xref: /llvm-project/llvm/test/Transforms/InstCombine/vscale_gep.ll (revision 462cb3cd6cecd0511ecaf0e3ebcaba455ece587d)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -S -passes=instcombine < %s | FileCheck %s
3
4; This test is used to verify we are not crashing at Assertion `CastInst::castIsValid(opc, C, Ty) && "Invalid constantexpr cast!".
5define <vscale x 2 x ptr> @gep_index_type_is_scalable(ptr %p) {
6; CHECK-LABEL: @gep_index_type_is_scalable(
7; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i8, ptr [[P:%.*]], <vscale x 2 x i64> undef
8; CHECK-NEXT:    ret <vscale x 2 x ptr> [[GEP]]
9;
10  %gep = getelementptr i8, ptr %p, <vscale x 2 x i64> undef
11  ret <vscale x 2 x ptr> %gep
12}
13
14; This test serves to verify code changes for "GEP.getNumIndices() == 1".
15define ptr @gep_num_of_indices_1(ptr %p) {
16; CHECK-LABEL: @gep_num_of_indices_1(
17; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
18; CHECK-NEXT:    [[TMP2:%.*]] = shl i64 [[TMP1]], 4
19; CHECK-NEXT:    [[GEP:%.*]] = getelementptr i8, ptr [[P:%.*]], i64 [[TMP2]]
20; CHECK-NEXT:    ret ptr [[GEP]]
21;
22  %gep = getelementptr <vscale x 4 x i32>, ptr %p, i64 1
23  ret ptr %gep
24}
25
26; This test serves to verify code changes for "GEP.getNumOperands() == 2".
27define void @gep_bitcast(ptr %p) {
28; CHECK-LABEL: @gep_bitcast(
29; CHECK-NEXT:    store <vscale x 16 x i8> zeroinitializer, ptr [[P:%.*]], align 16
30; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
31; CHECK-NEXT:    [[TMP2:%.*]] = shl i64 [[TMP1]], 4
32; CHECK-NEXT:    [[GEP2:%.*]] = getelementptr i8, ptr [[P]], i64 [[TMP2]]
33; CHECK-NEXT:    store <vscale x 16 x i8> zeroinitializer, ptr [[GEP2]], align 16
34; CHECK-NEXT:    ret void
35;
36  store <vscale x 16 x i8> zeroinitializer, ptr %p
37  %gep2 = getelementptr <vscale x 16 x i8>, ptr %p, i64 1
38  store <vscale x 16 x i8> zeroinitializer, ptr %gep2
39  ret void
40}
41
42; These tests serve to verify code changes when underlying gep ptr is alloca.
43; This test is to verify 'inbounds' is added when it's valid to accumulate constant offset.
44define i32 @gep_alloca_inbounds_vscale_zero() {
45; CHECK-LABEL: @gep_alloca_inbounds_vscale_zero(
46; CHECK-NEXT:    [[A:%.*]] = alloca <vscale x 4 x i32>, align 16
47; CHECK-NEXT:    [[TMP:%.*]] = getelementptr inbounds nuw i8, ptr [[A]], i64 8
48; CHECK-NEXT:    [[LOAD:%.*]] = load i32, ptr [[TMP]], align 4
49; CHECK-NEXT:    ret i32 [[LOAD]]
50;
51  %a = alloca <vscale x 4 x i32>
52  %tmp = getelementptr <vscale x 4 x i32>, ptr %a, i32 0, i32 2
53  %load = load i32, ptr %tmp
54  ret i32 %load
55}
56
57; This test is to verify 'inbounds' is not added when a constant offset can not be determined at compile-time.
58define i32 @gep_alloca_inbounds_vscale_nonzero() {
59; CHECK-LABEL: @gep_alloca_inbounds_vscale_nonzero(
60; CHECK-NEXT:    [[A:%.*]] = alloca <vscale x 4 x i32>, align 16
61; CHECK-NEXT:    [[TMP1:%.*]] = call i64 @llvm.vscale.i64()
62; CHECK-NEXT:    [[TMP2:%.*]] = shl i64 [[TMP1]], 4
63; CHECK-NEXT:    [[TMP_OFFS:%.*]] = or disjoint i64 [[TMP2]], 8
64; CHECK-NEXT:    [[TMP:%.*]] = getelementptr i8, ptr [[A]], i64 [[TMP_OFFS]]
65; CHECK-NEXT:    [[LOAD:%.*]] = load i32, ptr [[TMP]], align 4
66; CHECK-NEXT:    ret i32 [[LOAD]]
67;
68  %a = alloca <vscale x 4 x i32>
69  %tmp = getelementptr <vscale x 4 x i32>, ptr %a, i32 1, i32 2
70  %load = load i32, ptr %tmp
71  ret i32 %load
72}
73