| blob | 2856cfa01fad15defe3feec8368dd2993edfc36b |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc < %s -mtriple=x86_64-- | FileCheck %s
4 ;
5 ; 32-bit float to signed integer
6 ;
8 declare <4 x i1> @llvm.fptosi.sat.v4i1.v4f32(<4 x float>)
9 declare <4 x i8> @llvm.fptosi.sat.v4i8.v4f32(<4 x float>)
10 declare <4 x i16> @llvm.fptosi.sat.v4i16.v4f32(<4 x float>)
11 declare <4 x i32> @llvm.fptosi.sat.v4i32.v4f32(<4 x float>)
12 declare <4 x i64> @llvm.fptosi.sat.v4i64.v4f32(<4 x float>)
13 declare <4 x i128> @llvm.fptosi.sat.v4i128.v4f32(<4 x float>)
15 define <4 x i1> @test_signed_v4i1_v4f32(<4 x float> %f) nounwind {
16 ; CHECK-LABEL: test_signed_v4i1_v4f32:
17 ; CHECK: # %bb.0:
18 ; CHECK-NEXT: movaps %xmm0, %xmm1
19 ; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]
20 ; CHECK-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
21 ; CHECK-NEXT: xorl %eax, %eax
22 ; CHECK-NEXT: ucomiss %xmm1, %xmm1
23 ; CHECK-NEXT: maxss %xmm2, %xmm1
24 ; CHECK-NEXT: xorps %xmm3, %xmm3
25 ; CHECK-NEXT: minss %xmm3, %xmm1
26 ; CHECK-NEXT: cvttss2si %xmm1, %ecx
27 ; CHECK-NEXT: cmovpl %eax, %ecx
28 ; CHECK-NEXT: movd %ecx, %xmm1
29 ; CHECK-NEXT: movaps %xmm0, %xmm4
30 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm4 = xmm4[1],xmm0[1]
31 ; CHECK-NEXT: ucomiss %xmm4, %xmm4
32 ; CHECK-NEXT: maxss %xmm2, %xmm4
33 ; CHECK-NEXT: minss %xmm3, %xmm4
34 ; CHECK-NEXT: cvttss2si %xmm4, %ecx
35 ; CHECK-NEXT: cmovpl %eax, %ecx
36 ; CHECK-NEXT: movd %ecx, %xmm4
37 ; CHECK-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm1[0],xmm4[1],xmm1[1]
38 ; CHECK-NEXT: movaps %xmm0, %xmm1
39 ; CHECK-NEXT: maxss %xmm2, %xmm1
40 ; CHECK-NEXT: minss %xmm3, %xmm1
41 ; CHECK-NEXT: cvttss2si %xmm1, %ecx
42 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
43 ; CHECK-NEXT: cmovpl %eax, %ecx
44 ; CHECK-NEXT: movd %ecx, %xmm1
45 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
46 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
47 ; CHECK-NEXT: maxss %xmm2, %xmm0
48 ; CHECK-NEXT: minss %xmm3, %xmm0
49 ; CHECK-NEXT: cvttss2si %xmm0, %ecx
50 ; CHECK-NEXT: cmovpl %eax, %ecx
51 ; CHECK-NEXT: movd %ecx, %xmm0
52 ; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
53 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm4[0]
54 ; CHECK-NEXT: movdqa %xmm1, %xmm0
55 ; CHECK-NEXT: retq
56 %x = call <4 x i1> @llvm.fptosi.sat.v4i1.v4f32(<4 x float> %f)
57 ret <4 x i1> %x
58 }
60 define <4 x i8> @test_signed_v4i8_v4f32(<4 x float> %f) nounwind {
61 ; CHECK-LABEL: test_signed_v4i8_v4f32:
62 ; CHECK: # %bb.0:
63 ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
64 ; CHECK-NEXT: movaps %xmm1, %xmm3
65 ; CHECK-NEXT: maxss %xmm0, %xmm3
66 ; CHECK-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
67 ; CHECK-NEXT: movaps %xmm2, %xmm4
68 ; CHECK-NEXT: minss %xmm3, %xmm4
69 ; CHECK-NEXT: cvttss2si %xmm4, %eax
70 ; CHECK-NEXT: movzbl %al, %eax
71 ; CHECK-NEXT: movaps %xmm0, %xmm3
72 ; CHECK-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[1,1]
73 ; CHECK-NEXT: movaps %xmm1, %xmm4
74 ; CHECK-NEXT: maxss %xmm3, %xmm4
75 ; CHECK-NEXT: movaps %xmm2, %xmm3
76 ; CHECK-NEXT: minss %xmm4, %xmm3
77 ; CHECK-NEXT: cvttss2si %xmm3, %ecx
78 ; CHECK-NEXT: movzbl %cl, %ecx
79 ; CHECK-NEXT: shll $8, %ecx
80 ; CHECK-NEXT: orl %eax, %ecx
81 ; CHECK-NEXT: movaps %xmm0, %xmm3
82 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
83 ; CHECK-NEXT: movaps %xmm1, %xmm4
84 ; CHECK-NEXT: maxss %xmm3, %xmm4
85 ; CHECK-NEXT: movaps %xmm2, %xmm3
86 ; CHECK-NEXT: minss %xmm4, %xmm3
87 ; CHECK-NEXT: cvttss2si %xmm3, %eax
88 ; CHECK-NEXT: movzbl %al, %eax
89 ; CHECK-NEXT: shll $16, %eax
90 ; CHECK-NEXT: orl %ecx, %eax
91 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
92 ; CHECK-NEXT: maxss %xmm0, %xmm1
93 ; CHECK-NEXT: minss %xmm1, %xmm2
94 ; CHECK-NEXT: cvttss2si %xmm2, %ecx
95 ; CHECK-NEXT: shll $24, %ecx
96 ; CHECK-NEXT: orl %eax, %ecx
97 ; CHECK-NEXT: movd %ecx, %xmm0
98 ; CHECK-NEXT: retq
99 %x = call <4 x i8> @llvm.fptosi.sat.v4i8.v4f32(<4 x float> %f)
100 ret <4 x i8> %x
101 }
103 define <4 x i16> @test_signed_v4i16_v4f32(<4 x float> %f) nounwind {
104 ; CHECK-LABEL: test_signed_v4i16_v4f32:
105 ; CHECK: # %bb.0:
106 ; CHECK-NEXT: movaps %xmm0, %xmm1
107 ; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[1,1],xmm0[1,1]
108 ; CHECK-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
109 ; CHECK-NEXT: movaps %xmm2, %xmm3
110 ; CHECK-NEXT: maxss %xmm1, %xmm3
111 ; CHECK-NEXT: movss {{.*#+}} xmm4 = mem[0],zero,zero,zero
112 ; CHECK-NEXT: movaps %xmm4, %xmm1
113 ; CHECK-NEXT: minss %xmm3, %xmm1
114 ; CHECK-NEXT: cvttss2si %xmm1, %eax
115 ; CHECK-NEXT: movaps %xmm2, %xmm1
116 ; CHECK-NEXT: maxss %xmm0, %xmm1
117 ; CHECK-NEXT: movaps %xmm4, %xmm3
118 ; CHECK-NEXT: minss %xmm1, %xmm3
119 ; CHECK-NEXT: cvttss2si %xmm3, %ecx
120 ; CHECK-NEXT: movd %ecx, %xmm1
121 ; CHECK-NEXT: pinsrw $1, %eax, %xmm1
122 ; CHECK-NEXT: movaps %xmm0, %xmm3
123 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
124 ; CHECK-NEXT: movaps %xmm2, %xmm5
125 ; CHECK-NEXT: maxss %xmm3, %xmm5
126 ; CHECK-NEXT: movaps %xmm4, %xmm3
127 ; CHECK-NEXT: minss %xmm5, %xmm3
128 ; CHECK-NEXT: cvttss2si %xmm3, %eax
129 ; CHECK-NEXT: pinsrw $2, %eax, %xmm1
130 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
131 ; CHECK-NEXT: maxss %xmm0, %xmm2
132 ; CHECK-NEXT: minss %xmm2, %xmm4
133 ; CHECK-NEXT: cvttss2si %xmm4, %eax
134 ; CHECK-NEXT: pinsrw $3, %eax, %xmm1
135 ; CHECK-NEXT: movdqa %xmm1, %xmm0
136 ; CHECK-NEXT: retq
137 %x = call <4 x i16> @llvm.fptosi.sat.v4i16.v4f32(<4 x float> %f)
138 ret <4 x i16> %x
139 }
141 define <4 x i32> @test_signed_v4i32_v4f32(<4 x float> %f) nounwind {
142 ; CHECK-LABEL: test_signed_v4i32_v4f32:
143 ; CHECK: # %bb.0:
144 ; CHECK-NEXT: movaps %xmm0, %xmm1
145 ; CHECK-NEXT: shufps {{.*#+}} xmm1 = xmm1[3,3],xmm0[3,3]
146 ; CHECK-NEXT: cvttss2si %xmm1, %edx
147 ; CHECK-NEXT: movss {{.*#+}} xmm2 = mem[0],zero,zero,zero
148 ; CHECK-NEXT: ucomiss %xmm2, %xmm1
149 ; CHECK-NEXT: movl $2147483647, %eax # imm = 0x7FFFFFFF
150 ; CHECK-NEXT: cmoval %eax, %edx
151 ; CHECK-NEXT: xorl %ecx, %ecx
152 ; CHECK-NEXT: ucomiss %xmm1, %xmm1
153 ; CHECK-NEXT: cmovpl %ecx, %edx
154 ; CHECK-NEXT: movd %edx, %xmm1
155 ; CHECK-NEXT: movaps %xmm0, %xmm3
156 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1],xmm0[1]
157 ; CHECK-NEXT: cvttss2si %xmm3, %edx
158 ; CHECK-NEXT: ucomiss %xmm2, %xmm3
159 ; CHECK-NEXT: cmoval %eax, %edx
160 ; CHECK-NEXT: ucomiss %xmm3, %xmm3
161 ; CHECK-NEXT: cmovpl %ecx, %edx
162 ; CHECK-NEXT: movd %edx, %xmm3
163 ; CHECK-NEXT: punpckldq {{.*#+}} xmm3 = xmm3[0],xmm1[0],xmm3[1],xmm1[1]
164 ; CHECK-NEXT: cvttss2si %xmm0, %edx
165 ; CHECK-NEXT: ucomiss %xmm2, %xmm0
166 ; CHECK-NEXT: cmoval %eax, %edx
167 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
168 ; CHECK-NEXT: cmovpl %ecx, %edx
169 ; CHECK-NEXT: movd %edx, %xmm1
170 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
171 ; CHECK-NEXT: cvttss2si %xmm0, %edx
172 ; CHECK-NEXT: ucomiss %xmm2, %xmm0
173 ; CHECK-NEXT: cmoval %eax, %edx
174 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
175 ; CHECK-NEXT: cmovpl %ecx, %edx
176 ; CHECK-NEXT: movd %edx, %xmm0
177 ; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
178 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]
179 ; CHECK-NEXT: movdqa %xmm1, %xmm0
180 ; CHECK-NEXT: retq
181 %x = call <4 x i32> @llvm.fptosi.sat.v4i32.v4f32(<4 x float> %f)
182 ret <4 x i32> %x
183 }
185 define <4 x i64> @test_signed_v4i64_v4f32(<4 x float> %f) nounwind {
186 ; CHECK-LABEL: test_signed_v4i64_v4f32:
187 ; CHECK: # %bb.0:
188 ; CHECK-NEXT: cvttss2si %xmm0, %rdx
189 ; CHECK-NEXT: movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
190 ; CHECK-NEXT: ucomiss %xmm1, %xmm0
191 ; CHECK-NEXT: movabsq $9223372036854775807, %rax # imm = 0x7FFFFFFFFFFFFFFF
192 ; CHECK-NEXT: cmovaq %rax, %rdx
193 ; CHECK-NEXT: xorl %ecx, %ecx
194 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
195 ; CHECK-NEXT: cmovpq %rcx, %rdx
196 ; CHECK-NEXT: movq %rdx, %xmm2
197 ; CHECK-NEXT: movaps %xmm0, %xmm3
198 ; CHECK-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1],xmm0[1,1]
199 ; CHECK-NEXT: cvttss2si %xmm3, %rdx
200 ; CHECK-NEXT: ucomiss %xmm1, %xmm3
201 ; CHECK-NEXT: cmovaq %rax, %rdx
202 ; CHECK-NEXT: ucomiss %xmm3, %xmm3
203 ; CHECK-NEXT: cmovpq %rcx, %rdx
204 ; CHECK-NEXT: movq %rdx, %xmm3
205 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm2 = xmm2[0],xmm3[0]
206 ; CHECK-NEXT: movaps %xmm0, %xmm3
207 ; CHECK-NEXT: shufps {{.*#+}} xmm3 = xmm3[3,3],xmm0[3,3]
208 ; CHECK-NEXT: cvttss2si %xmm3, %rdx
209 ; CHECK-NEXT: ucomiss %xmm1, %xmm3
210 ; CHECK-NEXT: cmovaq %rax, %rdx
211 ; CHECK-NEXT: ucomiss %xmm3, %xmm3
212 ; CHECK-NEXT: cmovpq %rcx, %rdx
213 ; CHECK-NEXT: movq %rdx, %xmm3
214 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
215 ; CHECK-NEXT: cvttss2si %xmm0, %rdx
216 ; CHECK-NEXT: ucomiss %xmm1, %xmm0
217 ; CHECK-NEXT: cmovaq %rax, %rdx
218 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
219 ; CHECK-NEXT: cmovpq %rcx, %rdx
220 ; CHECK-NEXT: movq %rdx, %xmm1
221 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm3[0]
222 ; CHECK-NEXT: movdqa %xmm2, %xmm0
223 ; CHECK-NEXT: retq
224 %x = call <4 x i64> @llvm.fptosi.sat.v4i64.v4f32(<4 x float> %f)
225 ret <4 x i64> %x
226 }
228 define <4 x i128> @test_signed_v4i128_v4f32(<4 x float> %f) nounwind {
229 ; CHECK-LABEL: test_signed_v4i128_v4f32:
230 ; CHECK: # %bb.0:
231 ; CHECK-NEXT: pushq %rbp
232 ; CHECK-NEXT: pushq %r15
233 ; CHECK-NEXT: pushq %r14
234 ; CHECK-NEXT: pushq %r13
235 ; CHECK-NEXT: pushq %r12
236 ; CHECK-NEXT: pushq %rbx
237 ; CHECK-NEXT: subq $56, %rsp
238 ; CHECK-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill
239 ; CHECK-NEXT: movq %rdi, %rbx
240 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
241 ; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
242 ; CHECK-NEXT: callq __fixsfti@PLT
243 ; CHECK-NEXT: movq %rdx, %r15
244 ; CHECK-NEXT: xorl %r14d, %r14d
245 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
246 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
247 ; CHECK-NEXT: cmovbq %r14, %rax
248 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
249 ; CHECK-NEXT: cmovbq %rcx, %r15
250 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
251 ; CHECK-NEXT: movabsq $9223372036854775807, %rbp # imm = 0x7FFFFFFFFFFFFFFF
252 ; CHECK-NEXT: cmovaq %rbp, %r15
253 ; CHECK-NEXT: movq $-1, %rcx
254 ; CHECK-NEXT: cmovaq %rcx, %rax
255 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
256 ; CHECK-NEXT: cmovpq %r14, %rax
257 ; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
258 ; CHECK-NEXT: cmovpq %r14, %r15
259 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
260 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
261 ; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
262 ; CHECK-NEXT: callq __fixsfti@PLT
263 ; CHECK-NEXT: movq %rax, %r12
264 ; CHECK-NEXT: movq %rdx, %r13
265 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
266 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
267 ; CHECK-NEXT: cmovbq %r14, %r12
268 ; CHECK-NEXT: movabsq $-9223372036854775808, %rax # imm = 0x8000000000000000
269 ; CHECK-NEXT: cmovbq %rax, %r13
270 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
271 ; CHECK-NEXT: cmovaq %rbp, %r13
272 ; CHECK-NEXT: movq $-1, %rax
273 ; CHECK-NEXT: cmovaq %rax, %r12
274 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
275 ; CHECK-NEXT: cmovpq %r14, %r12
276 ; CHECK-NEXT: cmovpq %r14, %r13
277 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
278 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
279 ; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
280 ; CHECK-NEXT: callq __fixsfti@PLT
281 ; CHECK-NEXT: movq %rax, %rbp
282 ; CHECK-NEXT: movq %rdx, %r14
283 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
284 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
285 ; CHECK-NEXT: movl $0, %eax
286 ; CHECK-NEXT: cmovbq %rax, %rbp
287 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
288 ; CHECK-NEXT: cmovbq %rcx, %r14
289 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
290 ; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
291 ; CHECK-NEXT: cmovaq %rcx, %r14
292 ; CHECK-NEXT: movq $-1, %rcx
293 ; CHECK-NEXT: cmovaq %rcx, %rbp
294 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
295 ; CHECK-NEXT: cmovpq %rax, %rbp
296 ; CHECK-NEXT: cmovpq %rax, %r14
297 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
298 ; CHECK-NEXT: callq __fixsfti@PLT
299 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
300 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
301 ; CHECK-NEXT: movl $0, %esi
302 ; CHECK-NEXT: cmovbq %rsi, %rax
303 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
304 ; CHECK-NEXT: cmovbq %rcx, %rdx
305 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
306 ; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
307 ; CHECK-NEXT: cmovaq %rcx, %rdx
308 ; CHECK-NEXT: movq $-1, %rcx
309 ; CHECK-NEXT: cmovaq %rcx, %rax
310 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
311 ; CHECK-NEXT: cmovpq %rsi, %rax
312 ; CHECK-NEXT: movl $0, %ecx
313 ; CHECK-NEXT: cmovpq %rcx, %rdx
314 ; CHECK-NEXT: movq %rdx, 8(%rbx)
315 ; CHECK-NEXT: movq %rax, (%rbx)
316 ; CHECK-NEXT: movq %r14, 56(%rbx)
317 ; CHECK-NEXT: movq %rbp, 48(%rbx)
318 ; CHECK-NEXT: movq %r13, 40(%rbx)
319 ; CHECK-NEXT: movq %r12, 32(%rbx)
320 ; CHECK-NEXT: movq %r15, 24(%rbx)
321 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
322 ; CHECK-NEXT: movq %rax, 16(%rbx)
323 ; CHECK-NEXT: movq %rbx, %rax
324 ; CHECK-NEXT: addq $56, %rsp
325 ; CHECK-NEXT: popq %rbx
326 ; CHECK-NEXT: popq %r12
327 ; CHECK-NEXT: popq %r13
328 ; CHECK-NEXT: popq %r14
329 ; CHECK-NEXT: popq %r15
330 ; CHECK-NEXT: popq %rbp
331 ; CHECK-NEXT: retq
332 %x = call <4 x i128> @llvm.fptosi.sat.v4i128.v4f32(<4 x float> %f)
333 ret <4 x i128> %x
334 }
336 ;
337 ; 64-bit float to signed integer
338 ;
340 declare <2 x i1> @llvm.fptosi.sat.v2i1.v2f64(<2 x double>)
341 declare <2 x i8> @llvm.fptosi.sat.v2i8.v2f64(<2 x double>)
342 declare <2 x i16> @llvm.fptosi.sat.v2i16.v2f64(<2 x double>)
343 declare <2 x i32> @llvm.fptosi.sat.v2i32.v2f64(<2 x double>)
344 declare <2 x i64> @llvm.fptosi.sat.v2i64.v2f64(<2 x double>)
345 declare <2 x i128> @llvm.fptosi.sat.v2i128.v2f64(<2 x double>)
347 define <2 x i1> @test_signed_v2i1_v2f64(<2 x double> %f) nounwind {
348 ; CHECK-LABEL: test_signed_v2i1_v2f64:
349 ; CHECK: # %bb.0:
350 ; CHECK-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero
351 ; CHECK-NEXT: movapd %xmm0, %xmm1
352 ; CHECK-NEXT: maxsd %xmm2, %xmm1
353 ; CHECK-NEXT: xorpd %xmm3, %xmm3
354 ; CHECK-NEXT: minsd %xmm3, %xmm1
355 ; CHECK-NEXT: cvttsd2si %xmm1, %rax
356 ; CHECK-NEXT: xorl %ecx, %ecx
357 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
358 ; CHECK-NEXT: cmovpq %rcx, %rax
359 ; CHECK-NEXT: movq %rax, %xmm1
360 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
361 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
362 ; CHECK-NEXT: maxsd %xmm2, %xmm0
363 ; CHECK-NEXT: minsd %xmm3, %xmm0
364 ; CHECK-NEXT: cvttsd2si %xmm0, %rax
365 ; CHECK-NEXT: cmovpq %rcx, %rax
366 ; CHECK-NEXT: movq %rax, %xmm0
367 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
368 ; CHECK-NEXT: movdqa %xmm1, %xmm0
369 ; CHECK-NEXT: retq
370 %x = call <2 x i1> @llvm.fptosi.sat.v2i1.v2f64(<2 x double> %f)
371 ret <2 x i1> %x
372 }
374 define <2 x i8> @test_signed_v2i8_v2f64(<2 x double> %f) nounwind {
375 ; CHECK-LABEL: test_signed_v2i8_v2f64:
376 ; CHECK: # %bb.0:
377 ; CHECK-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
378 ; CHECK-NEXT: movapd %xmm1, %xmm2
379 ; CHECK-NEXT: maxsd %xmm0, %xmm2
380 ; CHECK-NEXT: movsd {{.*#+}} xmm3 = mem[0],zero
381 ; CHECK-NEXT: movapd %xmm3, %xmm4
382 ; CHECK-NEXT: minsd %xmm2, %xmm4
383 ; CHECK-NEXT: cvttsd2si %xmm4, %eax
384 ; CHECK-NEXT: movzbl %al, %eax
385 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
386 ; CHECK-NEXT: maxsd %xmm0, %xmm1
387 ; CHECK-NEXT: minsd %xmm1, %xmm3
388 ; CHECK-NEXT: cvttsd2si %xmm3, %ecx
389 ; CHECK-NEXT: shll $8, %ecx
390 ; CHECK-NEXT: orl %eax, %ecx
391 ; CHECK-NEXT: movd %ecx, %xmm0
392 ; CHECK-NEXT: retq
393 %x = call <2 x i8> @llvm.fptosi.sat.v2i8.v2f64(<2 x double> %f)
394 ret <2 x i8> %x
395 }
397 define <2 x i16> @test_signed_v2i16_v2f64(<2 x double> %f) nounwind {
398 ; CHECK-LABEL: test_signed_v2i16_v2f64:
399 ; CHECK: # %bb.0:
400 ; CHECK-NEXT: movsd {{.*#+}} xmm1 = mem[0],zero
401 ; CHECK-NEXT: movapd %xmm1, %xmm2
402 ; CHECK-NEXT: maxsd %xmm0, %xmm1
403 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
404 ; CHECK-NEXT: maxsd %xmm0, %xmm2
405 ; CHECK-NEXT: movsd {{.*#+}} xmm0 = mem[0],zero
406 ; CHECK-NEXT: movapd %xmm0, %xmm3
407 ; CHECK-NEXT: minsd %xmm2, %xmm3
408 ; CHECK-NEXT: cvttsd2si %xmm3, %eax
409 ; CHECK-NEXT: minsd %xmm1, %xmm0
410 ; CHECK-NEXT: cvttsd2si %xmm0, %ecx
411 ; CHECK-NEXT: movd %ecx, %xmm0
412 ; CHECK-NEXT: pinsrw $1, %eax, %xmm0
413 ; CHECK-NEXT: retq
414 %x = call <2 x i16> @llvm.fptosi.sat.v2i16.v2f64(<2 x double> %f)
415 ret <2 x i16> %x
416 }
418 define <2 x i32> @test_signed_v2i32_v2f64(<2 x double> %f) nounwind {
419 ; CHECK-LABEL: test_signed_v2i32_v2f64:
420 ; CHECK: # %bb.0:
421 ; CHECK-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero
422 ; CHECK-NEXT: movapd %xmm0, %xmm1
423 ; CHECK-NEXT: maxsd %xmm2, %xmm1
424 ; CHECK-NEXT: movsd {{.*#+}} xmm3 = mem[0],zero
425 ; CHECK-NEXT: minsd %xmm3, %xmm1
426 ; CHECK-NEXT: cvttsd2si %xmm1, %eax
427 ; CHECK-NEXT: xorl %ecx, %ecx
428 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
429 ; CHECK-NEXT: cmovpl %ecx, %eax
430 ; CHECK-NEXT: movd %eax, %xmm1
431 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
432 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
433 ; CHECK-NEXT: maxsd %xmm2, %xmm0
434 ; CHECK-NEXT: minsd %xmm3, %xmm0
435 ; CHECK-NEXT: cvttsd2si %xmm0, %eax
436 ; CHECK-NEXT: cmovpl %ecx, %eax
437 ; CHECK-NEXT: movd %eax, %xmm0
438 ; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
439 ; CHECK-NEXT: movdqa %xmm1, %xmm0
440 ; CHECK-NEXT: retq
441 %x = call <2 x i32> @llvm.fptosi.sat.v2i32.v2f64(<2 x double> %f)
442 ret <2 x i32> %x
443 }
445 define <2 x i64> @test_signed_v2i64_v2f64(<2 x double> %f) nounwind {
446 ; CHECK-LABEL: test_signed_v2i64_v2f64:
447 ; CHECK: # %bb.0:
448 ; CHECK-NEXT: cvttsd2si %xmm0, %rax
449 ; CHECK-NEXT: movsd {{.*#+}} xmm2 = mem[0],zero
450 ; CHECK-NEXT: ucomisd %xmm2, %xmm0
451 ; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
452 ; CHECK-NEXT: cmovaq %rcx, %rax
453 ; CHECK-NEXT: xorl %edx, %edx
454 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
455 ; CHECK-NEXT: cmovpq %rdx, %rax
456 ; CHECK-NEXT: movq %rax, %xmm1
457 ; CHECK-NEXT: unpckhpd {{.*#+}} xmm0 = xmm0[1,1]
458 ; CHECK-NEXT: cvttsd2si %xmm0, %rax
459 ; CHECK-NEXT: ucomisd %xmm2, %xmm0
460 ; CHECK-NEXT: cmovaq %rcx, %rax
461 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
462 ; CHECK-NEXT: cmovpq %rdx, %rax
463 ; CHECK-NEXT: movq %rax, %xmm0
464 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
465 ; CHECK-NEXT: movdqa %xmm1, %xmm0
466 ; CHECK-NEXT: retq
467 %x = call <2 x i64> @llvm.fptosi.sat.v2i64.v2f64(<2 x double> %f)
468 ret <2 x i64> %x
469 }
471 define <2 x i128> @test_signed_v2i128_v2f64(<2 x double> %f) nounwind {
472 ; CHECK-LABEL: test_signed_v2i128_v2f64:
473 ; CHECK: # %bb.0:
474 ; CHECK-NEXT: pushq %rbp
475 ; CHECK-NEXT: pushq %r15
476 ; CHECK-NEXT: pushq %r14
477 ; CHECK-NEXT: pushq %r13
478 ; CHECK-NEXT: pushq %r12
479 ; CHECK-NEXT: pushq %rbx
480 ; CHECK-NEXT: subq $40, %rsp
481 ; CHECK-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill
482 ; CHECK-NEXT: movq %rdi, %rbx
483 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
484 ; CHECK-NEXT: movaps %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
485 ; CHECK-NEXT: callq __fixdfti@PLT
486 ; CHECK-NEXT: movq %rax, %r14
487 ; CHECK-NEXT: movq %rdx, %r15
488 ; CHECK-NEXT: xorl %r12d, %r12d
489 ; CHECK-NEXT: movapd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
490 ; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
491 ; CHECK-NEXT: cmovbq %r12, %r14
492 ; CHECK-NEXT: movabsq $-9223372036854775808, %rax # imm = 0x8000000000000000
493 ; CHECK-NEXT: cmovbq %rax, %r15
494 ; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
495 ; CHECK-NEXT: movabsq $9223372036854775807, %rbp # imm = 0x7FFFFFFFFFFFFFFF
496 ; CHECK-NEXT: cmovaq %rbp, %r15
497 ; CHECK-NEXT: movq $-1, %r13
498 ; CHECK-NEXT: cmovaq %r13, %r14
499 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
500 ; CHECK-NEXT: cmovpq %r12, %r14
501 ; CHECK-NEXT: cmovpq %r12, %r15
502 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
503 ; CHECK-NEXT: callq __fixdfti@PLT
504 ; CHECK-NEXT: movapd (%rsp), %xmm0 # 16-byte Reload
505 ; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
506 ; CHECK-NEXT: cmovbq %r12, %rax
507 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
508 ; CHECK-NEXT: cmovbq %rcx, %rdx
509 ; CHECK-NEXT: ucomisd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
510 ; CHECK-NEXT: cmovaq %rbp, %rdx
511 ; CHECK-NEXT: cmovaq %r13, %rax
512 ; CHECK-NEXT: ucomisd %xmm0, %xmm0
513 ; CHECK-NEXT: cmovpq %r12, %rax
514 ; CHECK-NEXT: cmovpq %r12, %rdx
515 ; CHECK-NEXT: movq %rdx, 8(%rbx)
516 ; CHECK-NEXT: movq %rax, (%rbx)
517 ; CHECK-NEXT: movq %r15, 24(%rbx)
518 ; CHECK-NEXT: movq %r14, 16(%rbx)
519 ; CHECK-NEXT: movq %rbx, %rax
520 ; CHECK-NEXT: addq $40, %rsp
521 ; CHECK-NEXT: popq %rbx
522 ; CHECK-NEXT: popq %r12
523 ; CHECK-NEXT: popq %r13
524 ; CHECK-NEXT: popq %r14
525 ; CHECK-NEXT: popq %r15
526 ; CHECK-NEXT: popq %rbp
527 ; CHECK-NEXT: retq
528 %x = call <2 x i128> @llvm.fptosi.sat.v2i128.v2f64(<2 x double> %f)
529 ret <2 x i128> %x
530 }
532 ;
533 ; 16-bit float to signed integer
534 ;
536 declare <8 x i1> @llvm.fptosi.sat.v8i1.v8f16(<8 x half>)
537 declare <8 x i8> @llvm.fptosi.sat.v8i8.v8f16(<8 x half>)
538 declare <8 x i16> @llvm.fptosi.sat.v8i16.v8f16(<8 x half>)
539 declare <8 x i32> @llvm.fptosi.sat.v8i32.v8f16(<8 x half>)
540 declare <8 x i64> @llvm.fptosi.sat.v8i64.v8f16(<8 x half>)
541 declare <8 x i128> @llvm.fptosi.sat.v8i128.v8f16(<8 x half>)
543 define <8 x i1> @test_signed_v8i1_v8f16(<8 x half> %f) nounwind {
544 ; CHECK-LABEL: test_signed_v8i1_v8f16:
545 ; CHECK: # %bb.0:
546 ; CHECK-NEXT: pushq %rbp
547 ; CHECK-NEXT: pushq %rbx
548 ; CHECK-NEXT: subq $72, %rsp
549 ; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill
550 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
551 ; CHECK-NEXT: callq __extendhfsf2@PLT
552 ; CHECK-NEXT: cvttss2si %xmm0, %eax
553 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
554 ; CHECK-NEXT: movl $65535, %ebp # imm = 0xFFFF
555 ; CHECK-NEXT: cmovbl %ebp, %eax
556 ; CHECK-NEXT: xorl %ebx, %ebx
557 ; CHECK-NEXT: xorps %xmm1, %xmm1
558 ; CHECK-NEXT: ucomiss %xmm1, %xmm0
559 ; CHECK-NEXT: cmoval %ebx, %eax
560 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
561 ; CHECK-NEXT: cmovpl %ebx, %eax
562 ; CHECK-NEXT: movd %eax, %xmm0
563 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
564 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
565 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
566 ; CHECK-NEXT: callq __extendhfsf2@PLT
567 ; CHECK-NEXT: cvttss2si %xmm0, %eax
568 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
569 ; CHECK-NEXT: cmovbl %ebp, %eax
570 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
571 ; CHECK-NEXT: cmoval %ebx, %eax
572 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
573 ; CHECK-NEXT: cmovpl %ebx, %eax
574 ; CHECK-NEXT: movd %eax, %xmm0
575 ; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
576 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
577 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
578 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
579 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
580 ; CHECK-NEXT: callq __extendhfsf2@PLT
581 ; CHECK-NEXT: cvttss2si %xmm0, %eax
582 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
583 ; CHECK-NEXT: cmovbl %ebp, %eax
584 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
585 ; CHECK-NEXT: cmoval %ebx, %eax
586 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
587 ; CHECK-NEXT: cmovpl %ebx, %eax
588 ; CHECK-NEXT: movd %eax, %xmm0
589 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
590 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
591 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
592 ; CHECK-NEXT: callq __extendhfsf2@PLT
593 ; CHECK-NEXT: cvttss2si %xmm0, %eax
594 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
595 ; CHECK-NEXT: cmovbl %ebp, %eax
596 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
597 ; CHECK-NEXT: cmoval %ebx, %eax
598 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
599 ; CHECK-NEXT: cmovpl %ebx, %eax
600 ; CHECK-NEXT: movd %eax, %xmm0
601 ; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
602 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
603 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
604 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
605 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
606 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
607 ; CHECK-NEXT: psrlq $48, %xmm0
608 ; CHECK-NEXT: callq __extendhfsf2@PLT
609 ; CHECK-NEXT: cvttss2si %xmm0, %eax
610 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
611 ; CHECK-NEXT: cmovbl %ebp, %eax
612 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
613 ; CHECK-NEXT: cmoval %ebx, %eax
614 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
615 ; CHECK-NEXT: cmovpl %ebx, %eax
616 ; CHECK-NEXT: movd %eax, %xmm0
617 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
618 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
619 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
620 ; CHECK-NEXT: callq __extendhfsf2@PLT
621 ; CHECK-NEXT: cvttss2si %xmm0, %eax
622 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
623 ; CHECK-NEXT: cmovbl %ebp, %eax
624 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
625 ; CHECK-NEXT: cmoval %ebx, %eax
626 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
627 ; CHECK-NEXT: cmovpl %ebx, %eax
628 ; CHECK-NEXT: movd %eax, %xmm0
629 ; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
630 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
631 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
632 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
633 ; CHECK-NEXT: callq __extendhfsf2@PLT
634 ; CHECK-NEXT: cvttss2si %xmm0, %eax
635 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
636 ; CHECK-NEXT: cmovbl %ebp, %eax
637 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
638 ; CHECK-NEXT: cmoval %ebx, %eax
639 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
640 ; CHECK-NEXT: cmovpl %ebx, %eax
641 ; CHECK-NEXT: movd %eax, %xmm0
642 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
643 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
644 ; CHECK-NEXT: psrld $16, %xmm0
645 ; CHECK-NEXT: callq __extendhfsf2@PLT
646 ; CHECK-NEXT: cvttss2si %xmm0, %eax
647 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
648 ; CHECK-NEXT: cmovbl %ebp, %eax
649 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
650 ; CHECK-NEXT: cmoval %ebx, %eax
651 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
652 ; CHECK-NEXT: cmovpl %ebx, %eax
653 ; CHECK-NEXT: movd %eax, %xmm1
654 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
655 ; CHECK-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
656 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
657 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
658 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
659 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]
660 ; CHECK-NEXT: addq $72, %rsp
661 ; CHECK-NEXT: popq %rbx
662 ; CHECK-NEXT: popq %rbp
663 ; CHECK-NEXT: retq
664 %x = call <8 x i1> @llvm.fptosi.sat.v8i1.v8f16(<8 x half> %f)
665 ret <8 x i1> %x
666 }
668 define <8 x i8> @test_signed_v8i8_v8f16(<8 x half> %f) nounwind {
669 ; CHECK-LABEL: test_signed_v8i8_v8f16:
670 ; CHECK: # %bb.0:
671 ; CHECK-NEXT: pushq %rbp
672 ; CHECK-NEXT: pushq %r15
673 ; CHECK-NEXT: pushq %r14
674 ; CHECK-NEXT: pushq %r13
675 ; CHECK-NEXT: pushq %r12
676 ; CHECK-NEXT: pushq %rbx
677 ; CHECK-NEXT: subq $40, %rsp
678 ; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill
679 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
680 ; CHECK-NEXT: callq __extendhfsf2@PLT
681 ; CHECK-NEXT: cvttss2si %xmm0, %r12d
682 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
683 ; CHECK-NEXT: movl $128, %ebx
684 ; CHECK-NEXT: cmovbl %ebx, %r12d
685 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
686 ; CHECK-NEXT: movl $127, %ebp
687 ; CHECK-NEXT: cmoval %ebp, %r12d
688 ; CHECK-NEXT: xorl %r14d, %r14d
689 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
690 ; CHECK-NEXT: cmovpl %r14d, %r12d
691 ; CHECK-NEXT: shll $8, %r12d
692 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
693 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
694 ; CHECK-NEXT: callq __extendhfsf2@PLT
695 ; CHECK-NEXT: cvttss2si %xmm0, %eax
696 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
697 ; CHECK-NEXT: cmovbl %ebx, %eax
698 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
699 ; CHECK-NEXT: cmoval %ebp, %eax
700 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
701 ; CHECK-NEXT: cmovpl %r14d, %eax
702 ; CHECK-NEXT: movzbl %al, %r15d
703 ; CHECK-NEXT: orl %r12d, %r15d
704 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
705 ; CHECK-NEXT: callq __extendhfsf2@PLT
706 ; CHECK-NEXT: cvttss2si %xmm0, %eax
707 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
708 ; CHECK-NEXT: cmovbl %ebx, %eax
709 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
710 ; CHECK-NEXT: cmoval %ebp, %eax
711 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
712 ; CHECK-NEXT: cmovpl %r14d, %eax
713 ; CHECK-NEXT: movzbl %al, %r12d
714 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
715 ; CHECK-NEXT: psrld $16, %xmm0
716 ; CHECK-NEXT: callq __extendhfsf2@PLT
717 ; CHECK-NEXT: cvttss2si %xmm0, %eax
718 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
719 ; CHECK-NEXT: cmovbl %ebx, %eax
720 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
721 ; CHECK-NEXT: cmoval %ebp, %eax
722 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
723 ; CHECK-NEXT: cmovpl %r14d, %eax
724 ; CHECK-NEXT: movzbl %al, %r13d
725 ; CHECK-NEXT: shll $8, %r13d
726 ; CHECK-NEXT: orl %r12d, %r13d
727 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
728 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
729 ; CHECK-NEXT: callq __extendhfsf2@PLT
730 ; CHECK-NEXT: cvttss2si %xmm0, %eax
731 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
732 ; CHECK-NEXT: cmovbl %ebx, %eax
733 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
734 ; CHECK-NEXT: cmoval %ebp, %eax
735 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
736 ; CHECK-NEXT: cmovpl %r14d, %eax
737 ; CHECK-NEXT: movzbl %al, %r12d
738 ; CHECK-NEXT: shll $16, %r12d
739 ; CHECK-NEXT: orl %r13d, %r12d
740 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
741 ; CHECK-NEXT: psrlq $48, %xmm0
742 ; CHECK-NEXT: callq __extendhfsf2@PLT
743 ; CHECK-NEXT: cvttss2si %xmm0, %eax
744 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
745 ; CHECK-NEXT: cmovbl %ebx, %eax
746 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
747 ; CHECK-NEXT: cmoval %ebp, %eax
748 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
749 ; CHECK-NEXT: cmovpl %r14d, %eax
750 ; CHECK-NEXT: shll $24, %eax
751 ; CHECK-NEXT: orl %r12d, %eax
752 ; CHECK-NEXT: movd %eax, %xmm0
753 ; CHECK-NEXT: pinsrw $2, %r15d, %xmm0
754 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
755 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
756 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
757 ; CHECK-NEXT: callq __extendhfsf2@PLT
758 ; CHECK-NEXT: cvttss2si %xmm0, %r15d
759 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
760 ; CHECK-NEXT: cmovbl %ebx, %r15d
761 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
762 ; CHECK-NEXT: cmoval %ebp, %r15d
763 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
764 ; CHECK-NEXT: cmovpl %r14d, %r15d
765 ; CHECK-NEXT: shll $8, %r15d
766 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
767 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
768 ; CHECK-NEXT: callq __extendhfsf2@PLT
769 ; CHECK-NEXT: cvttss2si %xmm0, %eax
770 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
771 ; CHECK-NEXT: cmovbl %ebx, %eax
772 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
773 ; CHECK-NEXT: cmoval %ebp, %eax
774 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
775 ; CHECK-NEXT: cmovpl %r14d, %eax
776 ; CHECK-NEXT: movzbl %al, %eax
777 ; CHECK-NEXT: orl %r15d, %eax
778 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
779 ; CHECK-NEXT: pinsrw $3, %eax, %xmm0
780 ; CHECK-NEXT: addq $40, %rsp
781 ; CHECK-NEXT: popq %rbx
782 ; CHECK-NEXT: popq %r12
783 ; CHECK-NEXT: popq %r13
784 ; CHECK-NEXT: popq %r14
785 ; CHECK-NEXT: popq %r15
786 ; CHECK-NEXT: popq %rbp
787 ; CHECK-NEXT: retq
788 %x = call <8 x i8> @llvm.fptosi.sat.v8i8.v8f16(<8 x half> %f)
789 ret <8 x i8> %x
790 }
792 define <8 x i16> @test_signed_v8i16_v8f16(<8 x half> %f) nounwind {
793 ; CHECK-LABEL: test_signed_v8i16_v8f16:
794 ; CHECK: # %bb.0:
795 ; CHECK-NEXT: pushq %rbp
796 ; CHECK-NEXT: pushq %r14
797 ; CHECK-NEXT: pushq %rbx
798 ; CHECK-NEXT: subq $64, %rsp
799 ; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill
800 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
801 ; CHECK-NEXT: callq __extendhfsf2@PLT
802 ; CHECK-NEXT: cvttss2si %xmm0, %eax
803 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
804 ; CHECK-NEXT: movl $32768, %ebx # imm = 0x8000
805 ; CHECK-NEXT: cmovbl %ebx, %eax
806 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
807 ; CHECK-NEXT: movl $32767, %ebp # imm = 0x7FFF
808 ; CHECK-NEXT: cmoval %ebp, %eax
809 ; CHECK-NEXT: xorl %r14d, %r14d
810 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
811 ; CHECK-NEXT: cmovpl %r14d, %eax
812 ; CHECK-NEXT: movd %eax, %xmm0
813 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
814 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
815 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
816 ; CHECK-NEXT: callq __extendhfsf2@PLT
817 ; CHECK-NEXT: cvttss2si %xmm0, %eax
818 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
819 ; CHECK-NEXT: cmovbl %ebx, %eax
820 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
821 ; CHECK-NEXT: cmoval %ebp, %eax
822 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
823 ; CHECK-NEXT: cmovpl %r14d, %eax
824 ; CHECK-NEXT: movd %eax, %xmm0
825 ; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
826 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
827 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
828 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
829 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
830 ; CHECK-NEXT: callq __extendhfsf2@PLT
831 ; CHECK-NEXT: cvttss2si %xmm0, %eax
832 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
833 ; CHECK-NEXT: cmovbl %ebx, %eax
834 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
835 ; CHECK-NEXT: cmoval %ebp, %eax
836 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
837 ; CHECK-NEXT: cmovpl %r14d, %eax
838 ; CHECK-NEXT: movd %eax, %xmm0
839 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
840 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
841 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
842 ; CHECK-NEXT: callq __extendhfsf2@PLT
843 ; CHECK-NEXT: cvttss2si %xmm0, %eax
844 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
845 ; CHECK-NEXT: cmovbl %ebx, %eax
846 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
847 ; CHECK-NEXT: cmoval %ebp, %eax
848 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
849 ; CHECK-NEXT: cmovpl %r14d, %eax
850 ; CHECK-NEXT: movd %eax, %xmm0
851 ; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
852 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
853 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
854 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
855 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
856 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
857 ; CHECK-NEXT: psrlq $48, %xmm0
858 ; CHECK-NEXT: callq __extendhfsf2@PLT
859 ; CHECK-NEXT: cvttss2si %xmm0, %eax
860 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
861 ; CHECK-NEXT: cmovbl %ebx, %eax
862 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
863 ; CHECK-NEXT: cmoval %ebp, %eax
864 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
865 ; CHECK-NEXT: cmovpl %r14d, %eax
866 ; CHECK-NEXT: movd %eax, %xmm0
867 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
868 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
869 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
870 ; CHECK-NEXT: callq __extendhfsf2@PLT
871 ; CHECK-NEXT: cvttss2si %xmm0, %eax
872 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
873 ; CHECK-NEXT: cmovbl %ebx, %eax
874 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
875 ; CHECK-NEXT: cmoval %ebp, %eax
876 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
877 ; CHECK-NEXT: cmovpl %r14d, %eax
878 ; CHECK-NEXT: movd %eax, %xmm0
879 ; CHECK-NEXT: punpcklwd {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
880 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1],xmm0[2],mem[2],xmm0[3],mem[3]
881 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
882 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
883 ; CHECK-NEXT: callq __extendhfsf2@PLT
884 ; CHECK-NEXT: cvttss2si %xmm0, %eax
885 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
886 ; CHECK-NEXT: cmovbl %ebx, %eax
887 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
888 ; CHECK-NEXT: cmoval %ebp, %eax
889 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
890 ; CHECK-NEXT: cmovpl %r14d, %eax
891 ; CHECK-NEXT: movd %eax, %xmm0
892 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
893 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
894 ; CHECK-NEXT: psrld $16, %xmm0
895 ; CHECK-NEXT: callq __extendhfsf2@PLT
896 ; CHECK-NEXT: cvttss2si %xmm0, %eax
897 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
898 ; CHECK-NEXT: cmovbl %ebx, %eax
899 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
900 ; CHECK-NEXT: cmoval %ebp, %eax
901 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
902 ; CHECK-NEXT: cmovpl %r14d, %eax
903 ; CHECK-NEXT: movd %eax, %xmm1
904 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
905 ; CHECK-NEXT: punpcklwd {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1],xmm0[2],xmm1[2],xmm0[3],xmm1[3]
906 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
907 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
908 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
909 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]
910 ; CHECK-NEXT: addq $64, %rsp
911 ; CHECK-NEXT: popq %rbx
912 ; CHECK-NEXT: popq %r14
913 ; CHECK-NEXT: popq %rbp
914 ; CHECK-NEXT: retq
915 %x = call <8 x i16> @llvm.fptosi.sat.v8i16.v8f16(<8 x half> %f)
916 ret <8 x i16> %x
917 }
919 define <8 x i32> @test_signed_v8i32_v8f16(<8 x half> %f) nounwind {
920 ; CHECK-LABEL: test_signed_v8i32_v8f16:
921 ; CHECK: # %bb.0:
922 ; CHECK-NEXT: pushq %rbp
923 ; CHECK-NEXT: pushq %r14
924 ; CHECK-NEXT: pushq %rbx
925 ; CHECK-NEXT: subq $64, %rsp
926 ; CHECK-NEXT: movdqa %xmm0, (%rsp) # 16-byte Spill
927 ; CHECK-NEXT: psrlq $48, %xmm0
928 ; CHECK-NEXT: callq __extendhfsf2@PLT
929 ; CHECK-NEXT: cvttss2si %xmm0, %eax
930 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
931 ; CHECK-NEXT: movl $-2147483648, %ebx # imm = 0x80000000
932 ; CHECK-NEXT: cmovbl %ebx, %eax
933 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
934 ; CHECK-NEXT: movl $2147483647, %ebp # imm = 0x7FFFFFFF
935 ; CHECK-NEXT: cmoval %ebp, %eax
936 ; CHECK-NEXT: xorl %r14d, %r14d
937 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
938 ; CHECK-NEXT: cmovpl %r14d, %eax
939 ; CHECK-NEXT: movd %eax, %xmm0
940 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
941 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
942 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
943 ; CHECK-NEXT: callq __extendhfsf2@PLT
944 ; CHECK-NEXT: cvttss2si %xmm0, %eax
945 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
946 ; CHECK-NEXT: cmovbl %ebx, %eax
947 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
948 ; CHECK-NEXT: cmoval %ebp, %eax
949 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
950 ; CHECK-NEXT: cmovpl %r14d, %eax
951 ; CHECK-NEXT: movd %eax, %xmm0
952 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
953 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
954 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
955 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
956 ; CHECK-NEXT: callq __extendhfsf2@PLT
957 ; CHECK-NEXT: cvttss2si %xmm0, %eax
958 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
959 ; CHECK-NEXT: cmovbl %ebx, %eax
960 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
961 ; CHECK-NEXT: cmoval %ebp, %eax
962 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
963 ; CHECK-NEXT: cmovpl %r14d, %eax
964 ; CHECK-NEXT: movd %eax, %xmm0
965 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
966 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
967 ; CHECK-NEXT: psrld $16, %xmm0
968 ; CHECK-NEXT: callq __extendhfsf2@PLT
969 ; CHECK-NEXT: cvttss2si %xmm0, %eax
970 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
971 ; CHECK-NEXT: cmovbl %ebx, %eax
972 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
973 ; CHECK-NEXT: cmoval %ebp, %eax
974 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
975 ; CHECK-NEXT: cmovpl %r14d, %eax
976 ; CHECK-NEXT: movd %eax, %xmm0
977 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
978 ; CHECK-NEXT: punpckldq {{.*#+}} xmm1 = xmm1[0],xmm0[0],xmm1[1],xmm0[1]
979 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
980 ; CHECK-NEXT: # xmm1 = xmm1[0],mem[0]
981 ; CHECK-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
982 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
983 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
984 ; CHECK-NEXT: callq __extendhfsf2@PLT
985 ; CHECK-NEXT: cvttss2si %xmm0, %eax
986 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
987 ; CHECK-NEXT: cmovbl %ebx, %eax
988 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
989 ; CHECK-NEXT: cmoval %ebp, %eax
990 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
991 ; CHECK-NEXT: cmovpl %r14d, %eax
992 ; CHECK-NEXT: movd %eax, %xmm0
993 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
994 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
995 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
996 ; CHECK-NEXT: callq __extendhfsf2@PLT
997 ; CHECK-NEXT: cvttss2si %xmm0, %eax
998 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
999 ; CHECK-NEXT: cmovbl %ebx, %eax
1000 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1001 ; CHECK-NEXT: cmoval %ebp, %eax
1002 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1003 ; CHECK-NEXT: cmovpl %r14d, %eax
1004 ; CHECK-NEXT: movd %eax, %xmm0
1005 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
1006 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0],xmm0[1],mem[1]
1007 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1008 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
1009 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
1010 ; CHECK-NEXT: callq __extendhfsf2@PLT
1011 ; CHECK-NEXT: cvttss2si %xmm0, %eax
1012 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1013 ; CHECK-NEXT: cmovbl %ebx, %eax
1014 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1015 ; CHECK-NEXT: cmoval %ebp, %eax
1016 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1017 ; CHECK-NEXT: cmovpl %r14d, %eax
1018 ; CHECK-NEXT: movd %eax, %xmm0
1019 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1020 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
1021 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
1022 ; CHECK-NEXT: callq __extendhfsf2@PLT
1023 ; CHECK-NEXT: cvttss2si %xmm0, %eax
1024 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1025 ; CHECK-NEXT: cmovbl %ebx, %eax
1026 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1027 ; CHECK-NEXT: cmoval %ebp, %eax
1028 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1029 ; CHECK-NEXT: cmovpl %r14d, %eax
1030 ; CHECK-NEXT: movd %eax, %xmm1
1031 ; CHECK-NEXT: punpckldq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
1032 ; CHECK-NEXT: # xmm1 = xmm1[0],mem[0],xmm1[1],mem[1]
1033 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Folded Reload
1034 ; CHECK-NEXT: # xmm1 = xmm1[0],mem[0]
1035 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1036 ; CHECK-NEXT: addq $64, %rsp
1037 ; CHECK-NEXT: popq %rbx
1038 ; CHECK-NEXT: popq %r14
1039 ; CHECK-NEXT: popq %rbp
1040 ; CHECK-NEXT: retq
1041 %x = call <8 x i32> @llvm.fptosi.sat.v8i32.v8f16(<8 x half> %f)
1042 ret <8 x i32> %x
1043 }
1045 define <8 x i64> @test_signed_v8i64_v8f16(<8 x half> %f) nounwind {
1046 ; CHECK-LABEL: test_signed_v8i64_v8f16:
1047 ; CHECK: # %bb.0:
1048 ; CHECK-NEXT: pushq %r15
1049 ; CHECK-NEXT: pushq %r14
1050 ; CHECK-NEXT: pushq %rbx
1051 ; CHECK-NEXT: subq $80, %rsp
1052 ; CHECK-NEXT: movaps %xmm0, (%rsp) # 16-byte Spill
1053 ; CHECK-NEXT: callq __extendhfsf2@PLT
1054 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1055 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1056 ; CHECK-NEXT: movabsq $-9223372036854775808, %rbx # imm = 0x8000000000000000
1057 ; CHECK-NEXT: cmovbq %rbx, %rax
1058 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1059 ; CHECK-NEXT: movabsq $9223372036854775807, %r14 # imm = 0x7FFFFFFFFFFFFFFF
1060 ; CHECK-NEXT: cmovaq %r14, %rax
1061 ; CHECK-NEXT: xorl %r15d, %r15d
1062 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1063 ; CHECK-NEXT: cmovpq %r15, %rax
1064 ; CHECK-NEXT: movq %rax, %xmm0
1065 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1066 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
1067 ; CHECK-NEXT: psrld $16, %xmm0
1068 ; CHECK-NEXT: callq __extendhfsf2@PLT
1069 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1070 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1071 ; CHECK-NEXT: cmovbq %rbx, %rax
1072 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1073 ; CHECK-NEXT: cmovaq %r14, %rax
1074 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1075 ; CHECK-NEXT: cmovpq %r15, %rax
1076 ; CHECK-NEXT: movq %rax, %xmm0
1077 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
1078 ; CHECK-NEXT: punpcklqdq {{.*#+}} xmm1 = xmm1[0],xmm0[0]
1079 ; CHECK-NEXT: movdqa %xmm1, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1080 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
1081 ; CHECK-NEXT: psrlq $48, %xmm0
1082 ; CHECK-NEXT: callq __extendhfsf2@PLT
1083 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1084 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1085 ; CHECK-NEXT: cmovbq %rbx, %rax
1086 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1087 ; CHECK-NEXT: cmovaq %r14, %rax
1088 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1089 ; CHECK-NEXT: cmovpq %r15, %rax
1090 ; CHECK-NEXT: movq %rax, %xmm0
1091 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1092 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
1093 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
1094 ; CHECK-NEXT: callq __extendhfsf2@PLT
1095 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1096 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1097 ; CHECK-NEXT: cmovbq %rbx, %rax
1098 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1099 ; CHECK-NEXT: cmovaq %r14, %rax
1100 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1101 ; CHECK-NEXT: cmovpq %r15, %rax
1102 ; CHECK-NEXT: movq %rax, %xmm0
1103 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
1104 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]
1105 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1106 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
1107 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
1108 ; CHECK-NEXT: callq __extendhfsf2@PLT
1109 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1110 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1111 ; CHECK-NEXT: cmovbq %rbx, %rax
1112 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1113 ; CHECK-NEXT: cmovaq %r14, %rax
1114 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1115 ; CHECK-NEXT: cmovpq %r15, %rax
1116 ; CHECK-NEXT: movq %rax, %xmm0
1117 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1118 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
1119 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
1120 ; CHECK-NEXT: callq __extendhfsf2@PLT
1121 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1122 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1123 ; CHECK-NEXT: cmovbq %rbx, %rax
1124 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1125 ; CHECK-NEXT: cmovaq %r14, %rax
1126 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1127 ; CHECK-NEXT: cmovpq %r15, %rax
1128 ; CHECK-NEXT: movq %rax, %xmm0
1129 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Folded Reload
1130 ; CHECK-NEXT: # xmm0 = xmm0[0],mem[0]
1131 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1132 ; CHECK-NEXT: movdqa (%rsp), %xmm0 # 16-byte Reload
1133 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
1134 ; CHECK-NEXT: callq __extendhfsf2@PLT
1135 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1136 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1137 ; CHECK-NEXT: cmovbq %rbx, %rax
1138 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1139 ; CHECK-NEXT: cmovaq %r14, %rax
1140 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1141 ; CHECK-NEXT: cmovpq %r15, %rax
1142 ; CHECK-NEXT: movq %rax, %xmm0
1143 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1144 ; CHECK-NEXT: movaps (%rsp), %xmm0 # 16-byte Reload
1145 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
1146 ; CHECK-NEXT: callq __extendhfsf2@PLT
1147 ; CHECK-NEXT: cvttss2si %xmm0, %rax
1148 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1149 ; CHECK-NEXT: cmovbq %rbx, %rax
1150 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1151 ; CHECK-NEXT: cmovaq %r14, %rax
1152 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1153 ; CHECK-NEXT: cmovpq %r15, %rax
1154 ; CHECK-NEXT: movq %rax, %xmm3
1155 ; CHECK-NEXT: punpcklqdq {{[-0-9]+}}(%r{{[sb]}}p), %xmm3 # 16-byte Folded Reload
1156 ; CHECK-NEXT: # xmm3 = xmm3[0],mem[0]
1157 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1158 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm1 # 16-byte Reload
1159 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm2 # 16-byte Reload
1160 ; CHECK-NEXT: addq $80, %rsp
1161 ; CHECK-NEXT: popq %rbx
1162 ; CHECK-NEXT: popq %r14
1163 ; CHECK-NEXT: popq %r15
1164 ; CHECK-NEXT: retq
1165 %x = call <8 x i64> @llvm.fptosi.sat.v8i64.v8f16(<8 x half> %f)
1166 ret <8 x i64> %x
1167 }
1169 define <8 x i128> @test_signed_v8i128_v8f16(<8 x half> %f) nounwind {
1170 ; CHECK-LABEL: test_signed_v8i128_v8f16:
1171 ; CHECK: # %bb.0:
1172 ; CHECK-NEXT: pushq %rbp
1173 ; CHECK-NEXT: pushq %r15
1174 ; CHECK-NEXT: pushq %r14
1175 ; CHECK-NEXT: pushq %r13
1176 ; CHECK-NEXT: pushq %r12
1177 ; CHECK-NEXT: pushq %rbx
1178 ; CHECK-NEXT: subq $104, %rsp
1179 ; CHECK-NEXT: movdqa %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 16-byte Spill
1180 ; CHECK-NEXT: movq %rdi, %rbx
1181 ; CHECK-NEXT: psrld $16, %xmm0
1182 ; CHECK-NEXT: callq __extendhfsf2@PLT
1183 ; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill
1184 ; CHECK-NEXT: callq __fixsfti@PLT
1185 ; CHECK-NEXT: xorl %r12d, %r12d
1186 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1187 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1188 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1189 ; CHECK-NEXT: cmovbq %r12, %rax
1190 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
1191 ; CHECK-NEXT: cmovbq %rcx, %rdx
1192 ; CHECK-NEXT: movq %rcx, %r14
1193 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1194 ; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
1195 ; CHECK-NEXT: cmovaq %rcx, %rdx
1196 ; CHECK-NEXT: movq %rcx, %r15
1197 ; CHECK-NEXT: movq $-1, %rcx
1198 ; CHECK-NEXT: cmovaq %rcx, %rax
1199 ; CHECK-NEXT: movq $-1, %r13
1200 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1201 ; CHECK-NEXT: cmovpq %r12, %rax
1202 ; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1203 ; CHECK-NEXT: cmovpq %r12, %rdx
1204 ; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1205 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1206 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[1,1,1,1]
1207 ; CHECK-NEXT: callq __extendhfsf2@PLT
1208 ; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1209 ; CHECK-NEXT: callq __fixsfti@PLT
1210 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1211 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1212 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1213 ; CHECK-NEXT: cmovbq %r12, %rax
1214 ; CHECK-NEXT: cmovbq %r14, %rdx
1215 ; CHECK-NEXT: movq %r14, %rbp
1216 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1217 ; CHECK-NEXT: cmovaq %r15, %rdx
1218 ; CHECK-NEXT: cmovaq %r13, %rax
1219 ; CHECK-NEXT: movq $-1, %r14
1220 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1221 ; CHECK-NEXT: cmovpq %r12, %rax
1222 ; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1223 ; CHECK-NEXT: cmovpq %r12, %rdx
1224 ; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1225 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1226 ; CHECK-NEXT: psrlq $48, %xmm0
1227 ; CHECK-NEXT: callq __extendhfsf2@PLT
1228 ; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill
1229 ; CHECK-NEXT: callq __fixsfti@PLT
1230 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1231 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1232 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1233 ; CHECK-NEXT: cmovbq %r12, %rax
1234 ; CHECK-NEXT: cmovbq %rbp, %rdx
1235 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1236 ; CHECK-NEXT: cmovaq %r15, %rdx
1237 ; CHECK-NEXT: cmovaq %r14, %rax
1238 ; CHECK-NEXT: movq $-1, %r14
1239 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1240 ; CHECK-NEXT: cmovpq %r12, %rax
1241 ; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1242 ; CHECK-NEXT: cmovpq %r12, %rdx
1243 ; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1244 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1245 ; CHECK-NEXT: movhlps {{.*#+}} xmm0 = xmm0[1,1]
1246 ; CHECK-NEXT: callq __extendhfsf2@PLT
1247 ; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1248 ; CHECK-NEXT: callq __fixsfti@PLT
1249 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1250 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1251 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1252 ; CHECK-NEXT: cmovbq %r12, %rax
1253 ; CHECK-NEXT: cmovbq %rbp, %rdx
1254 ; CHECK-NEXT: movq %rbp, %r13
1255 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1256 ; CHECK-NEXT: cmovaq %r15, %rdx
1257 ; CHECK-NEXT: cmovaq %r14, %rax
1258 ; CHECK-NEXT: movq $-1, %r14
1259 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1260 ; CHECK-NEXT: cmovpq %r12, %rax
1261 ; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1262 ; CHECK-NEXT: cmovpq %r12, %rdx
1263 ; CHECK-NEXT: movq %rdx, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1264 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1265 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[10,11,12,13,14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
1266 ; CHECK-NEXT: callq __extendhfsf2@PLT
1267 ; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill
1268 ; CHECK-NEXT: callq __fixsfti@PLT
1269 ; CHECK-NEXT: movq %rdx, %rbp
1270 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1271 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1272 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1273 ; CHECK-NEXT: cmovbq %r12, %rax
1274 ; CHECK-NEXT: cmovbq %r13, %rbp
1275 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1276 ; CHECK-NEXT: cmovaq %r15, %rbp
1277 ; CHECK-NEXT: movq %r15, %r13
1278 ; CHECK-NEXT: cmovaq %r14, %rax
1279 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1280 ; CHECK-NEXT: cmovpq %r12, %rax
1281 ; CHECK-NEXT: movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
1282 ; CHECK-NEXT: cmovpq %r12, %rbp
1283 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1284 ; CHECK-NEXT: shufps {{.*#+}} xmm0 = xmm0[3,3,3,3]
1285 ; CHECK-NEXT: callq __extendhfsf2@PLT
1286 ; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1287 ; CHECK-NEXT: callq __fixsfti@PLT
1288 ; CHECK-NEXT: movq %rax, %r14
1289 ; CHECK-NEXT: movq %rdx, %r15
1290 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1291 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1292 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1293 ; CHECK-NEXT: cmovbq %r12, %r14
1294 ; CHECK-NEXT: movabsq $-9223372036854775808, %rax # imm = 0x8000000000000000
1295 ; CHECK-NEXT: cmovbq %rax, %r15
1296 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1297 ; CHECK-NEXT: cmovaq %r13, %r15
1298 ; CHECK-NEXT: movq $-1, %rax
1299 ; CHECK-NEXT: cmovaq %rax, %r14
1300 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1301 ; CHECK-NEXT: cmovpq %r12, %r14
1302 ; CHECK-NEXT: cmovpq %r12, %r15
1303 ; CHECK-NEXT: movdqa {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1304 ; CHECK-NEXT: psrldq {{.*#+}} xmm0 = xmm0[14,15],zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero
1305 ; CHECK-NEXT: callq __extendhfsf2@PLT
1306 ; CHECK-NEXT: movd %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Folded Spill
1307 ; CHECK-NEXT: callq __fixsfti@PLT
1308 ; CHECK-NEXT: movq %rax, %r12
1309 ; CHECK-NEXT: movq %rdx, %r13
1310 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1311 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1312 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1313 ; CHECK-NEXT: movl $0, %eax
1314 ; CHECK-NEXT: cmovbq %rax, %r12
1315 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
1316 ; CHECK-NEXT: cmovbq %rcx, %r13
1317 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1318 ; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
1319 ; CHECK-NEXT: cmovaq %rcx, %r13
1320 ; CHECK-NEXT: movq $-1, %rcx
1321 ; CHECK-NEXT: cmovaq %rcx, %r12
1322 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1323 ; CHECK-NEXT: cmovpq %rax, %r12
1324 ; CHECK-NEXT: cmovpq %rax, %r13
1325 ; CHECK-NEXT: movaps {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 16-byte Reload
1326 ; CHECK-NEXT: callq __extendhfsf2@PLT
1327 ; CHECK-NEXT: movss %xmm0, {{[-0-9]+}}(%r{{[sb]}}p) # 4-byte Spill
1328 ; CHECK-NEXT: callq __fixsfti@PLT
1329 ; CHECK-NEXT: movss {{[-0-9]+}}(%r{{[sb]}}p), %xmm0 # 4-byte Reload
1330 ; CHECK-NEXT: # xmm0 = mem[0],zero,zero,zero
1331 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1332 ; CHECK-NEXT: movabsq $-9223372036854775808, %rcx # imm = 0x8000000000000000
1333 ; CHECK-NEXT: cmovbq %rcx, %rdx
1334 ; CHECK-NEXT: movl $0, %esi
1335 ; CHECK-NEXT: cmovbq %rsi, %rax
1336 ; CHECK-NEXT: ucomiss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0
1337 ; CHECK-NEXT: movabsq $9223372036854775807, %rcx # imm = 0x7FFFFFFFFFFFFFFF
1338 ; CHECK-NEXT: cmovaq %rcx, %rdx
1339 ; CHECK-NEXT: movq $-1, %rcx
1340 ; CHECK-NEXT: cmovaq %rcx, %rax
1341 ; CHECK-NEXT: ucomiss %xmm0, %xmm0
1342 ; CHECK-NEXT: cmovpq %rsi, %rax
1343 ; CHECK-NEXT: movl $0, %ecx
1344 ; CHECK-NEXT: cmovpq %rcx, %rdx
1345 ; CHECK-NEXT: movq %rdx, 8(%rbx)
1346 ; CHECK-NEXT: movq %rax, (%rbx)
1347 ; CHECK-NEXT: movq %r13, 120(%rbx)
1348 ; CHECK-NEXT: movq %r12, 112(%rbx)
1349 ; CHECK-NEXT: movq %r15, 104(%rbx)
1350 ; CHECK-NEXT: movq %r14, 96(%rbx)
1351 ; CHECK-NEXT: movq %rbp, 88(%rbx)
1352 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1353 ; CHECK-NEXT: movq %rax, 80(%rbx)
1354 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1355 ; CHECK-NEXT: movq %rax, 72(%rbx)
1356 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1357 ; CHECK-NEXT: movq %rax, 64(%rbx)
1358 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1359 ; CHECK-NEXT: movq %rax, 56(%rbx)
1360 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1361 ; CHECK-NEXT: movq %rax, 48(%rbx)
1362 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1363 ; CHECK-NEXT: movq %rax, 40(%rbx)
1364 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1365 ; CHECK-NEXT: movq %rax, 32(%rbx)
1366 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1367 ; CHECK-NEXT: movq %rax, 24(%rbx)
1368 ; CHECK-NEXT: movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
1369 ; CHECK-NEXT: movq %rax, 16(%rbx)
1370 ; CHECK-NEXT: movq %rbx, %rax
1371 ; CHECK-NEXT: addq $104, %rsp
1372 ; CHECK-NEXT: popq %rbx
1373 ; CHECK-NEXT: popq %r12
1374 ; CHECK-NEXT: popq %r13
1375 ; CHECK-NEXT: popq %r14
1376 ; CHECK-NEXT: popq %r15
1377 ; CHECK-NEXT: popq %rbp
1378 ; CHECK-NEXT: retq
1379 %x = call <8 x i128> @llvm.fptosi.sat.v8i128.v8f16(<8 x half> %f)
1380 ret <8 x i128> %x
1381 }