds [20 x i32], [20 x i32]* %array, i64 0, i64 0
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  store i32 0, i32* %arrayidx, align 16, !noalias !2
+  store i32 43, i32* %_pA, align 16, !alias.scope !2
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  %arrayidx.lcssa = phi i32* [ %arrayidx, %for.cond ]
+  call void @g(i32* %arrayidx.lcssa) nounwind
+  ret void
+}
+
+define void @test_05(i32* nocapture %_pA) nounwind ssp {
+; CHECK-LABEL: @test_05(
+; CHECK: entry:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !13)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !13
+; CHECK: for.body:
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !5)
+; CHECK:   store i32 0, i32* %arrayidx, align 16, !noalias !5
+; CHECK:   store i32 43, i32* %_pA, align 16, !alias.scope !5
+; CHECK:   tail call void @llvm.experimental.noalias.scope.decl(metadata !15)
+; CHECK:   store i32 42, i32* %_pA, align 16, !alias.scope !15
+; CHECK: for.end:
+; CHECK:   store i32 44, i32* %_pA, align 16, !alias.scope !5
+
+entry:
+  %array = alloca [20 x i32], align 16
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.body, %entry
+  %i.0 = phi i32 [ 0, %entry ], [ %inc, %for.body ]
+  tail call void @llvm.experimental.noalias.scope.decl(metadata !2)
+  store i32 42, i32* %_pA, align 16, !alias.scope !2
+  %cmp = icmp slt i32 %i.0, 100
+  %arrayidx = getelementptr inbounds [20 x i32], [20 x i32]* %array, i64 0, i64 0
+  br i1 %cmp, label %for.body, label %for.end
+
+for.body:                                         ; preds = %for.cond
+  store i32 0, i32* %arrayidx, align 16, !noalias !2
+  store i32 43, i32* %_pA, align 16, !alias.scope !2
+  %inc = add nsw i32 %i.0, 1
+  br label %for.cond
+
+for.end:                                          ; preds = %for.cond
+  %arrayidx.lcssa = phi i32* [ %arrayidx, %for.cond ]
+  store i32 44, i32* %_pA, align 16, !alias.scope !2
+  call void @g(i32* %arrayidx.lcssa) nounwind
+  ret void
+}
+
+; Function Attrs: inaccessiblememonly nounwind
+declare void @llvm.experimental.noalias.scope.decl(metadata) #1
+
+attributes #0 = { 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" "unsafe-fp-math"="false" "use-soft-float"="false" }
+attributes #1 = { inaccessiblememonly nounwind }
+attributes #2 = { nounwind readnone speculatable }
+
+!llvm.module.flags = !{!0}
+!llvm.ident = !{!1}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{!"clang"}
+!2 = !{!3}
+!3 = distinct !{!3, !4, !"test_loop_rotate_XX: pA"}
+!4 = distinct !{!4, !"test_loop_rotate_XX"}
+
+; CHECK: !0 = !{i32 1, !"wchar_size", i32 4}
+; CHECK: !1 = !{!"clang"}
+; CHECK: !2 = !{!3}
+; CHECK: !3 = distinct !{!3, !4, !"test_loop_rotate_XX: pA:pre.rot"}
+; CHECK: !4 = distinct !{!4, !"test_loop_rotate_XX"}
+; CHECK: !5 = !{!6}
+; CHECK: !6 = distinct !{!6, !4, !"test_loop_rotate_XX: pA"}
+; CHECK: !7 = !{!8}
+; CHECK: !8 = distinct !{!8, !4, !"test_loop_rotate_XX: pA:h.rot"}
+; CHECK: !9 = !{!10}
+; CHECK: !10 = distinct !{!10, !4, !"test_loop_rotate_XX: pA:pre.rot"}
+; CHECK: !11 = !{!12}
+; CHECK: !12 = distinct !{!12, !4, !"test_loop_rotate_XX: pA:h.rot"}
+; CHECK: !13 = !{!14}
+; CHECK: !14 = distinct !{!14, !4, !"test_loop_rotate_XX: pA:pre.rot"}
+; CHECK: !15 = !{!16}
+; CHECK: !16 = distinct !{!16, !4, !"test_loop_rotate_XX: pA:h.rot"}