[InstCombine] Fold icmp of gep chain with base

nikic · nikic · commit 9ca8ba2061d2 · 2025-07-01T15:24:18.000+02:00
Fold icmp between a chain of geps and its based pointer. Previously
only a single gep was supported.

This will be extended to handle the case of two gep chains with a
common base in a followup.
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp
@@ -708,14 +708,14 @@ Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
     return I;
   };
 
-  Value *PtrBase = GEPLHS->getOperand(0);
-  if (PtrBase == RHS && CanFold(GEPLHS->getNoWrapFlags())) {
+  CommonPointerBase Base = CommonPointerBase::compute(GEPLHS, RHS);
+  if (Base.Ptr == RHS && CanFold(Base.LHSNW)) {
     // ((gep Ptr, OFFSET) cmp Ptr)   ---> (OFFSET cmp 0).
-    GEPNoWrapFlags NW = GEPLHS->getNoWrapFlags();
-    // Do not access GEPLHS after EmitGEPOffset, as the instruction may be
-    // destroyed.
-    Value *Offset = EmitGEPOffset(GEPLHS, /*RewriteGEP=*/true);
-    return NewICmp(NW, Offset, Constant::getNullValue(Offset->getType()));
+    Type *IdxTy = DL.getIndexType(GEPLHS->getType());
+    Value *Offset =
+        EmitGEPOffsets(Base.LHSGEPs, Base.LHSNW, IdxTy, /*RewriteGEPs=*/true);
+    return NewICmp(Base.LHSNW, Offset,
+                   Constant::getNullValue(Offset->getType()));
   }
 
   if (GEPLHS->isInBounds() && ICmpInst::isEquality(Cond) &&
@@ -752,6 +752,7 @@ Instruction *InstCombinerImpl::foldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
 
     // If the base pointers are different, but the indices are the same, just
     // compare the base pointer.
+    Value *PtrBase = GEPLHS->getOperand(0);
     if (PtrBase != GEPRHS->getOperand(0)) {
       bool IndicesTheSame =
           GEPLHS->getNumOperands() == GEPRHS->getNumOperands() &&
diff --git a/llvm/test/Transforms/InstCombine/getelementptr.ll b/llvm/test/Transforms/InstCombine/getelementptr.ll
@@ -688,9 +688,9 @@ define i32 @test28() nounwind  {
 ; CHECK-NEXT:    [[INDVAR:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[INDVAR_NEXT:%.*]], [[BB10]] ]
 ; CHECK-NEXT:    [[T12_REC:%.*]] = xor i32 [[INDVAR]], -1
 ; CHECK-NEXT:    [[TMP0:%.*]] = sext i32 [[T12_REC]] to i64
-; CHECK-NEXT:    [[T12:%.*]] = getelementptr inbounds [[STRUCT_X:%.*]], ptr [[T45]], i64 [[TMP0]]
+; CHECK-NEXT:    [[T12:%.*]] = getelementptr inbounds i8, ptr [[T45]], i64 [[TMP0]]
 ; CHECK-NEXT:    [[T16:%.*]] = call i32 (ptr, ...) @printf(ptr noundef nonnull dereferenceable(1) @.str1, ptr nonnull [[T12]]) #[[ATTR0]]
-; CHECK-NEXT:    [[T84:%.*]] = icmp eq ptr [[T12]], [[ORIENTATIONS]]
+; CHECK-NEXT:    [[T84:%.*]] = icmp eq i32 [[INDVAR]], 0
 ; CHECK-NEXT:    [[INDVAR_NEXT]] = add i32 [[INDVAR]], 1
 ; CHECK-NEXT:    br i1 [[T84]], label [[BB17:%.*]], label [[BB10]]
 ; CHECK:       bb17:
diff --git a/llvm/test/Transforms/InstCombine/icmp-gep.ll b/llvm/test/Transforms/InstCombine/icmp-gep.ll
@@ -788,9 +788,9 @@ define i1 @gep_diff_base_same_indices_nuw_nusw(ptr %x, ptr %y, i64 %z) {
 
 define i1 @gep_multiple_eq(ptr %base, i64 %idx, i64 %idx2) {
 ; CHECK-LABEL: @gep_multiple_eq(
-; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr i32, ptr [[BASE:%.*]], i64 [[IDX:%.*]]
-; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr i32, ptr [[GEP1]], i64 [[IDX2:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[GEP2]], [[BASE]]
+; CHECK-NEXT:    [[GEP1_IDX1:%.*]] = add i64 [[IDX:%.*]], [[IDX2:%.*]]
+; CHECK-NEXT:    [[DOTMASK:%.*]] = and i64 [[GEP1_IDX1]], 4611686018427387903
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[DOTMASK]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %gep1 = getelementptr i32, ptr %base, i64 %idx
@@ -801,9 +801,9 @@ define i1 @gep_multiple_eq(ptr %base, i64 %idx, i64 %idx2) {
 
 define i1 @gep_multiple_eq_commuted(ptr %base, i64 %idx, i64 %idx2) {
 ; CHECK-LABEL: @gep_multiple_eq_commuted(
-; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr i32, ptr [[BASE:%.*]], i64 [[IDX:%.*]]
-; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr i32, ptr [[GEP1]], i64 [[IDX2:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[BASE]], [[GEP2]]
+; CHECK-NEXT:    [[GEP1_IDX1:%.*]] = add i64 [[IDX:%.*]], [[IDX2:%.*]]
+; CHECK-NEXT:    [[DOTMASK:%.*]] = and i64 [[GEP1_IDX1]], 4611686018427387903
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i64 [[DOTMASK]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %gep1 = getelementptr i32, ptr %base, i64 %idx
@@ -814,9 +814,8 @@ define i1 @gep_multiple_eq_commuted(ptr %base, i64 %idx, i64 %idx2) {
 
 define i1 @gep_mugtiple_ugt_nuw(ptr %base, i64 %idx, i64 %idx2) {
 ; CHECK-LABEL: @gep_mugtiple_ugt_nuw(
-; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr nuw i32, ptr [[BASE:%.*]], i64 [[IDX:%.*]]
-; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr nuw i32, ptr [[GEP1]], i64 [[IDX2:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt ptr [[GEP2]], [[BASE]]
+; CHECK-NEXT:    [[GEP1_IDX1:%.*]] = sub i64 0, [[IDX2:%.*]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i64 [[IDX:%.*]], [[GEP1_IDX1]]
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %gep1 = getelementptr nuw i32, ptr %base, i64 %idx
@@ -840,9 +839,9 @@ define i1 @gep_mugtiple_ugt_not_all_nuw(ptr %base, i64 %idx, i64 %idx2) {
 
 define i1 @gep_mugtiple_ugt_inbounds_nusw(ptr %base, i64 %idx, i64 %idx2) {
 ; CHECK-LABEL: @gep_mugtiple_ugt_inbounds_nusw(
-; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr inbounds i32, ptr [[BASE:%.*]], i64 [[IDX:%.*]]
-; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr nusw i32, ptr [[GEP1]], i64 [[IDX2:%.*]]
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt ptr [[GEP2]], [[BASE]]
+; CHECK-NEXT:    [[GEP1_IDX1:%.*]] = add i64 [[IDX:%.*]], [[IDX2:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = shl i64 [[GEP1_IDX1]], 2
+; CHECK-NEXT:    [[CMP:%.*]] = icmp sgt i64 [[TMP1]], 0
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %gep1 = getelementptr inbounds i32, ptr %base, i64 %idx
diff --git a/llvm/test/Transforms/InstCombine/pr39908.ll b/llvm/test/Transforms/InstCombine/pr39908.ll
@@ -7,9 +7,7 @@ target datalayout = "p:32:32"
 
 define i1 @test(ptr %p, i32 %n) {
 ; CHECK-LABEL: @test(
-; CHECK-NEXT:    [[END:%.*]] = getelementptr inbounds [0 x %S], ptr [[P:%.*]], i32 0, i32 [[N:%.*]], i32 0, i32 0
-; CHECK-NEXT:    [[LAST:%.*]] = getelementptr inbounds i8, ptr [[END]], i32 -8
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[LAST]], [[P]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[N:%.*]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %end = getelementptr inbounds [0 x %S], ptr %p, i32 0, i32 %n, i32 0, i32 0
@@ -22,9 +20,7 @@ define i1 @test(ptr %p, i32 %n) {
 define i1 @test64(ptr %p, i64 %n) {
 ; CHECK-LABEL: @test64(
 ; CHECK-NEXT:    [[TMP1:%.*]] = trunc i64 [[N:%.*]] to i32
-; CHECK-NEXT:    [[END:%.*]] = getelementptr inbounds [0 x %S], ptr [[P:%.*]], i32 0, i32 [[TMP1]], i32 0, i32 0
-; CHECK-NEXT:    [[LAST:%.*]] = getelementptr inbounds i8, ptr [[END]], i32 -8
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[LAST]], [[P]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %end = getelementptr inbounds [0 x %S], ptr %p, i64 0, i64 %n, i32 0, i64 0
@@ -37,9 +33,7 @@ define i1 @test64(ptr %p, i64 %n) {
 define i1 @test64_overflow(ptr %p, i64 %n) {
 ; CHECK-LABEL: @test64_overflow(
 ; CHECK-NEXT:    [[TMP1:%.*]] = trunc i64 [[N:%.*]] to i32
-; CHECK-NEXT:    [[END:%.*]] = getelementptr inbounds [0 x %S], ptr [[P:%.*]], i32 0, i32 [[TMP1]], i32 0, i32 0
-; CHECK-NEXT:    [[LAST:%.*]] = getelementptr inbounds i8, ptr [[END]], i32 -8
-; CHECK-NEXT:    [[CMP:%.*]] = icmp eq ptr [[LAST]], [[P]]
+; CHECK-NEXT:    [[CMP:%.*]] = icmp eq i32 [[TMP1]], 1
 ; CHECK-NEXT:    ret i1 [[CMP]]
 ;
   %end = getelementptr inbounds [0 x %S], ptr %p, i64 0, i64 %n, i32 0, i64 8589934592
