review

Author: pskrgag

llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

@@ -1325,26 +1325,18 @@ Instruction *InstCombinerImpl::foldICmpWithZero(ICmpInst &Cmp) {
 //      icmp eq 0, (and num, val - 1)
 // For value being power of two
 Instruction *InstCombinerImpl::foldNextMultiply(ICmpInst &Cmp) {
-  Value *Op0 = Cmp.getOperand(0), *Op1 = Cmp.getOperand(1);
+  Value *Op0 = Cmp.getOperand(0);
   Value *Neg, *Add, *Num, *Mask, *Value;
-  CmpInst::Predicate Pred = Cmp.getPredicate();
-  const APInt *NegConst, *MaskConst, *NumCost;
-
-  if (!ICmpInst::isEquality(Pred))
-    return nullptr;
+  CmpPredicate Pred;
+  const APInt *NegConst, *MaskConst;
 
   // Match num + neg
-  if (!match(Op0, m_c_And(m_Value(Add), m_Value(Neg))))
+  if (!match(Op0, m_OneUse(m_c_And(m_Value(Add), m_Value(Neg)))))
     return nullptr;
 
   // Match num & mask and handle commutative care
-  if (!match(Add, m_c_Add(m_Value(Num), m_Value(Mask)))) {
-    if (match(Neg, m_c_Add(m_Value(Num), m_Value(Mask)))) {
-      std::swap(Add, Neg);
-    } else {
-      return nullptr;
-    }
-  }
+  if (!match(Op0, m_c_And(m_c_Add(m_Value(Num), m_Value(Mask)), m_Value(Neg))))
+    return nullptr;
 
   // Check the constant case
   if (match(Neg, m_APInt(NegConst)) && match(Mask, m_APInt(MaskConst))) {
@@ -1375,12 +1367,11 @@ Instruction *InstCombinerImpl::foldNextMultiply(ICmpInst &Cmp) {
       return nullptr;
   }
 
-  // Guard against weird special-case where Op1 gets optimized to constant.
-  // Leave it constant fonder.
-  if (match(Op1, m_APInt(NumCost)))
+  // Verify that Add and Num are connected by ICmp.
+  if (!match(&Cmp, m_c_ICmp(Pred, m_Value(Add), m_Specific(Num))))
     return nullptr;
 
-  if (!match(Op1, m_Value(Num)))
+  if (!ICmpInst::isEquality(Pred))
     return nullptr;
 
   // Create new icmp eq (num & (val - 1)), 0

llvm/test/Transforms/InstCombine/icmp-add.ll

@@ -3301,6 +3301,8 @@ entry:
   ret i1 %cmp
 }
 
+; PR 152851
+
 define i1 @val_is_aligend_sub(i32 %num, i32 %val) {
 ; CHECK-LABEL: @val_is_aligend_sub(
 ; CHECK-NEXT:    [[TMP1:%.*]] = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 [[NUM:%.*]])
@@ -3406,8 +3408,8 @@ define i1 @val_is_aligend_const_pow2(i32 %num) {
 }
 
 ; Should not work for non-power-of-two cases
-define i1 @val_is_aligend_const_non-pow2(i32 %num) {
-; CHECK-LABEL: @val_is_aligend_const_non-pow2(
+define i1 @val_is_aligend_const_non_pow2(i32 %num) {
+; CHECK-LABEL: @val_is_aligend_const_non_pow2(
 ; CHECK-NEXT:    [[NUM_BIASED:%.*]] = add i32 [[NUM:%.*]], 6
 ; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED]], -7
 ; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], [[NUM]]
@@ -3436,3 +3438,62 @@ define i1 @val_is_aligend_non_pow(i32 %num, i32 %val) {
   %_0 = icmp eq i32 %_2.sroa.0.0, %num
   ret i1 %_0
 }
+
+define i1 @val_is_aligend_const_pow2_multiuse(i32 %num) {
+; CHECK-LABEL: @val_is_aligend_const_pow2_multiuse(
+; CHECK-NEXT:    [[NUM_BIASED:%.*]] = add i32 [[NUM:%.*]], 4095
+; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED]], -4096
+; CHECK-NEXT:    call void @use(i32 [[_2_SROA_0_0]])
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], [[NUM]]
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %num.biased = add i32 %num, 4095
+  %_2.sroa.0.0 = and i32 %num.biased, -4096
+  call void @use(i32 %_2.sroa.0.0)
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}
+
+define i1 @val_is_aligend_const_pow2_multiuse1(i32 %num) {
+; CHECK-LABEL: @val_is_aligend_const_pow2_multiuse1(
+; CHECK-NEXT:    [[NUM_BIASED:%.*]] = add i32 [[NUM:%.*]], 4095
+; CHECK-NEXT:    call void @use(i32 [[NUM_BIASED]])
+; CHECK-NEXT:    [[TMP1:%.*]] = and i32 [[NUM]], 4095
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[TMP1]], 0
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %num.biased = add i32 %num, 4095
+  call void @use(i32 %num.biased)
+  %_2.sroa.0.0 = and i32 %num.biased, -4096
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}
+
+define i1 @val_is_aligend_add_multiuse(i32 %num, i32 %val) {
+; CHECK-LABEL: @val_is_aligend_add_multiuse(
+; CHECK-NEXT:    [[TMP1:%.*]] = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 [[VAL:%.*]])
+; CHECK-NEXT:    [[POW:%.*]] = icmp eq i32 [[TMP1]], 1
+; CHECK-NEXT:    call void @llvm.assume(i1 [[POW]])
+; CHECK-NEXT:    [[MASK:%.*]] = add i32 [[VAL]], -1
+; CHECK-NEXT:    [[NEG:%.*]] = sub nsw i32 0, [[VAL]]
+; CHECK-NEXT:    [[NUM_BIASED:%.*]] = add i32 [[NUM:%.*]], [[MASK]]
+; CHECK-NEXT:    [[_2_SROA_0_0:%.*]] = and i32 [[NUM_BIASED]], [[NEG]]
+; CHECK-NEXT:    call void @use(i32 [[_2_SROA_0_0]])
+; CHECK-NEXT:    [[_0:%.*]] = icmp eq i32 [[_2_SROA_0_0]], [[NUM]]
+; CHECK-NEXT:    ret i1 [[_0]]
+;
+  %1 = tail call range(i32 1, 33) i32 @llvm.ctpop.i32(i32 %val)
+  %pow = icmp eq i32 %1, 1
+  call void @llvm.assume(i1 %pow)
+
+  %mask = add i32 %val, -1
+  %neg = sub nsw i32 0, %val
+
+  %num.biased = add i32 %num, %mask
+  %_2.sroa.0.0 = and i32 %num.biased, %neg
+
+  call void @use(i32 %_2.sroa.0.0)
+
+  %_0 = icmp eq i32 %_2.sroa.0.0, %num
+  ret i1 %_0
+}