test/Transforms/DeadStoreElimination/trivial-dse-calls.ll

; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -passes=dse -S < %s | FileCheck %s

declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture)
declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture)

declare void @unknown()
declare void @f(ptr)
declare void @f2(ptr, ptr)
declare ptr @f3(ptr, ptr)

; Basic case for DSEing a trivially dead writing call
define void @test_dead() {
; CHECK-LABEL: @test_dead(
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  ret void
}

; Add in canonical lifetime intrinsics
define void @test_lifetime() {
; CHECK-LABEL: @test_lifetime(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @llvm.lifetime.start.p0(i64 4, ptr [[A]])
; CHECK-NEXT:    call void @llvm.lifetime.end.p0(i64 4, ptr [[A]])
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @llvm.lifetime.start.p0(i64 4, ptr %a)
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  call void @llvm.lifetime.end.p0(i64 4, ptr %a)
  ret void
}

; Add some unknown calls just to point out that this is use based, not
; instruction order sensitive
define void @test_lifetime2() {
; CHECK-LABEL: @test_lifetime2(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @llvm.lifetime.start.p0(i64 4, ptr [[A]])
; CHECK-NEXT:    call void @unknown()
; CHECK-NEXT:    call void @unknown()
; CHECK-NEXT:    call void @llvm.lifetime.end.p0(i64 4, ptr [[A]])
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @llvm.lifetime.start.p0(i64 4, ptr %a)
  call void @unknown()
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  call void @unknown()
  call void @llvm.lifetime.end.p0(i64 4, ptr %a)
  ret void
}

; As long as the result is unused, we can even remove reads of the alloca
; itself since the write will be dropped.
define void @test_dead_readwrite() {
; CHECK-LABEL: @test_dead_readwrite(
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f(ptr nocapture %a) argmemonly nounwind willreturn
  ret void
}

define i32 @test_neg_read_after() {
; CHECK-LABEL: @test_neg_read_after(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR1:[0-9]+]]
; CHECK-NEXT:    [[RES:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT:    ret i32 [[RES]]
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  %res = load i32, ptr %a
  ret i32 %res
}


define void @test_neg_infinite_loop() {
; CHECK-LABEL: @test_neg_infinite_loop(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR2:[0-9]+]]
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind
  ret void
}

define void @test_neg_throw() {
; CHECK-LABEL: @test_neg_throw(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR3:[0-9]+]]
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly willreturn
  ret void
}

define void @test_neg_extra_write() {
; CHECK-LABEL: @test_neg_extra_write(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR4:[0-9]+]]
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) nounwind willreturn
  ret void
}

; In this case, we can't remove a1 because we need to preserve the write to
; a2, and if we leave the call around, we need memory to pass to the first arg.
define void @test_neg_unmodeled_write() {
; CHECK-LABEL: @test_neg_unmodeled_write(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    [[A2:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f2(ptr nocapture writeonly [[A]], ptr [[A2]]) #[[ATTR1]]
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  %a2 = alloca i32, align 4
  call void @f2(ptr nocapture writeonly %a, ptr %a2) argmemonly nounwind willreturn
  ret void
}

define i32 @test_neg_captured_by_call() {
; CHECK-LABEL: @test_neg_captured_by_call(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    [[A2:%.*]] = alloca ptr, align 4
; CHECK-NEXT:    call void @f2(ptr writeonly [[A]], ptr [[A2]]) #[[ATTR1]]
; CHECK-NEXT:    [[A_COPY_CAST:%.*]] = load ptr, ptr [[A2]], align 8
; CHECK-NEXT:    [[RES:%.*]] = load i32, ptr [[A_COPY_CAST]], align 4
; CHECK-NEXT:    ret i32 [[RES]]
;
  %a = alloca i32, align 4
  %a2 = alloca ptr, align 4
  call void @f2(ptr writeonly %a, ptr %a2) argmemonly nounwind willreturn
  %a_copy_cast = load ptr, ptr %a2
  %res = load i32, ptr %a_copy_cast
  ret i32 %res
}

define i32 @test_neg_captured_before() {
; CHECK-LABEL: @test_neg_captured_before(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    [[A2:%.*]] = alloca ptr, align 4
; CHECK-NEXT:    store ptr [[A]], ptr [[A2]], align 8
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR1]]
; CHECK-NEXT:    [[A_COPY_CAST:%.*]] = load ptr, ptr [[A2]], align 8
; CHECK-NEXT:    [[RES:%.*]] = load i32, ptr [[A_COPY_CAST]], align 4
; CHECK-NEXT:    ret i32 [[RES]]
;
  %a = alloca i32, align 4
  %a2 = alloca ptr, align 4
  store ptr %a, ptr %a2
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  %a_copy_cast = load ptr, ptr %a2
  %res = load i32, ptr %a_copy_cast
  ret i32 %res
}

; Callee might be dead, but op bundle has unknown semantics and thus isn't.
define void @test_new_op_bundle() {
; CHECK-LABEL: @test_new_op_bundle(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR1]] [ "unknown"(ptr [[A]]) ]
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn ["unknown" (ptr %a)]
  ret void
}

; Show that reading from unrelated memory is okay
define void @test_unreleated_read() {
; CHECK-LABEL: @test_unreleated_read(
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  %a2 = alloca i32, align 4
  call void @f2(ptr nocapture writeonly %a, ptr nocapture readonly %a2) argmemonly nounwind willreturn
  ret void
}

; Removing a capture is also okay. The capture can only be in the return value
; (which is unused) or written into the dead out parameter.
define void @test_unrelated_capture() {
; CHECK-LABEL: @test_unrelated_capture(
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  %a2 = alloca i32, align 4
  call ptr @f3(ptr nocapture writeonly %a, ptr readonly %a2) argmemonly nounwind willreturn
  ret void
}

; Cannot remove call, as %a2 is captured via the return value.
define i8 @test_neg_unrelated_capture_used_via_return() {
; CHECK-LABEL: @test_neg_unrelated_capture_used_via_return(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    [[A2:%.*]] = alloca i32, align 4
; CHECK-NEXT:    [[CAPTURE:%.*]] = call ptr @f3(ptr nocapture writeonly [[A]], ptr readonly [[A2]]) #[[ATTR1]]
; CHECK-NEXT:    [[V:%.*]] = load i8, ptr [[CAPTURE]], align 1
; CHECK-NEXT:    ret i8 [[V]]
;
  %a = alloca i32, align 4
  %a2 = alloca i32, align 4
  %capture = call ptr @f3(ptr nocapture writeonly %a, ptr readonly %a2) argmemonly nounwind willreturn
  %v = load i8, ptr %capture
  ret i8 %v
}

; As long as the result is unused, we can even remove reads of the alloca
; itself since the write will be dropped.
define void @test_self_read() {
; CHECK-LABEL: @test_self_read(
; CHECK-NEXT:    ret void
;
  %a = alloca i32, align 4
  call void @f2(ptr nocapture writeonly %a, ptr nocapture readonly %a) argmemonly nounwind willreturn
  ret void
}

; We can remove the call because while we don't know the size of the write done
; by the call, we do know the following store writes to the entire contents of
; the alloca.
define i32 @test_dse_overwrite() {
; CHECK-LABEL: @test_dse_overwrite(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    store i32 0, ptr [[A]], align 4
; CHECK-NEXT:    [[V:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT:    ret i32 [[V]]
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  store i32 0, ptr %a
  %v = load i32, ptr %a
  ret i32 %v
}

; Negative case where we can read part of the value written by @f.
define i32 @test_neg_dse_partial_overwrite() {
; CHECK-LABEL: @test_neg_dse_partial_overwrite(
; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A]]) #[[ATTR1]]
; CHECK-NEXT:    store i8 0, ptr [[A]], align 1
; CHECK-NEXT:    [[V:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT:    ret i32 [[V]]
;
  %a = alloca i32, align 4
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  store i8 0, ptr %a
  %v = load i32, ptr %a
  ret i32 %v
}

; Negative case where we don't know the size of a, and thus can't use the
; full overwrite reasoning
define i32 @test_neg_dse_unsized(ptr %a) {
; CHECK-LABEL: @test_neg_dse_unsized(
; CHECK-NEXT:    call void @f(ptr nocapture writeonly [[A:%.*]]) #[[ATTR1]]
; CHECK-NEXT:    store i32 0, ptr [[A]], align 4
; CHECK-NEXT:    [[V:%.*]] = load i32, ptr [[A]], align 4
; CHECK-NEXT:    ret i32 [[V]]
;
  call void @f(ptr writeonly nocapture %a) argmemonly nounwind willreturn
  store i32 0, ptr %a
  %v = load i32, ptr %a
  ret i32 %v
}

@G = global i8 0

; Same as test_dse_overwrite, but with a non-alloca object.
define void @test_dse_non_alloca() {
; CHECK-LABEL: @test_dse_non_alloca(
; CHECK-NEXT:    store i8 0, ptr @G, align 1
; CHECK-NEXT:    ret void
;
  call void @f(ptr writeonly nocapture @G) argmemonly nounwind willreturn
  store i8 0, ptr @G
  ret void
}