1*cff8d716Sv01dXYZ; RUN: opt < %s -S -passes='print<scalar-evolution>,loop-unroll<peeling;full-unroll-max=0>,print<scalar-evolution>' 2>&1 | FileCheck %s 2*cff8d716Sv01dXYZ; 3*cff8d716Sv01dXYZ; This test ensures that (extractvalue 0 (with-overflow-inst op0, op1)) 4*cff8d716Sv01dXYZ; is invalidated by LoopPeel when the operands of with-overflow-inst 5*cff8d716Sv01dXYZ; are changed. 6*cff8d716Sv01dXYZ; 7*cff8d716Sv01dXYZ; In the following case, LoopPeel modifies the CFG into another one 8*cff8d716Sv01dXYZ; with %bb7 not dominating %bb2 and %bb3 although %extractvalue is 9*cff8d716Sv01dXYZ; still the step for the %bb3 loop. %call has been modified and uses 10*cff8d716Sv01dXYZ; different operands but the SCEV value for %extractvalue has not been 11*cff8d716Sv01dXYZ; invalidated and still refers to %load in its SCEV operands 12*cff8d716Sv01dXYZ; (SCEV(%extractvalue) := -2 + -2 * %load). 13*cff8d716Sv01dXYZ; 14*cff8d716Sv01dXYZ; When LoopUnroll tries to compute the SCEV for the %bb3 Phi, the 15*cff8d716Sv01dXYZ; stale data for %extractvalue is used whereas %load is not available 16*cff8d716Sv01dXYZ; in %bb3 which is wrong. 17*cff8d716Sv01dXYZ; 18*cff8d716Sv01dXYZ; for more details and nice pictures: https://github.com/llvm/llvm-project/issues/97586 19*cff8d716Sv01dXYZ; 20*cff8d716Sv01dXYZ; Although the reason for the bug was in forgetValue, it is still relevant to 21*cff8d716Sv01dXYZ; test if LoopPeel invalidates %extractvalue after changing %call. 22*cff8d716Sv01dXYZ; 23*cff8d716Sv01dXYZ; forgetValue only walks the users, so calling it on the IV Phis does not 24*cff8d716Sv01dXYZ; invalidate %extractvalue (thus forgetLoop does not invalidate it too). 25*cff8d716Sv01dXYZ; It has to be done by LoopPeel itself. 26*cff8d716Sv01dXYZ 27*cff8d716Sv01dXYZ 28*cff8d716Sv01dXYZdefine void @loop_peeling_smul_with_overflow() { 29*cff8d716Sv01dXYZ; before loop-unroll 30*cff8d716Sv01dXYZ; CHECK: Classifying expressions for: @loop_peeling_smul_with_overflow 31*cff8d716Sv01dXYZ; CHECK: %extractvalue = extractvalue { i32, i1 } %call, 0 32*cff8d716Sv01dXYZ; CHECK-NEXT: --> (-2 + (-2 * %load)) 33*cff8d716Sv01dXYZ; CHECK: %phi4 = phi i32 [ %add, %bb3 ], [ 0, %bb2 ] 34*cff8d716Sv01dXYZ; CHECK-NEXT: --> {0,+,(-2 + (-2 * %load))}<nuw><nsw><%bb3> 35*cff8d716Sv01dXYZ; after loop-unroll 36*cff8d716Sv01dXYZ; CHECK: Classifying expressions for: @loop_peeling_smul_with_overflow 37*cff8d716Sv01dXYZ; CHECK: %extractvalue = extractvalue { i32, i1 } %call, 0 38*cff8d716Sv01dXYZ; CHECK-NEXT: --> (-2 * %add8.lcssa) 39*cff8d716Sv01dXYZ; CHECK: %phi4 = phi i32 [ %add, %bb3 ], [ 0, %bb2 ] 40*cff8d716Sv01dXYZ; CHECK-NEXT: --> {0,+,(-2 * %add8.lcssa)}<nuw><nsw><%bb3> 41*cff8d716Sv01dXYZ; 42*cff8d716Sv01dXYZbb: 43*cff8d716Sv01dXYZ br label %bb1 44*cff8d716Sv01dXYZ 45*cff8d716Sv01dXYZbb1: ; preds = %bb3, %bb 46*cff8d716Sv01dXYZ %phi = phi i32 [ 0, %bb ], [ %phi4, %bb3 ] 47*cff8d716Sv01dXYZ br label %bb5 48*cff8d716Sv01dXYZ 49*cff8d716Sv01dXYZbb2: ; preds = %bb7 50*cff8d716Sv01dXYZ %call = call { i32, i1 } @llvm.smul.with.overflow.i32(i32 %add8, i32 -2) 51*cff8d716Sv01dXYZ %extractvalue = extractvalue { i32, i1 } %call, 0 52*cff8d716Sv01dXYZ br label %bb3 53*cff8d716Sv01dXYZ 54*cff8d716Sv01dXYZbb3: ; preds = %bb3, %bb2 55*cff8d716Sv01dXYZ %phi4 = phi i32 [ %add, %bb3 ], [ 0, %bb2 ] 56*cff8d716Sv01dXYZ %add = add i32 %extractvalue, %phi4 57*cff8d716Sv01dXYZ br i1 false, label %bb3, label %bb1 58*cff8d716Sv01dXYZ 59*cff8d716Sv01dXYZbb5: ; preds = %bb7, %bb1 60*cff8d716Sv01dXYZ %phi6 = phi i32 [ 1, %bb1 ], [ 0, %bb7 ] 61*cff8d716Sv01dXYZ %icmp = icmp eq i32 %phi, 0 62*cff8d716Sv01dXYZ br i1 %icmp, label %bb9, label %bb7 63*cff8d716Sv01dXYZ 64*cff8d716Sv01dXYZbb7: ; preds = %bb5 65*cff8d716Sv01dXYZ %load = load i32, ptr addrspace(1) null, align 4 66*cff8d716Sv01dXYZ %add8 = add i32 %load, 1 67*cff8d716Sv01dXYZ br i1 false, label %bb2, label %bb5 68*cff8d716Sv01dXYZ 69*cff8d716Sv01dXYZbb9: ; preds = %bb5 70*cff8d716Sv01dXYZ ret void 71*cff8d716Sv01dXYZ} 72