[InstCombine] Don't canonicalize (gep i8* X, -(ptrtoint Y)) as (inttoptr (sub...

It's just a wrong thing to do.

We introduce inttoptr where there were none, which results in
loosing all provenance information because we no longer have a GEP{i,},
and pessimize all future optimizations,
because we are basically not allowed to look past `inttoptr`.

(gep i8* X, -(ptrtoint Y))  *is* the canonical form.
So just drop this fold.

Noticed while reviewing D98120.

diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index 61dd387..5481162 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -2173,16 +2173,6 @@ Instruction *InstCombinerImpl::visitGetElementPtrInst(GetElementPtrInst &GEP) {
       }
 
       if (Matched) {
-        // Canonicalize (gep i8* X, -(ptrtoint Y))
-        // to (inttoptr (sub (ptrtoint X), (ptrtoint Y)))
-        // The GEP pattern is emitted by the SCEV expander for certain kinds of
-        // pointer arithmetic.
-        if (match(V, m_Neg(m_PtrToInt(m_Value())))) {
-          Operator *Index = cast<Operator>(V);
-          Value *PtrToInt = Builder.CreatePtrToInt(PtrOp, Index->getType());
-          Value *NewSub = Builder.CreateSub(PtrToInt, Index->getOperand(1));
-          return CastInst::Create(Instruction::IntToPtr, NewSub, GEPType);
-        }
         // Canonicalize (gep i8* X, (ptrtoint Y)-(ptrtoint X))
         // to (bitcast Y)
         Value *Y;

diff --git a/llvm/test/Transforms/InstCombine/getelementptr.ll b/llvm/test/Transforms/InstCombine/getelementptr.ll
index 877f200..f04ba82 100644 (file)
--- a/llvm/test/Transforms/InstCombine/getelementptr.ll
+++ b/llvm/test/Transforms/InstCombine/getelementptr.ll
@@ -1028,9 +1028,8 @@ define i16 @test41([3 x i32] addrspace(1)* %array) {
 define i8* @test42i(i8* %c1, i8* %c2) {
 ; CHECK-LABEL: @test42i(
 ; CHECK-NEXT:    [[PTRTOINT:%.*]] = ptrtoint i8* [[C1:%.*]] to i64
-; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i8* [[C2:%.*]] to i64
-; CHECK-NEXT:    [[TMP2:%.*]] = sub i64 [[TMP1]], [[PTRTOINT]]
-; CHECK-NEXT:    [[GEP:%.*]] = inttoptr i64 [[TMP2]] to i8*
+; CHECK-NEXT:    [[SUB:%.*]] = sub i64 0, [[PTRTOINT]]
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds i8, i8* [[C2:%.*]], i64 [[SUB]]
 ; CHECK-NEXT:    ret i8* [[GEP]]
 ;
   %ptrtoint = ptrtoint i8* %c1 to i64
@@ -1043,9 +1042,8 @@ define i8* @test42i(i8* %c1, i8* %c2) {
 define i8* @test42(i8* %c1, i8* %c2) {
 ; CHECK-LABEL: @test42(
 ; CHECK-NEXT:    [[PTRTOINT:%.*]] = ptrtoint i8* [[C1:%.*]] to i64
-; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i8* [[C2:%.*]] to i64
-; CHECK-NEXT:    [[TMP2:%.*]] = sub i64 [[TMP1]], [[PTRTOINT]]
-; CHECK-NEXT:    [[GEP:%.*]] = inttoptr i64 [[TMP2]] to i8*
+; CHECK-NEXT:    [[SUB:%.*]] = sub i64 0, [[PTRTOINT]]
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i8, i8* [[C2:%.*]], i64 [[SUB]]
 ; CHECK-NEXT:    ret i8* [[GEP]]
 ;
   %ptrtoint = ptrtoint i8* %c1 to i64
@@ -1058,9 +1056,9 @@ define i8* @test42(i8* %c1, i8* %c2) {
 define i16* @test43i(i16* %c1, i16* %c2) {
 ; CHECK-LABEL: @test43i(
 ; CHECK-NEXT:    [[PTRTOINT:%.*]] = ptrtoint i16* [[C1:%.*]] to i64
-; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint i16* [[C2:%.*]] to i64
-; CHECK-NEXT:    [[TMP2:%.*]] = sub i64 [[TMP1]], [[PTRTOINT]]
-; CHECK-NEXT:    [[GEP:%.*]] = inttoptr i64 [[TMP2]] to i16*
+; CHECK-NEXT:    [[SUB:%.*]] = sub i64 0, [[PTRTOINT]]
+; CHECK-NEXT:    [[SHR:%.*]] = ashr i64 [[SUB]], 1
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds i16, i16* [[C2:%.*]], i64 [[SHR]]
 ; CHECK-NEXT:    ret i16* [[GEP]]
 ;
   %ptrtoint = ptrtoint i16* %c1 to i64
@@ -1074,9 +1072,9 @@ define i16* @test43i(i16* %c1, i16* %c2) {
 define %struct.C* @test44i(%struct.C* %c1, %struct.C* %c2) {
 ; CHECK-LABEL: @test44i(
 ; CHECK-NEXT:    [[PTRTOINT:%.*]] = ptrtoint %struct.C* [[C1:%.*]] to i64
-; CHECK-NEXT:    [[TMP1:%.*]] = ptrtoint %struct.C* [[C2:%.*]] to i64
-; CHECK-NEXT:    [[TMP2:%.*]] = sub i64 [[TMP1]], [[PTRTOINT]]
-; CHECK-NEXT:    [[GEP:%.*]] = inttoptr i64 [[TMP2]] to %struct.C*
+; CHECK-NEXT:    [[SUB:%.*]] = sub i64 0, [[PTRTOINT]]
+; CHECK-NEXT:    [[SHR:%.*]] = sdiv i64 [[SUB]], 7
+; CHECK-NEXT:    [[GEP:%.*]] = getelementptr inbounds [[STRUCT_C:%.*]], %struct.C* [[C2:%.*]], i64 [[SHR]]
 ; CHECK-NEXT:    ret %struct.C* [[GEP]]
 ;
   %ptrtoint = ptrtoint %struct.C* %c1 to i64