[LSR] Don't allow zero quotient as scale ref. PR56160

[mkazantsev]

Summary

Scale reg should never be zero, so when the quotient is zero, we
cannot assign it there. Limit this transform to avoid this situation.

[eopXD]

Looks good to me. I personally prefer a more comprehensive test case filename.
Thank you for the patch!

[mkazantsev]

Well, using bug IDs in test names is a common practice, and the idea is to refer the bug that this patch fixes.

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

for (int64_t Factor : Factors) {
     for (size_t i = 0, e = Base.BaseRegs.size(); i != e; ++i) {
       const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Base.BaseRegs[i]);
       if (AR && (AR->getLoop() == L || LU.AllFixupsOutsideLoop)) {
         const SCEV *FactorS = SE.getConstant(IntTy, Factor);
         if (FactorS->isZero())
           continue;
         // Divide out the factor, ignoring high bits, since we'll be
         // scaling the value back up in the end.
-        if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true)) {
+        if (const SCEV *Quotient = getExactSDiv(AR, FactorS, SE, true))
+          if (!Quotient->isZero()) {
             // TODO: This could be optimized to avoid all the copying.
             Formula F = Base;
             F.ScaledReg = Quotient;
             F.deleteBaseReg(F.BaseRegs[i]);
             // The canonical representation of 1*reg is reg, which is already in
             // Base. In that case, do not try to insert the formula, it will be
             // rejected anyway.
             if (F.Scale == 1 && (F.BaseRegs.empty() ||
                                  (AR->getLoop() != L && LU.AllFixupsOutsideLoop)))
               continue;
             // If AllFixupsOutsideLoop is true and F.Scale is 1, we may generate
             // non canonical Formula with ScaledReg's loop not being L.
             if (F.Scale == 1 && LU.AllFixupsOutsideLoop)
               F.canonicalize(*L);
             (void)InsertFormula(LU, LUIdx, F);
           }
       }
     }
   }
 }
 
 /// Generate reuse formulae from different IV types.
 void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
   // Don't bother truncating symbolic values.

llvm/test/Transforms/LoopStrengthReduce/pr56160.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -loop-reduce %s | FileCheck %s
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128-ni:1-p2:32:8:8:32-ni:2"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Make sure we don't crash.
+define void @test() {
+; CHECK-LABEL: @test(
+; CHECK-NEXT:  bb:
+; CHECK-NEXT:    [[TMP:%.*]] = load atomic i64, i64 addrspace(1)* undef unordered, align 8
+; CHECK-NEXT:    br label [[BB5:%.*]]
+; CHECK:       bb2:
+; CHECK-NEXT:    [[TMP0:%.*]] = trunc i64 [[TMP]] to i32
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[TMP0]], undef
+; CHECK-NEXT:    [[TMP2:%.*]] = add i32 [[TMP1]], 2
+; CHECK-NEXT:    [[TMP3:%.*]] = sub i32 [[TMP2]], [[TMP3:%.*]]
+; CHECK-NEXT:    [[TMP4:%.*]] = add i64 [[TMP]], undef
+; CHECK-NEXT:    [[TMP5:%.*]] = add i64 [[TMP4]], -4294967294
+; CHECK-NEXT:    [[TMP6:%.*]] = trunc i64 [[TMP5]] to i32
+; CHECK-NEXT:    [[TMP7:%.*]] = sub i32 [[TMP3]], [[TMP6]]
+; CHECK-NEXT:    [[TMP8:%.*]] = sub i32 0, [[TMP7]]
+; CHECK-NEXT:    [[TMP3]] = add i32 undef, [[TMP8]]
+; CHECK-NEXT:    unreachable
+; CHECK:       bb5:
+; CHECK-NEXT:    br label [[BB5]]
+;
+bb:
+  %tmp = load atomic i64, i64 addrspace(1)* undef unordered, align 8
+  %tmp1 = sub i64 4294967294, undef
+  br label %bb5
+
+bb2:                                              ; No predecessors!
+  %tmp3 = add i32 undef, %tmp24
+  unreachable
+
+bb5:                                              ; preds = %bb5, %bb
+  %tmp6 = phi i64 [ %tmp18, %bb5 ], [ %tmp, %bb ]
+  %tmp7 = phi i32 [ %tmp19, %bb5 ], [ undef, %bb ]
+  %tmp8 = phi i32 [ %tmp24, %bb5 ], [ undef, %bb ]
+  %tmp9 = sub i32 %tmp8, undef
+  %tmp10 = zext i32 %tmp9 to i64
+  %tmp11 = add i32 %tmp7, 1
+  %tmp12 = zext i32 %tmp11 to i64
+  %tmp13 = add i64 %tmp1, %tmp12
+  %tmp14 = add i64 %tmp6, %tmp10
+  %tmp15 = sub i64 %tmp14, %tmp13
+  %tmp16 = trunc i64 %tmp15 to i32
+  %tmp17 = add i32 undef, %tmp16
+  %tmp18 = add i64 %tmp6, 2
+  %tmp19 = add i32 %tmp7, 2
+  %tmp20 = xor i64 %tmp6, -1
+  %tmp21 = add i64 %tmp1, %tmp20
+  %tmp22 = trunc i64 %tmp21 to i32
+  %tmp23 = add i32 %tmp19, %tmp22
+  %tmp24 = add i32 %tmp17, %tmp23
+  br label %bb5
+}