Reapply [InstCombine] Simplify and/or of icmp eq with op replacement (#70335)

Relative to the first attempt, this contains two changes:

First, we only handle the case where one side simplifies to true or
false, instead of calling simplification recursively. The previous
approach would return poison if one operand simplified to poison
(under the equality assumption), which is incorrect.

Second, we do not fold llvm.is.constant in simplifyWithOpReplaced().
We may be assuming that a value is constant, if the equality holds,
but it may not actually be constant. This is nominally just a QoI
issue, but the std::list implementation in libstdc++ relies on the
precise behavior in a way that causes miscompiles.

-----

and/or in logical (select) form benefit from generic simplifications via
simplifyWithOpReplaced(). However, the corresponding fold for plain
and/or currently does not exist.

Similar to selects, there are two general cases for this fold
(illustrated with `and`, but there are `or` conjugates).

The basic case is something like `(a == b) & c`, where the replacement
of a with b or b with a inside c allows it to fold to true or false.
Then the whole operation will fold to either false or `a == b`.

The second case is something like `(a != b) & c`, where the replacement
inside c allows it to fold to false. In that case, the operand can be
replaced with c, because in the case where a == b (and thus the icmp is
false), c itself will already be false.

As the test diffs show, this catches quite a lot of patterns in existing
test coverage. This also obsoletes quite a few existing special-case
and/or of icmp folds we have (e.g. simplifyAndOrOfICmpsWithLimitConst),
but I haven't removed anything as part of this patch in the interest of
risk mitigation.

Fixes #69050.
Fixes #69091.

Author: nikic

llvm/lib/Analysis/InstructionSimplify.cpp

@@ -2026,6 +2026,58 @@ static Value *simplifyAndOrOfCmps(const SimplifyQuery &Q, Value *Op0,
   return nullptr;
 }
 
+static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
+                                     const SimplifyQuery &Q,
+                                     bool AllowRefinement,
+                                     SmallVectorImpl<Instruction *> *DropFlags,
+                                     unsigned MaxRecurse);
+
+static Value *simplifyAndOrWithICmpEq(unsigned Opcode, Value *Op0, Value *Op1,
+                                      const SimplifyQuery &Q,
+                                      unsigned MaxRecurse) {
+  assert((Opcode == Instruction::And || Opcode == Instruction::Or) &&
+         "Must be and/or");
+  ICmpInst::Predicate Pred;
+  Value *A, *B;
+  if (!match(Op0, m_ICmp(Pred, m_Value(A), m_Value(B))) ||
+      !ICmpInst::isEquality(Pred))
+    return nullptr;
+
+  auto Simplify = [&](Value *Res) -> Value * {
+    Constant *Absorber = ConstantExpr::getBinOpAbsorber(Opcode, Res->getType());
+
+    // and (icmp eq a, b), x implies (a==b) inside x.
+    // or (icmp ne a, b), x implies (a==b) inside x.
+    // If x simplifies to true/false, we can simplify the and/or.
+    if (Pred ==
+        (Opcode == Instruction::And ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_NE)) {
+      if (Res == Absorber)
+        return Absorber;
+      if (Res == ConstantExpr::getBinOpIdentity(Opcode, Res->getType()))
+        return Op0;
+      return nullptr;
+    }
+
+    // If we have and (icmp ne a, b), x and for a==b we can simplify x to false,
+    // then we can drop the icmp, as x will already be false in the case where
+    // the icmp is false. Similar for or and true.
+    if (Res == Absorber)
+      return Op1;
+    return nullptr;
+  };
+
+  if (Value *Res =
+          simplifyWithOpReplaced(Op1, A, B, Q, /* AllowRefinement */ true,
+                                 /* DropFlags */ nullptr, MaxRecurse))
+    return Simplify(Res);
+  if (Value *Res =
+          simplifyWithOpReplaced(Op1, B, A, Q, /* AllowRefinement */ true,
+                                 /* DropFlags */ nullptr, MaxRecurse))
+    return Simplify(Res);
+
+  return nullptr;
+}
+
 /// Given a bitwise logic op, check if the operands are add/sub with a common
 /// source value and inverted constant (identity: C - X -> ~(X + ~C)).
 static Value *simplifyLogicOfAddSub(Value *Op0, Value *Op1,
@@ -2160,6 +2212,13 @@ static Value *simplifyAndInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
       isKnownToBeAPowerOfTwo(Op0, Q.DL, /*OrZero*/ true, 0, Q.AC, Q.CxtI, Q.DT))
     return Constant::getNullValue(Op0->getType());
 
+  if (Value *V =
+          simplifyAndOrWithICmpEq(Instruction::And, Op0, Op1, Q, MaxRecurse))
+    return V;
+  if (Value *V =
+          simplifyAndOrWithICmpEq(Instruction::And, Op1, Op0, Q, MaxRecurse))
+    return V;
+
   if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, true))
     return V;
 
@@ -2436,6 +2495,13 @@ static Value *simplifyOrInst(Value *Op0, Value *Op1, const SimplifyQuery &Q,
       match(Op0, m_LShr(m_Specific(X), m_Specific(Y))))
     return Op1;
 
+  if (Value *V =
+          simplifyAndOrWithICmpEq(Instruction::Or, Op0, Op1, Q, MaxRecurse))
+    return V;
+  if (Value *V =
+          simplifyAndOrWithICmpEq(Instruction::Or, Op1, Op0, Q, MaxRecurse))
+    return V;
+
   if (Value *V = simplifyAndOrOfCmps(Q, Op0, Op1, false))
     return V;
 
@@ -4337,6 +4403,10 @@ static Value *simplifyWithOpReplaced(Value *V, Value *Op, Value *RepOp,
       return nullptr;
   }
 
+  // Don't fold away llvm.is.constant checks based on assumptions.
+  if (match(I, m_Intrinsic<Intrinsic::is_constant>()))
+    return nullptr;
+
   // Replace Op with RepOp in instruction operands.
   SmallVector<Value *, 8> NewOps;
   bool AnyReplaced = false;

llvm/test/CodeGen/PowerPC/pr45448.ll

@@ -20,20 +20,16 @@ define hidden void @julia_tryparse_internal_45896() #0 {
 ; CHECK-NEXT:  .LBB0_6: # %fail194
 ; CHECK-NEXT:  .LBB0_7: # %L670
 ; CHECK-NEXT:    li r5, -3
-; CHECK-NEXT:    cmpdi r3, 0
 ; CHECK-NEXT:    sradi r4, r3, 63
 ; CHECK-NEXT:    rldic r5, r5, 4, 32
-; CHECK-NEXT:    crnot 4*cr5+lt, eq
 ; CHECK-NEXT:    mulhdu r3, r3, r5
 ; CHECK-NEXT:    maddld r6, r4, r5, r3
 ; CHECK-NEXT:    cmpld cr1, r6, r3
 ; CHECK-NEXT:    mulhdu. r3, r4, r5
-; CHECK-NEXT:    bc 4, 4*cr5+lt, .LBB0_10
-; CHECK-NEXT:  # %bb.8: # %L670
 ; CHECK-NEXT:    crorc 4*cr5+lt, 4*cr1+lt, eq
-; CHECK-NEXT:    bc 4, 4*cr5+lt, .LBB0_10
-; CHECK-NEXT:  # %bb.9: # %L917
-; CHECK-NEXT:  .LBB0_10: # %L994
+; CHECK-NEXT:    bc 4, 4*cr5+lt, .LBB0_9
+; CHECK-NEXT:  # %bb.8: # %L917
+; CHECK-NEXT:  .LBB0_9: # %L994
 top:
   %0 = load i64, ptr undef, align 8
   %1 = icmp ne i64 %0, 0

llvm/test/Transforms/InstCombine/div-by-0-guard-before-smul_ov.ll

@@ -47,11 +47,7 @@ define i1 @n2_wrong_size(i4 %size0, i4 %size1, i4 %nmemb) {
 
 define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n3_wrong_pred(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[SMUL:%.*]] = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[SMUL_OV:%.*]] = extractvalue { i4, i1 } [[SMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = and i1 [[SMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp eq i4 %size, 0 ; not 'ne'
   %smul = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 %size, i4 %nmemb)
@@ -63,10 +59,7 @@ define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 define i1 @n4_not_and(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n4_not_and(
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[SMUL:%.*]] = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[SMUL_OV:%.*]] = extractvalue { i4, i1 } [[SMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = or i1 [[SMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %cmp = icmp ne i4 %size, 0
   %smul = tail call { i4, i1 } @llvm.smul.with.overflow.i4(i4 %size, i4 %nmemb)

llvm/test/Transforms/InstCombine/div-by-0-guard-before-umul_ov.ll

@@ -47,11 +47,7 @@ define i1 @n2_wrong_size(i4 %size0, i4 %size1, i4 %nmemb) {
 
 define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n3_wrong_pred(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[UMUL:%.*]] = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[UMUL_OV:%.*]] = extractvalue { i4, i1 } [[UMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = and i1 [[UMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 false
 ;
   %cmp = icmp eq i4 %size, 0 ; not 'ne'
   %umul = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 %size, i4 %nmemb)
@@ -63,10 +59,7 @@ define i1 @n3_wrong_pred(i4 %size, i4 %nmemb) {
 define i1 @n4_not_and(i4 %size, i4 %nmemb) {
 ; CHECK-LABEL: @n4_not_and(
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i4 [[SIZE:%.*]], 0
-; CHECK-NEXT:    [[UMUL:%.*]] = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 [[SIZE]], i4 [[NMEMB:%.*]])
-; CHECK-NEXT:    [[UMUL_OV:%.*]] = extractvalue { i4, i1 } [[UMUL]], 1
-; CHECK-NEXT:    [[AND:%.*]] = or i1 [[UMUL_OV]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[AND]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %cmp = icmp ne i4 %size, 0
   %umul = tail call { i4, i1 } @llvm.umul.with.overflow.i4(i4 %size, i4 %nmemb)

llvm/test/Transforms/InstCombine/ispow2.ll

@@ -392,9 +392,7 @@ define i1 @is_pow2_ctpop_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @is_pow2_ctpop_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i32 [[T0]], 2
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ugt i32 %t0, 2
@@ -946,9 +944,7 @@ define i1 @is_pow2or0_ctpop_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @is_pow2or0_ctpop_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 1
-; CHECK-NEXT:    [[ISZERO:%.*]] = icmp eq i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[ISZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 1
@@ -959,11 +955,7 @@ define i1 @is_pow2or0_ctpop_wrong_pred1(i32 %x) {
 
 define i1 @is_pow2or0_ctpop_wrong_pred2(i32 %x) {
 ; CHECK-LABEL: @is_pow2or0_ctpop_wrong_pred2(
-; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 1
-; CHECK-NEXT:    [[ISZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[ISZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 1
@@ -1149,9 +1141,7 @@ define i1 @isnot_pow2nor0_ctpop_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @isnot_pow2nor0_ctpop_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[T0]], 1
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp eq i32 %t0, 1
@@ -1162,11 +1152,7 @@ define i1 @isnot_pow2nor0_ctpop_wrong_pred1(i32 %x) {
 
 define i1 @isnot_pow2nor0_ctpop_wrong_pred2(i32 %x) {
 ; CHECK-LABEL: @isnot_pow2nor0_ctpop_wrong_pred2(
-; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]]), !range [[RNG0]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[T0]], 1
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp eq i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = and i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 false
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp eq i32 %t0, 1

llvm/test/Transforms/InstSimplify/and-or-icmp-ctpop.ll

@@ -40,11 +40,7 @@ define <2 x i1> @eq_or_non_0_commute(<2 x i32> %x) {
 
 define i1 @eq_or_non_0_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @eq_or_non_0_wrong_pred1(
-; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]])
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 10
-; CHECK-NEXT:    [[NOTZERO:%.*]] = icmp ne i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[NOTZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 true
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 10
@@ -90,9 +86,7 @@ define i1 @ne_and_is_0_wrong_pred1(i32 %x) {
 ; CHECK-LABEL: @ne_and_is_0_wrong_pred1(
 ; CHECK-NEXT:    [[T0:%.*]] = tail call i32 @llvm.ctpop.i32(i32 [[X:%.*]])
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[T0]], 10
-; CHECK-NEXT:    [[ISZERO:%.*]] = icmp eq i32 [[X]], 0
-; CHECK-NEXT:    [[R:%.*]] = or i1 [[ISZERO]], [[CMP]]
-; CHECK-NEXT:    ret i1 [[R]]
+; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %t0 = tail call i32 @llvm.ctpop.i32(i32 %x)
   %cmp = icmp ne i32 %t0, 10