xref: /llvm-project/llvm/test/Transforms/SLPVectorizer/X86/gep.ll (revision 38fffa630ee80163dc65e759392ad29798905679)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s -passes=slp-vectorizer -S |FileCheck %s
3; RUN: opt < %s -aa-pipeline=basic-aa -passes=slp-vectorizer -S |FileCheck %s
4target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
5target triple = "x86_64-unknown-unknown"
6
7; Test if SLP can handle GEP expressions.
8; The test perform the following action:
9;   x->first  = y->first  + 16
10;   x->second = y->second + 16
11
12define void @foo1 (ptr noalias %x, ptr noalias %y) {
13; CHECK-LABEL: @foo1(
14; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds { ptr, ptr }, ptr [[Y:%.*]], i64 0, i32 0
15; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds { ptr, ptr }, ptr [[X:%.*]], i64 0, i32 0
16; CHECK-NEXT:    [[TMP4:%.*]] = load <2 x ptr>, ptr [[TMP1]], align 8
17; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr i32, <2 x ptr> [[TMP4]], <2 x i64> splat (i64 16)
18; CHECK-NEXT:    store <2 x ptr> [[TMP5]], ptr [[TMP2]], align 8
19; CHECK-NEXT:    ret void
20;
21  %1 = getelementptr inbounds { ptr, ptr }, ptr %y, i64 0, i32 0
22  %2 = load ptr, ptr %1, align 8
23  %3 = getelementptr inbounds i32, ptr %2, i64 16
24  %4 = getelementptr inbounds { ptr, ptr }, ptr %x, i64 0, i32 0
25  store ptr %3, ptr %4, align 8
26  %5 = getelementptr inbounds { ptr, ptr }, ptr %y, i64 0, i32 1
27  %6 = load ptr, ptr %5, align 8
28  %7 = getelementptr inbounds i32, ptr %6, i64 16
29  %8 = getelementptr inbounds { ptr, ptr }, ptr %x, i64 0, i32 1
30  store ptr %7, ptr %8, align 8
31  ret void
32}
33
34; Test that we don't vectorize GEP expressions if indexes are not constants.
35; We can't produce an efficient code in that case.
36define void @foo2 (ptr noalias %x, ptr noalias %y, i32 %i) {
37; CHECK-LABEL: @foo2(
38; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds { ptr, ptr }, ptr [[Y:%.*]], i64 0, i32 0
39; CHECK-NEXT:    [[TMP2:%.*]] = load ptr, ptr [[TMP1]], align 8
40; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[TMP2]], i32 [[I:%.*]]
41; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds { ptr, ptr }, ptr [[X:%.*]], i64 0, i32 0
42; CHECK-NEXT:    store ptr [[TMP3]], ptr [[TMP4]], align 8
43; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds { ptr, ptr }, ptr [[Y]], i64 0, i32 1
44; CHECK-NEXT:    [[TMP6:%.*]] = load ptr, ptr [[TMP5]], align 8
45; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i32, ptr [[TMP6]], i32 [[I]]
46; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds { ptr, ptr }, ptr [[X]], i64 0, i32 1
47; CHECK-NEXT:    store ptr [[TMP7]], ptr [[TMP8]], align 8
48; CHECK-NEXT:    ret void
49;
50  %1 = getelementptr inbounds { ptr, ptr }, ptr %y, i64 0, i32 0
51  %2 = load ptr, ptr %1, align 8
52  %3 = getelementptr inbounds i32, ptr %2, i32 %i
53  %4 = getelementptr inbounds { ptr, ptr }, ptr %x, i64 0, i32 0
54  store ptr %3, ptr %4, align 8
55  %5 = getelementptr inbounds { ptr, ptr }, ptr %y, i64 0, i32 1
56  %6 = load ptr, ptr %5, align 8
57  %7 = getelementptr inbounds i32, ptr %6, i32 %i
58  %8 = getelementptr inbounds { ptr, ptr }, ptr %x, i64 0, i32 1
59  store ptr %7, ptr %8, align 8
60  ret void
61}
62