1; RUN: llc < %s -asm-verbose=false -verify-machineinstrs -disable-wasm-fallthrough-return-opt -wasm-disable-explicit-locals -wasm-keep-registers -mattr=+simd128 | FileCheck %s
2
3; Test that SIMD shifts can be lowered correctly even with shift
4; values that are more complex than plain splats.
5
6target datalayout = "e-m:e-p:32:32-i64:64-n32:64-S128"
7target triple = "wasm32-unknown-unknown"
8
9;; TODO: Optimize this further by scalarizing the add
10
11; CHECK-LABEL: shl_add:
12; CHECK-NEXT: .functype shl_add (v128, i32, i32) -> (v128)
13; CHECK-NEXT: i8x16.splat $push1=, $1
14; CHECK-NEXT: i8x16.splat $push0=, $2
15; CHECK-NEXT: i8x16.add $push2=, $pop1, $pop0
16; CHECK-NEXT: i8x16.extract_lane_u $push3=, $pop2, 0
17; CHECK-NEXT: i8x16.shl $push4=, $0, $pop3
18; CHECK-NEXT: return $pop4
19define <16 x i8> @shl_add(<16 x i8> %v, i8 %a, i8 %b) {
20  %t1 = insertelement <16 x i8> undef, i8 %a, i32 0
21  %va = shufflevector <16 x i8> %t1, <16 x i8> undef, <16 x i32> zeroinitializer
22  %t2 = insertelement <16 x i8> undef, i8 %b, i32 0
23  %vb = shufflevector <16 x i8> %t2, <16 x i8> undef, <16 x i32> zeroinitializer
24  %shift = add <16 x i8> %va, %vb
25  %r = shl <16 x i8> %v, %shift
26  ret <16 x i8> %r
27}
28
29; CHECK-LABEL: shl_abs:
30; CHECK-NEXT: .functype shl_abs (v128, i32) -> (v128)
31; CHECK-NEXT: i8x16.extract_lane_u $push8=, $0, 0
32; CHECK-NEXT: i8x16.splat $push0=, $1
33; CHECK-NEXT: i8x16.abs $push98=, $pop0
34; CHECK-NEXT: local.tee $push97=, $2=, $pop98
35; CHECK-NEXT: i8x16.extract_lane_u $push6=, $pop97, 0
36; CHECK-NEXT: i32.const $push2=, 7
37; CHECK-NEXT: i32.and $push7=, $pop6, $pop2
38; CHECK-NEXT: i32.shl $push9=, $pop8, $pop7
39; CHECK-NEXT: i8x16.splat $push10=, $pop9
40; CHECK-NEXT: i8x16.extract_lane_u $push4=, $0, 1
41; CHECK-NEXT: i8x16.extract_lane_u $push1=, $2, 1
42; CHECK-NEXT: i32.const $push96=, 7
43; CHECK-NEXT: i32.and $push3=, $pop1, $pop96
44; CHECK-NEXT: i32.shl $push5=, $pop4, $pop3
45; CHECK-NEXT: i8x16.replace_lane $push11=, $pop10, 1, $pop5
46; ...
47; CHECK:      i8x16.extract_lane_u $push79=, $0, 15
48; CHECK-NEXT: i8x16.extract_lane_u $push77=, $2, 15
49; CHECK-NEXT: i32.const $push82=, 7
50; CHECK-NEXT: i32.and $push78=, $pop77, $pop82
51; CHECK-NEXT: i32.shl $push80=, $pop79, $pop78
52; CHECK-NEXT: i8x16.replace_lane $push81=, $pop76, 15, $pop80
53; CHECK-NEXT: return $pop81
54define <16 x i8> @shl_abs(<16 x i8> %v, i8 %a) {
55  %t1 = insertelement <16 x i8> undef, i8 %a, i32 0
56  %va = shufflevector <16 x i8> %t1, <16 x i8> undef, <16 x i32> zeroinitializer
57  %nva = sub <16 x i8> zeroinitializer, %va
58  %c = icmp sgt <16 x i8> %va, zeroinitializer
59  %shift = select <16 x i1> %c, <16 x i8> %va, <16 x i8> %nva
60  %r = shl <16 x i8> %v, %shift
61  ret <16 x i8> %r
62}
63
64; CHECK-LABEL: shl_abs_add:
65; CHECK-NEXT: .functype shl_abs_add (v128, i32, i32) -> (v128)
66; CHECK-NEXT: i8x16.extract_lane_u $push11=, $0, 0
67; CHECK-NEXT: i8x16.splat $push1=, $1
68; CHECK-NEXT: i8x16.splat $push0=, $2
69; CHECK-NEXT: i8x16.add $push2=, $pop1, $pop0
70; CHECK-NEXT: v8x16.shuffle $push3=, $pop2, $0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
71; CHECK-NEXT: i8x16.abs $push101=, $pop3
72; CHECK-NEXT: local.tee $push100=, $3=, $pop101
73; CHECK-NEXT: i8x16.extract_lane_u $push9=, $pop100, 0
74; CHECK-NEXT: i32.const $push5=, 7
75; CHECK-NEXT: i32.and $push10=, $pop9, $pop5
76; CHECK-NEXT: i32.shl $push12=, $pop11, $pop10
77; CHECK-NEXT: i8x16.splat $push13=, $pop12
78; CHECK-NEXT: i8x16.extract_lane_u $push7=, $0, 1
79; CHECK-NEXT: i8x16.extract_lane_u $push4=, $3, 1
80; CHECK-NEXT: i32.const $push99=, 7
81; CHECK-NEXT: i32.and $push6=, $pop4, $pop99
82; CHECK-NEXT: i32.shl $push8=, $pop7, $pop6
83; CHECK-NEXT: i8x16.replace_lane $push14=, $pop13, 1, $pop8
84; ...
85; CHECK:      i8x16.extract_lane_u $push82=, $0, 15
86; CHECK-NEXT: i8x16.extract_lane_u $push80=, $3, 15
87; CHECK-NEXT: i32.const $push85=, 7
88; CHECK-NEXT: i32.and $push81=, $pop80, $pop85
89; CHECK-NEXT: i32.shl $push83=, $pop82, $pop81
90; CHECK-NEXT: i8x16.replace_lane $push84=, $pop79, 15, $pop83
91; CHECK-NEXT: return $pop84
92define <16 x i8> @shl_abs_add(<16 x i8> %v, i8 %a, i8 %b) {
93  %t1 = insertelement <16 x i8> undef, i8 %a, i32 0
94  %va = shufflevector <16 x i8> %t1, <16 x i8> undef, <16 x i32> zeroinitializer
95  %t2 = insertelement <16 x i8> undef, i8 %b, i32 0
96  %vb = shufflevector <16 x i8> %t2, <16 x i8> undef, <16 x i32> zeroinitializer
97  %vadd = add <16 x i8> %va, %vb
98  %nvadd = sub <16 x i8> zeroinitializer, %vadd
99  %c = icmp sgt <16 x i8> %vadd, zeroinitializer
100  %shift = select <16 x i1> %c, <16 x i8> %vadd, <16 x i8> %nvadd
101  %r = shl <16 x i8> %v, %shift
102  ret <16 x i8> %r
103}
104