xref: /llvm-project/llvm/test/CodeGen/AArch64/machine-combiner-subadd.ll (revision db158c7c830807caeeb0691739c41f1d522029e9)
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=aarch64-linux-gnu %s -o - | FileCheck %s
3
4; The test cases in this file check following transformation if the right form
5; can reduce latency.
6;     A - (B + C)  ==>   (A - B) - C
7
8; 32 bit version.
9define i32 @test1(i32 %a, i32 %b, i32 %c) {
10; CHECK-LABEL: test1:
11; CHECK:       // %bb.0: // %entry
12; CHECK-NEXT:    add w8, w0, #100
13; CHECK-NEXT:    orr w9, w2, #0x80
14; CHECK-NEXT:    eor w10, w1, w8, lsl #8
15; CHECK-NEXT:    sub w8, w9, w8
16; CHECK-NEXT:    sub w8, w8, w10
17; CHECK-NEXT:    eor w0, w8, w10, asr #13
18; CHECK-NEXT:    ret
19entry:
20  %c1  = or  i32 %c, 128
21  %a1  = add i32 %a, 100
22  %shl = shl i32 %a1, 8
23  %xor = xor i32 %shl, %b
24  %add = add i32 %xor, %a1
25  %sub = sub i32 %c1, %add
26  %shr = ashr i32 %xor, 13
27  %xor2 = xor i32 %sub, %shr
28  ret i32 %xor2
29}
30
31; 64 bit version.
32define i64 @test2(i64 %a, i64 %b, i64 %c) {
33; CHECK-LABEL: test2:
34; CHECK:       // %bb.0: // %entry
35; CHECK-NEXT:    add x8, x0, #100
36; CHECK-NEXT:    orr x9, x2, #0x80
37; CHECK-NEXT:    eor x10, x1, x8, lsl #8
38; CHECK-NEXT:    sub x8, x9, x8
39; CHECK-NEXT:    sub x8, x8, x10
40; CHECK-NEXT:    eor x0, x8, x10, asr #13
41; CHECK-NEXT:    ret
42entry:
43  %c1  = or  i64 %c, 128
44  %a1  = add i64 %a, 100
45  %shl = shl i64 %a1, 8
46  %xor = xor i64 %shl, %b
47  %add = add i64 %xor, %a1
48  %sub = sub i64 %c1, %add
49  %shr = ashr i64 %xor, 13
50  %xor2 = xor i64 %sub, %shr
51  ret i64 %xor2
52}
53
54; Negative test. The right form can't reduce latency.
55define i32 @test3(i32 %a, i32 %b, i32 %c) {
56; CHECK-LABEL: test3:
57; CHECK:       // %bb.0: // %entry
58; CHECK-NEXT:    add w8, w0, #100
59; CHECK-NEXT:    orr w9, w2, #0x80
60; CHECK-NEXT:    eor w10, w1, w8, lsl #8
61; CHECK-NEXT:    add w8, w9, w8
62; CHECK-NEXT:    sub w8, w10, w8
63; CHECK-NEXT:    eor w0, w8, w10, asr #13
64; CHECK-NEXT:    ret
65entry:
66  %c1  = or  i32 %c, 128
67  %a1  = add i32 %a, 100
68  %shl = shl i32 %a1, 8
69  %xor = xor i32 %shl, %b
70  %add = add i32 %c1, %a1
71  %sub = sub i32 %xor, %add
72  %shr = ashr i32 %xor, 13
73  %xor2 = xor i32 %sub, %shr
74  ret i32 %xor2
75}
76
77