[IRCE] Use SCEVExpander to modify loop bound

[dantrushin]

Summary

IRCE pass checks that it can calculate loop bounds by checking
SCEV availability at loop entry. However it is possible that loop
bound SCEV is loop invariant, but instruction used to compute it
resides within loop. In such case adjusting loop bound in preheader
using IRBuilder leads to malformed SSA.
Use SCEVExpander instead to generate proper instructions.

[merge_guards_bot]

Unit tests: pass. 62225 tests passed, 0 failed and 816 were skipped.

clang-tidy: pass.

clang-format: fail. Please format your changes with clang-format by running git-clang-format HEAD^ or applying this patch.

Build artifacts: diff.json, clang-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

Pre-merge checks is in beta. Report issue. Please join beta or enable it for your project.

[dantrushin]

Fix clang-format error

[merge_guards_bot]

Unit tests: pass. 62225 tests passed, 0 failed and 816 were skipped.

clang-tidy: pass.

clang-format: pass.

Build artifacts: diff.json, clang-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

Pre-merge checks is in beta. Report issue. Please join beta or enable it for your project.

[mkazantsev]

I am a bit worried about that. As far as I remember RightValue needs to be in preheader because it is used in checks that decide whether or not we should go to pre/postloops. Does your case generate them? If yes, please add checks to make it explicit. If not, please construct a similar test with pre/postloop and make sure it passes with the fix. If it doesn't, it means that IRCE is illegal in this situation, or that we need some advanced hoisting for these conditions.

llvm/test/Transforms/IRCE/non-loop-invariant-rhs-instr.ll

5

Please add CHECK-LABEL: test so that we could add more here later (maybe makes sense to rename it to test_01).

[mkazantsev]

llvm/test/Transforms/IRCE/non-loop-invariant-rhs-instr.ll

2

Please add run command for new pass manager.

[dantrushin]

Make sure RightValue is always properly instantiated in preheader;
Updated test;

[dantrushin]

In D73496#1848948, @mkazantsev wrote:

I am a bit worried about that. As far as I remember RightValue needs to be in preheader because it is used in checks that decide whether or not we should go to pre/postloops. Does your case generate them? If yes, please add checks to make it explicit. If not, please construct a similar test with pre/postloop and make sure it passes with the fix. If it doesn't, it means that IRCE is illegal in this situation, or that we need some advanced hoisting for these conditions.

This is exactly a purpose of this patch - ensure that RightValue is instantiated in preheader.
I added one more check to handle a case when we do not need to increment/decrement it.

[merge_guards_bot]

Unit tests: pass. 62331 tests passed, 0 failed and 838 were skipped.

clang-tidy: pass.

clang-format: fail. Please format your changes with clang-format by running git-clang-format HEAD^ or applying this patch.

Build artifacts: diff.json, clang-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

Pre-merge checks is in beta. Report issue. Please join beta or enable it for your project.

[mkazantsev]

LGTM (pls make sure that clang-format is fixed).

[dantrushin]

In D73496#1854173, @mkazantsev wrote:

LGTM (pls make sure that clang-format is fixed).

Thanks. Could you approve it then? (I fixed clang-format locally already, just did not upload such a trivial change)

[mkazantsev]

Done.

llvm/lib/Transforms/Scalar/InductiveRangeCheckElimination.cpp

LoopStructure::parseLoopStructure(ScalarEvolution &SE,
   assert(!StepCI->isZero() && "Zero step?");
   bool IsIncreasing = !StepCI->isNegative();
   bool IsSignedPredicate;
   const SCEV *StartNext = IndVarBase->getStart();
   const SCEV *Addend = SE.getNegativeSCEV(IndVarBase->getStepRecurrence(SE));
   const SCEV *IndVarStart = SE.getAddExpr(StartNext, Addend);
   const SCEV *Step = SE.getSCEV(StepCI);
 
-  ConstantInt *One = ConstantInt::get(IndVarTy, 1);
+  const SCEV *FixedRightSCEV = nullptr;
   if (IsIncreasing) {
     bool DecreasedRightValueByOne = false;
     if (StepCI->isOne()) {
       // Try to turn eq/ne predicates to those we can work with.
       if (Pred == ICmpInst::ICMP_NE && LatchBrExitIdx == 1)
         // while (++i != len) {         while (++i < len) {
         //   ...                 --->     ...
         // }                            }
if (IsIncreasing) {
     if (!isSafeIncreasingBound(IndVarStart, RightSCEV, Step, Pred,
                                LatchBrExitIdx, &L, SE)) {
       FailureReason = "Unsafe loop bounds";
       return None;
     }
     if (LatchBrExitIdx == 0) {
       // We need to increase the right value unless we have already decreased
       // it virtually when we replaced EQ with SGT.
-      if (!DecreasedRightValueByOne) {
-        IRBuilder<> B(Preheader->getTerminator());
-        RightValue = B.CreateAdd(RightValue, One);
-      }
+      if (!DecreasedRightValueByOne)
+        FixedRightSCEV =
+          SE.getAddExpr(RightSCEV, SE.getOne(RightSCEV->getType()));
     } else {
       assert(!DecreasedRightValueByOne &&
              "Right value can be decreased only for LatchBrExitIdx == 0!");
     }
   } else {
     bool IncreasedRightValueByOne = false;
     if (StepCI->isMinusOne()) {
       // Try to turn eq/ne predicates to those we can work with.
if (!isSafeDecreasingBound(IndVarStart, RightSCEV, Step, Pred,
                                LatchBrExitIdx, &L, SE)) {
       FailureReason = "Unsafe bounds";
       return None;
     }
 
     if (LatchBrExitIdx == 0) {
       // We need to decrease the right value unless we have already increased
       // it virtually when we replaced EQ with SLT.
-      if (!IncreasedRightValueByOne) {
-        IRBuilder<> B(Preheader->getTerminator());
-        RightValue = B.CreateSub(RightValue, One);
-      }
+      if (!IncreasedRightValueByOne)
+        FixedRightSCEV =
+            SE.getMinusSCEV(RightSCEV, SE.getOne(RightSCEV->getType()));
     } else {
       assert(!IncreasedRightValueByOne &&
              "Right value can be increased only for LatchBrExitIdx == 0!");
     }
   }
   BasicBlock *LatchExit = LatchBr->getSuccessor(LatchBrExitIdx);
 
   assert(SE.getLoopDisposition(LatchCount, &L) ==
              ScalarEvolution::LoopInvariant &&
          "loop variant exit count doesn't make sense!");
 
   assert(!L.contains(LatchExit) && "expected an exit block!");
   const DataLayout &DL = Preheader->getModule()->getDataLayout();
-  Value *IndVarStartV =
-      SCEVExpander(SE, DL, "irce")
-          .expandCodeFor(IndVarStart, IndVarTy, Preheader->getTerminator());
+  SCEVExpander Expander(SE, DL, "irce");
+  Instruction *Ins = Preheader->getTerminator();
+
+  if (FixedRightSCEV)
+    RightValue =
+        Expander.expandCodeFor(FixedRightSCEV, FixedRightSCEV->getType(), Ins);
+
+  Value *IndVarStartV = Expander.expandCodeFor(IndVarStart, IndVarTy, Ins);
   IndVarStartV->setName("indvar.start");
 
   LoopStructure Result;
 
   Result.Tag = "main";
   Result.Header = Header;
   Result.Latch = Latch;
   Result.LatchBr = LatchBr;

llvm/test/Transforms/IRCE/non-loop-invariant-rhs-instr.ll

+; RUN: opt -irce-print-changed-loops -S -verify-loop-info -irce -verify < %s 2>&1 | FileCheck %s
+
+; this test must compile successfully
+; CHECK: constrained Loop
+define i32 @test(i32 %A, i64 %Len, i32 *%array) {
+
+preheader:
+  %tripcheck = icmp sgt i64 %Len, 2
+  br i1 %tripcheck, label %loop, label %zero
+
+loop:
+  %indvar = phi i64 [ 1, %preheader ], [ %indvar.next, %latch ]
+  %0 = icmp ult i64 %indvar, %Len
+  br i1 %0, label %guarded, label %deopt
+
+guarded:
+  %addr = getelementptr inbounds i32, i32* %array, i64 %indvar
+  %res = load i32, i32* %addr
+  %cmp = icmp eq i32 %res, 0
+  br i1 %cmp, label %zero, label %latch
+
+latch:
+  %indvar.next = add nuw nsw i64 %indvar, 2
+  %res2 = mul i32 %res, 3
+; NOTE: this is loop invariant value, but not loop invariant instruction!
+  %1 = zext i32 %A to i64
+  %cmp2 = icmp ugt i64 %indvar.next, %1
+  br i1 %cmp2, label %loopexit, label %loop
+
+loopexit:
+  ret i32 %res2
+
+zero:
+  ret i32 0
+
+deopt:
+  ret i32 1
+
+}