exedepsfix-broadcast.ll
6.31 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc < %s -mtriple=x86_64-apple-macosx -mattr=+avx2 -enable-no-signed-zeros-fp-math | FileCheck %s
; Check that the ExeDepsFix pass correctly fixes the domain for broadcast instructions.
; <rdar://problem/16354675>
define <4 x float> @ExeDepsFix_broadcastss(<4 x float> %arg, <4 x float> %arg2) {
; CHECK-LABEL: ExeDepsFix_broadcastss:
; CHECK: ## %bb.0:
; CHECK-NEXT: vbroadcastss {{.*}}(%rip), %xmm2
; CHECK-NEXT: vandps %xmm2, %xmm0, %xmm0
; CHECK-NEXT: vmaxps %xmm1, %xmm0, %xmm0
; CHECK-NEXT: retq
%bitcast = bitcast <4 x float> %arg to <4 x i32>
%and = and <4 x i32> %bitcast, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>
%floatcast = bitcast <4 x i32> %and to <4 x float>
%max_is_x = fcmp oge <4 x float> %floatcast, %arg2
%max = select <4 x i1> %max_is_x, <4 x float> %floatcast, <4 x float> %arg2
ret <4 x float> %max
}
define <8 x float> @ExeDepsFix_broadcastss256(<8 x float> %arg, <8 x float> %arg2) {
; CHECK-LABEL: ExeDepsFix_broadcastss256:
; CHECK: ## %bb.0:
; CHECK-NEXT: vbroadcastss {{.*}}(%rip), %ymm2
; CHECK-NEXT: vandps %ymm2, %ymm0, %ymm0
; CHECK-NEXT: vmaxps %ymm1, %ymm0, %ymm0
; CHECK-NEXT: retq
%bitcast = bitcast <8 x float> %arg to <8 x i32>
%and = and <8 x i32> %bitcast, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>
%floatcast = bitcast <8 x i32> %and to <8 x float>
%max_is_x = fcmp oge <8 x float> %floatcast, %arg2
%max = select <8 x i1> %max_is_x, <8 x float> %floatcast, <8 x float> %arg2
ret <8 x float> %max
}
define <4 x float> @ExeDepsFix_broadcastss_inreg(<4 x float> %arg, <4 x float> %arg2, i32 %broadcastvalue) {
; CHECK-LABEL: ExeDepsFix_broadcastss_inreg:
; CHECK: ## %bb.0:
; CHECK-NEXT: vmovd %edi, %xmm2
; CHECK-NEXT: vpbroadcastd %xmm2, %xmm2
; CHECK-NEXT: vpand %xmm2, %xmm0, %xmm0
; CHECK-NEXT: vmaxps %xmm1, %xmm0, %xmm0
; CHECK-NEXT: retq
%bitcast = bitcast <4 x float> %arg to <4 x i32>
%in = insertelement <4 x i32> undef, i32 %broadcastvalue, i32 0
%mask = shufflevector <4 x i32> %in, <4 x i32> undef, <4 x i32> zeroinitializer
%and = and <4 x i32> %bitcast, %mask
%floatcast = bitcast <4 x i32> %and to <4 x float>
%max_is_x = fcmp oge <4 x float> %floatcast, %arg2
%max = select <4 x i1> %max_is_x, <4 x float> %floatcast, <4 x float> %arg2
ret <4 x float> %max
}
define <8 x float> @ExeDepsFix_broadcastss256_inreg(<8 x float> %arg, <8 x float> %arg2, i32 %broadcastvalue) {
; CHECK-LABEL: ExeDepsFix_broadcastss256_inreg:
; CHECK: ## %bb.0:
; CHECK-NEXT: vmovd %edi, %xmm2
; CHECK-NEXT: vpbroadcastd %xmm2, %ymm2
; CHECK-NEXT: vpand %ymm2, %ymm0, %ymm0
; CHECK-NEXT: vmaxps %ymm1, %ymm0, %ymm0
; CHECK-NEXT: retq
%bitcast = bitcast <8 x float> %arg to <8 x i32>
%in = insertelement <8 x i32> undef, i32 %broadcastvalue, i32 0
%mask = shufflevector <8 x i32> %in, <8 x i32> undef, <8 x i32> zeroinitializer
%and = and <8 x i32> %bitcast, %mask
%floatcast = bitcast <8 x i32> %and to <8 x float>
%max_is_x = fcmp oge <8 x float> %floatcast, %arg2
%max = select <8 x i1> %max_is_x, <8 x float> %floatcast, <8 x float> %arg2
ret <8 x float> %max
}
; In that case the broadcast is directly folded into vandpd.
define <2 x double> @ExeDepsFix_broadcastsd(<2 x double> %arg, <2 x double> %arg2) {
; CHECK-LABEL: ExeDepsFix_broadcastsd:
; CHECK: ## %bb.0:
; CHECK-NEXT: vandpd {{.*}}(%rip), %xmm0, %xmm0
; CHECK-NEXT: vmaxpd %xmm1, %xmm0, %xmm0
; CHECK-NEXT: retq
%bitcast = bitcast <2 x double> %arg to <2 x i64>
%and = and <2 x i64> %bitcast, <i64 2147483647, i64 2147483647>
%floatcast = bitcast <2 x i64> %and to <2 x double>
%max_is_x = fcmp oge <2 x double> %floatcast, %arg2
%max = select <2 x i1> %max_is_x, <2 x double> %floatcast, <2 x double> %arg2
ret <2 x double> %max
}
define <4 x double> @ExeDepsFix_broadcastsd256(<4 x double> %arg, <4 x double> %arg2) {
; CHECK-LABEL: ExeDepsFix_broadcastsd256:
; CHECK: ## %bb.0:
; CHECK-NEXT: vbroadcastsd {{.*}}(%rip), %ymm2
; CHECK-NEXT: vandpd %ymm2, %ymm0, %ymm0
; CHECK-NEXT: vmaxpd %ymm1, %ymm0, %ymm0
; CHECK-NEXT: retq
%bitcast = bitcast <4 x double> %arg to <4 x i64>
%and = and <4 x i64> %bitcast, <i64 2147483647, i64 2147483647, i64 2147483647, i64 2147483647>
%floatcast = bitcast <4 x i64> %and to <4 x double>
%max_is_x = fcmp oge <4 x double> %floatcast, %arg2
%max = select <4 x i1> %max_is_x, <4 x double> %floatcast, <4 x double> %arg2
ret <4 x double> %max
}
; ExeDepsFix works top down, thus it coalesces vpunpcklqdq domain with
; vpand and there is nothing more you can do to match vmaxpd.
define <2 x double> @ExeDepsFix_broadcastsd_inreg(<2 x double> %arg, <2 x double> %arg2, i64 %broadcastvalue) {
; CHECK-LABEL: ExeDepsFix_broadcastsd_inreg:
; CHECK: ## %bb.0:
; CHECK-NEXT: vmovq %rdi, %xmm2
; CHECK-NEXT: vpbroadcastq %xmm2, %xmm2
; CHECK-NEXT: vpand %xmm2, %xmm0, %xmm0
; CHECK-NEXT: vmaxpd %xmm1, %xmm0, %xmm0
; CHECK-NEXT: retq
%bitcast = bitcast <2 x double> %arg to <2 x i64>
%in = insertelement <2 x i64> undef, i64 %broadcastvalue, i32 0
%mask = shufflevector <2 x i64> %in, <2 x i64> undef, <2 x i32> zeroinitializer
%and = and <2 x i64> %bitcast, %mask
%floatcast = bitcast <2 x i64> %and to <2 x double>
%max_is_x = fcmp oge <2 x double> %floatcast, %arg2
%max = select <2 x i1> %max_is_x, <2 x double> %floatcast, <2 x double> %arg2
ret <2 x double> %max
}
define <4 x double> @ExeDepsFix_broadcastsd256_inreg(<4 x double> %arg, <4 x double> %arg2, i64 %broadcastvalue) {
; CHECK-LABEL: ExeDepsFix_broadcastsd256_inreg:
; CHECK: ## %bb.0:
; CHECK-NEXT: vmovq %rdi, %xmm2
; CHECK-NEXT: vpbroadcastq %xmm2, %ymm2
; CHECK-NEXT: vpand %ymm2, %ymm0, %ymm0
; CHECK-NEXT: vmaxpd %ymm1, %ymm0, %ymm0
; CHECK-NEXT: retq
%bitcast = bitcast <4 x double> %arg to <4 x i64>
%in = insertelement <4 x i64> undef, i64 %broadcastvalue, i32 0
%mask = shufflevector <4 x i64> %in, <4 x i64> undef, <4 x i32> zeroinitializer
%and = and <4 x i64> %bitcast, %mask
%floatcast = bitcast <4 x i64> %and to <4 x double>
%max_is_x = fcmp oge <4 x double> %floatcast, %arg2
%max = select <4 x i1> %max_is_x, <4 x double> %floatcast, <4 x double> %arg2
ret <4 x double> %max
}