[InstCombine] Preserve inbounds when folding select of GEP

[nikic]

Summary

The select base, (gep base, offset) to gep base, select (0, offset) fold used to drop inbounds, because the gep base, 0 this introduces might not be inbounds. After the semantics change in D154051, such a GEP is always considered inbounds, in which allows us to preserve the flag here.

As the PhaseOrdering test demonstrates, this can result in major optimization improvements.

Depends on: D154051

[nlopes]

LGTM

llvm/lib/Transforms/InstCombine/InstCombineSelect.cpp

if (isa<VectorType>(CondVal->getType()) && !isa<VectorType>(Idx->getType()))
       return nullptr;
     Type *ElementType = Gep->getResultElementType();
     Value *NewT = Idx;
     Value *NewF = Constant::getNullValue(Idx->getType());
     if (Swap)
       std::swap(NewT, NewF);
     Value *NewSI =
         Builder.CreateSelect(CondVal, NewT, NewF, SI.getName() + ".idx", &SI);
+    if (Gep->isInBounds())
+      return GetElementPtrInst::CreateInBounds(ElementType, Ptr, {NewSI});
     return GetElementPtrInst::Create(ElementType, Ptr, {NewSI});
   };
   if (auto *TrueGep = dyn_cast<GetElementPtrInst>(TrueVal))
     if (auto *NewGep = SelectGepWithBase(TrueGep, FalseVal, false))
       return NewGep;
   if (auto *FalseGep = dyn_cast<GetElementPtrInst>(FalseVal))
     if (auto *NewGep = SelectGepWithBase(FalseGep, TrueVal, true))
       return NewGep;

llvm/test/Transforms/InstCombine/select-gep.ll

;
   ret ptr %select
 }
 
 ; PR50183
 define ptr @test2a(ptr %p, i64 %x, i64 %y) {
 ; CHECK-LABEL: @test2a(
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i64 [[X:%.*]], [[Y:%.*]]
 ; CHECK-NEXT:    [[SELECT_IDX:%.*]] = select i1 [[CMP]], i64 [[X]], i64 0
-; CHECK-NEXT:    [[SELECT:%.*]] = getelementptr i32, ptr [[P:%.*]], i64 [[SELECT_IDX]]
+; CHECK-NEXT:    [[SELECT:%.*]] = getelementptr inbounds i32, ptr [[P:%.*]], i64 [[SELECT_IDX]]
 ; CHECK-NEXT:    ret ptr [[SELECT]]
 ;
   %gep = getelementptr inbounds i32, ptr %p, i64 %x
   %cmp = icmp ugt i64 %x, %y
   %select = select i1 %cmp, ptr %gep, ptr %p
   ret ptr %select
 }
 
 ; PR50183
 define ptr @test2b(ptr %p, i64 %x, i64 %y) {
 ; CHECK-LABEL: @test2b(
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ugt i64 [[X:%.*]], [[Y:%.*]]
 ; CHECK-NEXT:    [[SELECT_IDX:%.*]] = select i1 [[CMP]], i64 0, i64 [[X]]
-; CHECK-NEXT:    [[SELECT:%.*]] = getelementptr i32, ptr [[P:%.*]], i64 [[SELECT_IDX]]
+; CHECK-NEXT:    [[SELECT:%.*]] = getelementptr inbounds i32, ptr [[P:%.*]], i64 [[SELECT_IDX]]
 ; CHECK-NEXT:    ret ptr [[SELECT]]
 ;
   %gep = getelementptr inbounds i32, ptr %p, i64 %x
   %cmp = icmp ugt i64 %x, %y
   %select = select i1 %cmp, ptr %p, ptr %gep
   ret ptr %select
 }
 
 ; PR51069
 define ptr @test2c(ptr %p, i64 %x, i64 %y) {
 ; CHECK-LABEL: @test2c(
 ; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr inbounds i32, ptr [[P:%.*]], i64 [[X:%.*]]
 ; CHECK-NEXT:    [[ICMP:%.*]] = icmp ugt i64 [[X]], [[Y:%.*]]
 ; CHECK-NEXT:    [[SEL_IDX:%.*]] = select i1 [[ICMP]], i64 0, i64 6
-; CHECK-NEXT:    [[SEL:%.*]] = getelementptr i32, ptr [[GEP1]], i64 [[SEL_IDX]]
+; CHECK-NEXT:    [[SEL:%.*]] = getelementptr inbounds i32, ptr [[GEP1]], i64 [[SEL_IDX]]
 ; CHECK-NEXT:    ret ptr [[SEL]]
 ;
   %gep1 = getelementptr inbounds i32, ptr %p, i64 %x
   %gep2 = getelementptr inbounds i32, ptr %gep1, i64 6
   %icmp = icmp ugt i64 %x, %y
   %sel = select i1 %icmp, ptr %gep1, ptr %gep2
   ret ptr %sel
 }
 
 ; PR51069
 define ptr @test2d(ptr %p, i64 %x, i64 %y) {
 ; CHECK-LABEL: @test2d(
 ; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr inbounds i32, ptr [[P:%.*]], i64 [[X:%.*]]
 ; CHECK-NEXT:    [[ICMP:%.*]] = icmp ugt i64 [[X]], [[Y:%.*]]
 ; CHECK-NEXT:    [[SEL_IDX:%.*]] = select i1 [[ICMP]], i64 6, i64 0
-; CHECK-NEXT:    [[SEL:%.*]] = getelementptr i32, ptr [[GEP1]], i64 [[SEL_IDX]]
+; CHECK-NEXT:    [[SEL:%.*]] = getelementptr inbounds i32, ptr [[GEP1]], i64 [[SEL_IDX]]
 ; CHECK-NEXT:    ret ptr [[SEL]]
 ;
   %gep1 = getelementptr inbounds i32, ptr %p, i64 %x
   %gep2 = getelementptr inbounds i32, ptr %gep1, i64 6
   %icmp = icmp ugt i64 %x, %y
   %sel = select i1 %icmp, ptr %gep2, ptr %gep1
   ret ptr %sel
 }

 }
 
 ; PR51069 - multiple uses
 define ptr @test6(ptr %p, i64 %x, i64 %y) {
 ; CHECK-LABEL: @test6(
 ; CHECK-NEXT:    [[GEP1:%.*]] = getelementptr inbounds i32, ptr [[P:%.*]], i64 [[X:%.*]]
 ; CHECK-NEXT:    [[ICMP:%.*]] = icmp ugt i64 [[X]], [[Y:%.*]]
 ; CHECK-NEXT:    [[SEL_IDX:%.*]] = select i1 [[ICMP]], i64 [[Y]], i64 0
-; CHECK-NEXT:    [[SEL:%.*]] = getelementptr i32, ptr [[GEP1]], i64 [[SEL_IDX]]
+; CHECK-NEXT:    [[SEL:%.*]] = getelementptr inbounds i32, ptr [[GEP1]], i64 [[SEL_IDX]]
 ; CHECK-NEXT:    call void @use_i32p(ptr [[GEP1]])
 ; CHECK-NEXT:    ret ptr [[SEL]]
 ;
   %gep1 = getelementptr inbounds i32, ptr %p, i64 %x
   %gep2 = getelementptr inbounds i32, ptr %gep1, i64 %y
   %icmp = icmp ugt i64 %x, %y
   %sel = select i1 %icmp, ptr %gep2, ptr %gep1
   call void @use_i32p(ptr %gep1)

llvm/test/Transforms/InstCombine/stpncpy-1.ll

 ; ANY: @[[STR_9:[a-zA-Z0-9_$"\\.-]+]] = private unnamed_addr constant [10 x i8] c"1234\00\00\00\00\00\00", align 1
 ;.
 define void @fold_stpncpy_overlap(ptr %dst, i64 %n) {
 ; ANY-LABEL: @fold_stpncpy_overlap(
 ; ANY-NEXT:    call void @sink(ptr [[DST:%.*]], ptr [[DST]])
 ; ANY-NEXT:    [[STXNCPY_CHAR0:%.*]] = load i8, ptr [[DST]], align 1
 ; ANY-NEXT:    [[STPNCPY_CHAR0CMP:%.*]] = icmp ne i8 [[STXNCPY_CHAR0]], 0
 ; ANY-NEXT:    [[STPNCPY_SEL_IDX:%.*]] = zext i1 [[STPNCPY_CHAR0CMP]] to i64
-; ANY-NEXT:    [[STPNCPY_SEL:%.*]] = getelementptr i8, ptr [[DST]], i64 [[STPNCPY_SEL_IDX]]
-; ANY-NEXT:    call void @sink(ptr nonnull [[DST]], ptr [[STPNCPY_SEL]])
+; ANY-NEXT:    [[STPNCPY_SEL:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 [[STPNCPY_SEL_IDX]]
+; ANY-NEXT:    call void @sink(ptr nonnull [[DST]], ptr nonnull [[STPNCPY_SEL]])
 ; ANY-NEXT:    ret void
 ;
 ; Fold stpncpy(D, D, 0) to just D.
   %es_0 = call ptr @stpncpy(ptr %dst, ptr %dst, i64 0)
   call void @sink(ptr %dst, ptr %es_0)
 
 ; Fold stpncpy(D, D, 1) to D + (*D != '\0').
   %es_1 = call ptr @stpncpy(ptr %dst, ptr %dst, i64 1)

 
 define void @fold_stpncpy_s(ptr %dst, ptr %src) {
 ; ANY-LABEL: @fold_stpncpy_s(
 ; ANY-NEXT:    call void @sink(ptr [[DST:%.*]], ptr [[DST]])
 ; ANY-NEXT:    [[STXNCPY_CHAR0:%.*]] = load i8, ptr [[SRC:%.*]], align 1
 ; ANY-NEXT:    store i8 [[STXNCPY_CHAR0]], ptr [[DST]], align 1
 ; ANY-NEXT:    [[STPNCPY_CHAR0CMP:%.*]] = icmp ne i8 [[STXNCPY_CHAR0]], 0
 ; ANY-NEXT:    [[STPNCPY_SEL_IDX:%.*]] = zext i1 [[STPNCPY_CHAR0CMP]] to i64
-; ANY-NEXT:    [[STPNCPY_SEL:%.*]] = getelementptr i8, ptr [[DST]], i64 [[STPNCPY_SEL_IDX]]
-; ANY-NEXT:    call void @sink(ptr nonnull [[DST]], ptr [[STPNCPY_SEL]])
+; ANY-NEXT:    [[STPNCPY_SEL:%.*]] = getelementptr inbounds i8, ptr [[DST]], i64 [[STPNCPY_SEL_IDX]]
+; ANY-NEXT:    call void @sink(ptr nonnull [[DST]], ptr nonnull [[STPNCPY_SEL]])
 ; ANY-NEXT:    ret void
 ;
 ; Fold stpncpy(D, S, 0) to just D.
   %es_0 = call ptr @stpncpy(ptr %dst, ptr %src, i64 0)
   call void @sink(ptr %dst, ptr %es_0)
 
 ; Transform stpncpy(D, "", 1) to *D = '\0, D.
   %es_1 = call ptr @stpncpy(ptr %dst, ptr %src, i64 1)

llvm/test/Transforms/PhaseOrdering/gep-null-compare-in-loop.ll

bb:
   store i32 3, ptr %getelementptr1, align 4
   %getelementptr2 = getelementptr i32, ptr %alloca, i32 3
   %call = call i32 @loop(ptr %alloca, ptr %getelementptr2)
   ret i32 %call
 }
 
 define i32 @using_malloc() {
 ; CHECK-LABEL: define i32 @using_malloc
-; CHECK-SAME: () local_unnamed_addr #[[ATTR1:[0-9]+]] {
+; CHECK-SAME: () local_unnamed_addr #[[ATTR0]] {
 ; CHECK-NEXT:  bb:
-; CHECK-NEXT:    [[ALLOC:%.*]] = tail call dereferenceable_or_null(64) ptr @malloc(i64 64)
-; CHECK-NEXT:    store i32 1, ptr [[ALLOC]], align 4
-; CHECK-NEXT:    [[GETELEMENTPTR:%.*]] = getelementptr i32, ptr [[ALLOC]], i64 1
-; CHECK-NEXT:    store i32 2, ptr [[GETELEMENTPTR]], align 4
-; CHECK-NEXT:    [[GETELEMENTPTR1:%.*]] = getelementptr i32, ptr [[ALLOC]], i64 2
-; CHECK-NEXT:    store i32 3, ptr [[GETELEMENTPTR1]], align 4
-; CHECK-NEXT:    [[SPEC_SELECT_I:%.*]] = getelementptr i32, ptr [[ALLOC]], i64 1
-; CHECK-NEXT:    [[LOAD_I:%.*]] = load i32, ptr [[ALLOC]], align 4
-; CHECK-NEXT:    [[SPEC_SELECT_I_1:%.*]] = getelementptr i32, ptr [[ALLOC]], i64 2
-; CHECK-NEXT:    [[LOAD_I_1:%.*]] = load i32, ptr [[SPEC_SELECT_I]], align 4
-; CHECK-NEXT:    [[ADD_I_1:%.*]] = add i32 [[LOAD_I_1]], [[LOAD_I]]
-; CHECK-NEXT:    [[LOAD_I_2:%.*]] = load i32, ptr [[SPEC_SELECT_I_1]], align 4
-; CHECK-NEXT:    [[ADD_I_2:%.*]] = add i32 [[LOAD_I_2]], [[ADD_I_1]]
-; CHECK-NEXT:    ret i32 [[ADD_I_2]]
+; CHECK-NEXT:    ret i32 6
 ;
 bb:
   %alloc = call dereferenceable_or_null(64) ptr @malloc(i64 64)
   store i32 1, ptr %alloc, align 4
   %getelementptr = getelementptr i32, ptr %alloc, i64 1
   store i32 2, ptr %getelementptr, align 4
   %getelementptr1 = getelementptr i32, ptr %alloc, i64 2
   store i32 3, ptr %getelementptr1, align 4
   %getelementptr2 = getelementptr i32, ptr %alloc, i64 3
   %call = call i32 @loop(ptr %alloc, ptr %getelementptr2)
   ret i32 %call
 }
 
 declare ptr @malloc(i64)
 
 declare void @free(ptr)