| blob | a173fd40feb280a28db4d72c4761afc59d8c3365 |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=aarch64 -mattr=+sme < %s | FileCheck %s
4 declare void @private_za_callee()
5 declare void @private_za_preserved_callee() "aarch64_pstate_za_preserved"
6 declare float @llvm.cos.f32(float)
8 ; Test lazy-save mechanism for a single callee.
9 define void @test_lazy_save_1_callee() nounwind "aarch64_pstate_za_shared" {
10 ; CHECK-LABEL: test_lazy_save_1_callee:
11 ; CHECK: // %bb.0:
12 ; CHECK-NEXT: stp x29, x30, [sp, #-16]! // 16-byte Folded Spill
13 ; CHECK-NEXT: mov x29, sp
14 ; CHECK-NEXT: sub sp, sp, #16
15 ; CHECK-NEXT: rdsvl x8, #1
16 ; CHECK-NEXT: mov x9, sp
17 ; CHECK-NEXT: msub x9, x8, x8, x9
18 ; CHECK-NEXT: mov sp, x9
19 ; CHECK-NEXT: sub x10, x29, #16
20 ; CHECK-NEXT: stur wzr, [x29, #-4]
21 ; CHECK-NEXT: sturh wzr, [x29, #-6]
22 ; CHECK-NEXT: stur x9, [x29, #-16]
23 ; CHECK-NEXT: sturh w8, [x29, #-8]
24 ; CHECK-NEXT: msr TPIDR2_EL0, x10
25 ; CHECK-NEXT: bl private_za_callee
26 ; CHECK-NEXT: smstart za
27 ; CHECK-NEXT: mrs x8, TPIDR2_EL0
28 ; CHECK-NEXT: sub x0, x29, #16
29 ; CHECK-NEXT: cbnz x8, .LBB0_2
30 ; CHECK-NEXT: // %bb.1:
31 ; CHECK-NEXT: bl __arm_tpidr2_restore
32 ; CHECK-NEXT: .LBB0_2:
33 ; CHECK-NEXT: msr TPIDR2_EL0, xzr
34 ; CHECK-NEXT: mov sp, x29
35 ; CHECK-NEXT: ldp x29, x30, [sp], #16 // 16-byte Folded Reload
36 ; CHECK-NEXT: ret
37 call void @private_za_callee()
38 ret void
39 }
41 ; Test lazy-save mechanism for multiple callees.
42 define void @test_lazy_save_2_callees() nounwind "aarch64_pstate_za_shared" {
43 ; CHECK-LABEL: test_lazy_save_2_callees:
44 ; CHECK: // %bb.0:
45 ; CHECK-NEXT: stp x29, x30, [sp, #-32]! // 16-byte Folded Spill
46 ; CHECK-NEXT: stp x20, x19, [sp, #16] // 16-byte Folded Spill
47 ; CHECK-NEXT: mov x29, sp
48 ; CHECK-NEXT: sub sp, sp, #16
49 ; CHECK-NEXT: rdsvl x19, #1
50 ; CHECK-NEXT: mov x8, sp
51 ; CHECK-NEXT: msub x8, x19, x19, x8
52 ; CHECK-NEXT: mov sp, x8
53 ; CHECK-NEXT: sub x20, x29, #16
54 ; CHECK-NEXT: stur wzr, [x29, #-4]
55 ; CHECK-NEXT: sturh wzr, [x29, #-6]
56 ; CHECK-NEXT: stur x8, [x29, #-16]
57 ; CHECK-NEXT: sturh w19, [x29, #-8]
58 ; CHECK-NEXT: msr TPIDR2_EL0, x20
59 ; CHECK-NEXT: bl private_za_callee
60 ; CHECK-NEXT: smstart za
61 ; CHECK-NEXT: mrs x8, TPIDR2_EL0
62 ; CHECK-NEXT: sub x0, x29, #16
63 ; CHECK-NEXT: cbnz x8, .LBB1_2
64 ; CHECK-NEXT: // %bb.1:
65 ; CHECK-NEXT: bl __arm_tpidr2_restore
66 ; CHECK-NEXT: .LBB1_2:
67 ; CHECK-NEXT: msr TPIDR2_EL0, xzr
68 ; CHECK-NEXT: sturh w19, [x29, #-8]
69 ; CHECK-NEXT: msr TPIDR2_EL0, x20
70 ; CHECK-NEXT: bl private_za_callee
71 ; CHECK-NEXT: smstart za
72 ; CHECK-NEXT: mrs x8, TPIDR2_EL0
73 ; CHECK-NEXT: sub x0, x29, #16
74 ; CHECK-NEXT: cbnz x8, .LBB1_4
75 ; CHECK-NEXT: // %bb.3:
76 ; CHECK-NEXT: bl __arm_tpidr2_restore
77 ; CHECK-NEXT: .LBB1_4:
78 ; CHECK-NEXT: msr TPIDR2_EL0, xzr
79 ; CHECK-NEXT: mov sp, x29
80 ; CHECK-NEXT: ldp x20, x19, [sp, #16] // 16-byte Folded Reload
81 ; CHECK-NEXT: ldp x29, x30, [sp], #32 // 16-byte Folded Reload
82 ; CHECK-NEXT: ret
83 call void @private_za_callee()
84 call void @private_za_callee()
85 ret void
86 }
88 ; Test a call of an intrinsic that gets expanded to a library call.
89 define float @test_lazy_save_expanded_intrinsic(float %a) nounwind "aarch64_pstate_za_shared" {
90 ; CHECK-LABEL: test_lazy_save_expanded_intrinsic:
91 ; CHECK: // %bb.0:
92 ; CHECK-NEXT: stp x29, x30, [sp, #-16]! // 16-byte Folded Spill
93 ; CHECK-NEXT: mov x29, sp
94 ; CHECK-NEXT: sub sp, sp, #16
95 ; CHECK-NEXT: rdsvl x8, #1
96 ; CHECK-NEXT: mov x9, sp
97 ; CHECK-NEXT: msub x9, x8, x8, x9
98 ; CHECK-NEXT: mov sp, x9
99 ; CHECK-NEXT: sub x10, x29, #16
100 ; CHECK-NEXT: stur wzr, [x29, #-4]
101 ; CHECK-NEXT: sturh wzr, [x29, #-6]
102 ; CHECK-NEXT: stur x9, [x29, #-16]
103 ; CHECK-NEXT: sturh w8, [x29, #-8]
104 ; CHECK-NEXT: msr TPIDR2_EL0, x10
105 ; CHECK-NEXT: bl cosf
106 ; CHECK-NEXT: smstart za
107 ; CHECK-NEXT: mrs x8, TPIDR2_EL0
108 ; CHECK-NEXT: sub x0, x29, #16
109 ; CHECK-NEXT: cbnz x8, .LBB2_2
110 ; CHECK-NEXT: // %bb.1:
111 ; CHECK-NEXT: bl __arm_tpidr2_restore
112 ; CHECK-NEXT: .LBB2_2:
113 ; CHECK-NEXT: msr TPIDR2_EL0, xzr
114 ; CHECK-NEXT: mov sp, x29
115 ; CHECK-NEXT: ldp x29, x30, [sp], #16 // 16-byte Folded Reload
116 ; CHECK-NEXT: ret
117 %res = call float @llvm.cos.f32(float %a)
118 ret float %res
119 }
121 ; Test a combination of streaming-compatible -> normal call with lazy-save.
122 define void @test_lazy_save_and_conditional_smstart() nounwind "aarch64_pstate_za_shared" "aarch64_pstate_sm_compatible" {
123 ; CHECK-LABEL: test_lazy_save_and_conditional_smstart:
124 ; CHECK: // %bb.0:
125 ; CHECK-NEXT: stp d15, d14, [sp, #-96]! // 16-byte Folded Spill
126 ; CHECK-NEXT: stp d13, d12, [sp, #16] // 16-byte Folded Spill
127 ; CHECK-NEXT: stp d11, d10, [sp, #32] // 16-byte Folded Spill
128 ; CHECK-NEXT: stp d9, d8, [sp, #48] // 16-byte Folded Spill
129 ; CHECK-NEXT: stp x29, x30, [sp, #64] // 16-byte Folded Spill
130 ; CHECK-NEXT: add x29, sp, #64
131 ; CHECK-NEXT: str x19, [sp, #80] // 8-byte Folded Spill
132 ; CHECK-NEXT: sub sp, sp, #16
133 ; CHECK-NEXT: rdsvl x8, #1
134 ; CHECK-NEXT: mov x9, sp
135 ; CHECK-NEXT: msub x9, x8, x8, x9
136 ; CHECK-NEXT: mov sp, x9
137 ; CHECK-NEXT: sub x10, x29, #80
138 ; CHECK-NEXT: stur wzr, [x29, #-68]
139 ; CHECK-NEXT: sturh wzr, [x29, #-70]
140 ; CHECK-NEXT: stur x9, [x29, #-80]
141 ; CHECK-NEXT: sturh w8, [x29, #-72]
142 ; CHECK-NEXT: msr TPIDR2_EL0, x10
143 ; CHECK-NEXT: bl __arm_sme_state
144 ; CHECK-NEXT: and x19, x0, #0x1
145 ; CHECK-NEXT: tbz w19, #0, .LBB3_2
146 ; CHECK-NEXT: // %bb.1:
147 ; CHECK-NEXT: smstop sm
148 ; CHECK-NEXT: .LBB3_2:
149 ; CHECK-NEXT: bl private_za_callee
150 ; CHECK-NEXT: tbz w19, #0, .LBB3_4
151 ; CHECK-NEXT: // %bb.3:
152 ; CHECK-NEXT: smstart sm
153 ; CHECK-NEXT: .LBB3_4:
154 ; CHECK-NEXT: smstart za
155 ; CHECK-NEXT: mrs x8, TPIDR2_EL0
156 ; CHECK-NEXT: sub x0, x29, #80
157 ; CHECK-NEXT: cbnz x8, .LBB3_6
158 ; CHECK-NEXT: // %bb.5:
159 ; CHECK-NEXT: bl __arm_tpidr2_restore
160 ; CHECK-NEXT: .LBB3_6:
161 ; CHECK-NEXT: msr TPIDR2_EL0, xzr
162 ; CHECK-NEXT: sub sp, x29, #64
163 ; CHECK-NEXT: ldp x29, x30, [sp, #64] // 16-byte Folded Reload
164 ; CHECK-NEXT: ldr x19, [sp, #80] // 8-byte Folded Reload
165 ; CHECK-NEXT: ldp d9, d8, [sp, #48] // 16-byte Folded Reload
166 ; CHECK-NEXT: ldp d11, d10, [sp, #32] // 16-byte Folded Reload
167 ; CHECK-NEXT: ldp d13, d12, [sp, #16] // 16-byte Folded Reload
168 ; CHECK-NEXT: ldp d15, d14, [sp], #96 // 16-byte Folded Reload
169 ; CHECK-NEXT: ret
170 call void @private_za_callee()
171 ret void
172 }
175 ; Test lazy-save mechanism for an aarch64_pstate_za_shared caller
176 ; calling a callee with aarch64_pstate_za_preserved.
177 define void @za_shared_caller_za_preserved_callee() nounwind "aarch64_pstate_za_shared" "aarch64_pstate_sm_compatible" {
178 ; CHECK-LABEL: za_shared_caller_za_preserved_callee:
179 ; CHECK: // %bb.0:
180 ; CHECK-NEXT: stp d15, d14, [sp, #-96]! // 16-byte Folded Spill
181 ; CHECK-NEXT: stp d13, d12, [sp, #16] // 16-byte Folded Spill
182 ; CHECK-NEXT: stp d11, d10, [sp, #32] // 16-byte Folded Spill
183 ; CHECK-NEXT: stp d9, d8, [sp, #48] // 16-byte Folded Spill
184 ; CHECK-NEXT: stp x29, x30, [sp, #64] // 16-byte Folded Spill
185 ; CHECK-NEXT: add x29, sp, #64
186 ; CHECK-NEXT: str x19, [sp, #80] // 8-byte Folded Spill
187 ; CHECK-NEXT: sub sp, sp, #16
188 ; CHECK-NEXT: rdsvl x8, #1
189 ; CHECK-NEXT: mov x9, sp
190 ; CHECK-NEXT: msub x8, x8, x8, x9
191 ; CHECK-NEXT: mov sp, x8
192 ; CHECK-NEXT: sub x9, x29, #80
193 ; CHECK-NEXT: stp x8, xzr, [x29, #-80]
194 ; CHECK-NEXT: msr TPIDR2_EL0, x9
195 ; CHECK-NEXT: bl __arm_sme_state
196 ; CHECK-NEXT: and x19, x0, #0x1
197 ; CHECK-NEXT: tbz w19, #0, .LBB4_2
198 ; CHECK-NEXT: // %bb.1:
199 ; CHECK-NEXT: smstop sm
200 ; CHECK-NEXT: .LBB4_2:
201 ; CHECK-NEXT: bl private_za_preserved_callee
202 ; CHECK-NEXT: tbz w19, #0, .LBB4_4
203 ; CHECK-NEXT: // %bb.3:
204 ; CHECK-NEXT: smstart sm
205 ; CHECK-NEXT: .LBB4_4:
206 ; CHECK-NEXT: msr TPIDR2_EL0, xzr
207 ; CHECK-NEXT: sub sp, x29, #64
208 ; CHECK-NEXT: ldp x29, x30, [sp, #64] // 16-byte Folded Reload
209 ; CHECK-NEXT: ldr x19, [sp, #80] // 8-byte Folded Reload
210 ; CHECK-NEXT: ldp d9, d8, [sp, #48] // 16-byte Folded Reload
211 ; CHECK-NEXT: ldp d11, d10, [sp, #32] // 16-byte Folded Reload
212 ; CHECK-NEXT: ldp d13, d12, [sp, #16] // 16-byte Folded Reload
213 ; CHECK-NEXT: ldp d15, d14, [sp], #96 // 16-byte Folded Reload
214 ; CHECK-NEXT: ret
215 call void @private_za_preserved_callee()
216 ret void
217 }