[ConstraintElim] Apply add with neg constant first during decomp. (llvm#164791)

We have dedicated decomposition logic for (add %x, -C), but if we have
(add nsw %x, -C) we will first apply the generic logic for NSWAdd, which
gives worse results for negative constants in practice.

Update the code to first apply the pattern with negative constants.

Helps to remove a number of runtime checks in practice:
dtcxzyw/llvm-opt-benchmark#2968

Alive2 proofs for the test changes: https://alive2.llvm.org/ce/z/JfR2Ma

PR: llvm#164791

Author: fhahn

llvm/lib/Transforms/Scalar/ConstraintElimination.cpp

@@ -614,6 +614,16 @@ static Decomposition decompose(Value *V,
     return {V, IsKnownNonNegative};
   }
 
+  if (match(V, m_Add(m_Value(Op0), m_ConstantInt(CI))) && CI->isNegative() &&
+      canUseSExt(CI)) {
+    Preconditions.emplace_back(
+        CmpInst::ICMP_UGE, Op0,
+        ConstantInt::get(Op0->getType(), CI->getSExtValue() * -1));
+    if (auto Decomp = MergeResults(Op0, CI, true))
+      return *Decomp;
+    return {V, IsKnownNonNegative};
+  }
+
   if (match(V, m_NSWAdd(m_Value(Op0), m_Value(Op1)))) {
     if (!isKnownNonNegative(Op0, DL))
       Preconditions.emplace_back(CmpInst::ICMP_SGE, Op0,
@@ -627,16 +637,6 @@ static Decomposition decompose(Value *V,
     return {V, IsKnownNonNegative};
   }
 
-  if (match(V, m_Add(m_Value(Op0), m_ConstantInt(CI))) && CI->isNegative() &&
-      canUseSExt(CI)) {
-    Preconditions.emplace_back(
-        CmpInst::ICMP_UGE, Op0,
-        ConstantInt::get(Op0->getType(), CI->getSExtValue() * -1));
-    if (auto Decomp = MergeResults(Op0, CI, true))
-      return *Decomp;
-    return {V, IsKnownNonNegative};
-  }
-
   // Decompose or as an add if there are no common bits between the operands.
   if (match(V, m_DisjointOr(m_Value(Op0), m_ConstantInt(CI)))) {
     if (auto Decomp = MergeResults(Op0, CI, IsSigned))

llvm/test/Transforms/ConstraintElimination/add-nsw.ll

@@ -757,8 +757,7 @@ define i1 @add_neg_1_known_sge_ult_1(i32 %a) {
 ; CHECK-NEXT:    [[A_SGE:%.*]] = icmp sge i32 [[A:%.*]], 1
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[A_SGE]])
 ; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[A]], -1
-; CHECK-NEXT:    [[C:%.*]] = icmp ult i32 [[SUB]], [[A]]
-; CHECK-NEXT:    ret i1 [[C]]
+; CHECK-NEXT:    ret i1 true
 ;
 entry:
   %a.sge = icmp sge i32 %a, 1
@@ -823,8 +822,7 @@ define i1 @add_neg_3_known_sge_ult_1(i32 %a) {
 ; CHECK-NEXT:    [[A_SGE:%.*]] = icmp sge i32 [[A:%.*]], 3
 ; CHECK-NEXT:    call void @llvm.assume(i1 [[A_SGE]])
 ; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[A]], -3
-; CHECK-NEXT:    [[C:%.*]] = icmp ult i32 [[SUB]], [[A]]
-; CHECK-NEXT:    ret i1 [[C]]
+; CHECK-NEXT:    ret i1 true
 ;
 entry:
   %a.sge = icmp sge i32 %a, 3

llvm/test/Transforms/ConstraintElimination/gep-arithmetic-add.ll

@@ -389,8 +389,7 @@ define i1 @gep_count_add_1_sge_known_ult_1(i32 %count, ptr %p) {
 ; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[COUNT]], -1
 ; CHECK-NEXT:    [[SUB_EXT:%.*]] = zext i32 [[SUB]] to i64
 ; CHECK-NEXT:    [[GEP_SUB:%.*]] = getelementptr inbounds i32, ptr [[P]], i64 [[SUB_EXT]]
-; CHECK-NEXT:    [[C:%.*]] = icmp ult ptr [[GEP_SUB]], [[GEP_COUNT]]
-; CHECK-NEXT:    ret i1 [[C]]
+; CHECK-NEXT:    ret i1 true
 ;
 entry:
   %sge = icmp sge i32 %count, 1
@@ -415,8 +414,7 @@ define i1 @gep_count_add_1_sge_known_uge_1(i32 %count, ptr %p) {
 ; CHECK-NEXT:    [[SUB:%.*]] = add nsw i32 [[COUNT]], -1
 ; CHECK-NEXT:    [[SUB_EXT:%.*]] = zext i32 [[SUB]] to i64
 ; CHECK-NEXT:    [[GEP_SUB:%.*]] = getelementptr inbounds i32, ptr [[P]], i64 [[SUB_EXT]]
-; CHECK-NEXT:    [[C:%.*]] = icmp uge ptr [[GEP_SUB]], [[P]]
-; CHECK-NEXT:    ret i1 [[C]]
+; CHECK-NEXT:    ret i1 true
 ;
 entry:
   %sge = icmp sge i32 %count, 1