xref: /llvm-project/llvm/test/CodeGen/X86/smin.ll (revision 7b3bbd83c0c24087072ec5b22a76799ab31f87d5) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=x86_64-linux | FileCheck %s --check-prefixes=X64,SSE
3; RUN: llc < %s -mtriple=x86_64-linux -mattr=avx | FileCheck %s --check-prefixes=X64,AVX,AVX1
4; RUN: llc < %s -mtriple=x86_64-linux -mattr=avx2 | FileCheck %s --check-prefixes=X64,AVX,AVX2
5; RUN: llc < %s -mtriple=i686 -mattr=cmov | FileCheck %s --check-prefix=X86
6
7declare i8 @llvm.smin.i8(i8, i8)
8declare i16 @llvm.smin.i16(i16, i16)
9declare i24 @llvm.smin.i24(i24, i24)
10declare i32 @llvm.smin.i32(i32, i32)
11declare i64 @llvm.smin.i64(i64, i64)
12declare i128 @llvm.smin.i128(i128, i128)
13
14declare <1 x i32> @llvm.smin.v1i32(<1 x i32>, <1 x i32>)
15declare <2 x i32> @llvm.smin.v2i32(<2 x i32>, <2 x i32>)
16declare <3 x i32> @llvm.smin.v3i32(<3 x i32>, <3 x i32>)
17declare <4 x i32> @llvm.smin.v4i32(<4 x i32>, <4 x i32>)
18declare <8 x i32> @llvm.smin.v8i32(<8 x i32>, <8 x i32>)
19
20declare <8 x i16> @llvm.smin.v8i16(<8 x i16>, <8 x i16>)
21declare <16 x i8> @llvm.smin.v16i8(<16 x i8>, <16 x i8>)
22
23define i8 @test_i8(i8 %a, i8 %b) nounwind {
24; X64-LABEL: test_i8:
25; X64:       # %bb.0:
26; X64-NEXT:    movl %esi, %eax
27; X64-NEXT:    cmpb %al, %dil
28; X64-NEXT:    cmovll %edi, %eax
29; X64-NEXT:    # kill: def $al killed $al killed $eax
30; X64-NEXT:    retq
31;
32; X86-LABEL: test_i8:
33; X86:       # %bb.0:
34; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
35; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
36; X86-NEXT:    cmpb %al, %cl
37; X86-NEXT:    cmovll %ecx, %eax
38; X86-NEXT:    # kill: def $al killed $al killed $eax
39; X86-NEXT:    retl
40  %r = call i8 @llvm.smin.i8(i8 %a, i8 %b)
41  ret i8 %r
42}
43
44define i16 @test_i16(i16 %a, i16 %b) nounwind {
45; X64-LABEL: test_i16:
46; X64:       # %bb.0:
47; X64-NEXT:    movl %esi, %eax
48; X64-NEXT:    cmpw %ax, %di
49; X64-NEXT:    cmovll %edi, %eax
50; X64-NEXT:    # kill: def $ax killed $ax killed $eax
51; X64-NEXT:    retq
52;
53; X86-LABEL: test_i16:
54; X86:       # %bb.0:
55; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
56; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
57; X86-NEXT:    cmpw %ax, %cx
58; X86-NEXT:    cmovll %ecx, %eax
59; X86-NEXT:    # kill: def $ax killed $ax killed $eax
60; X86-NEXT:    retl
61  %r = call i16 @llvm.smin.i16(i16 %a, i16 %b)
62  ret i16 %r
63}
64
65define i24 @test_i24(i24 %a, i24 %b) nounwind {
66; X64-LABEL: test_i24:
67; X64:       # %bb.0:
68; X64-NEXT:    movl %edi, %eax
69; X64-NEXT:    shll $8, %esi
70; X64-NEXT:    sarl $8, %esi
71; X64-NEXT:    shll $8, %eax
72; X64-NEXT:    sarl $8, %eax
73; X64-NEXT:    cmpl %esi, %eax
74; X64-NEXT:    cmovgel %esi, %eax
75; X64-NEXT:    retq
76;
77; X86-LABEL: test_i24:
78; X86:       # %bb.0:
79; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
80; X86-NEXT:    shll $8, %ecx
81; X86-NEXT:    sarl $8, %ecx
82; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
83; X86-NEXT:    shll $8, %eax
84; X86-NEXT:    sarl $8, %eax
85; X86-NEXT:    cmpl %ecx, %eax
86; X86-NEXT:    cmovgel %ecx, %eax
87; X86-NEXT:    retl
88  %r = call i24 @llvm.smin.i24(i24 %a, i24 %b)
89  ret i24 %r
90}
91
92define i32 @test_i32(i32 %a, i32 %b) nounwind {
93; X64-LABEL: test_i32:
94; X64:       # %bb.0:
95; X64-NEXT:    movl %esi, %eax
96; X64-NEXT:    cmpl %esi, %edi
97; X64-NEXT:    cmovll %edi, %eax
98; X64-NEXT:    retq
99;
100; X86-LABEL: test_i32:
101; X86:       # %bb.0:
102; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
103; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
104; X86-NEXT:    cmpl %eax, %ecx
105; X86-NEXT:    cmovll %ecx, %eax
106; X86-NEXT:    retl
107  %r = call i32 @llvm.smin.i32(i32 %a, i32 %b)
108  ret i32 %r
109}
110
111define i64 @test_i64(i64 %a, i64 %b) nounwind {
112; X64-LABEL: test_i64:
113; X64:       # %bb.0:
114; X64-NEXT:    movq %rsi, %rax
115; X64-NEXT:    cmpq %rsi, %rdi
116; X64-NEXT:    cmovlq %rdi, %rax
117; X64-NEXT:    retq
118;
119; X86-LABEL: test_i64:
120; X86:       # %bb.0:
121; X86-NEXT:    pushl %edi
122; X86-NEXT:    pushl %esi
123; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
124; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
125; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
126; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
127; X86-NEXT:    cmpl %eax, %ecx
128; X86-NEXT:    movl %esi, %edi
129; X86-NEXT:    sbbl %edx, %edi
130; X86-NEXT:    cmovll %ecx, %eax
131; X86-NEXT:    cmovll %esi, %edx
132; X86-NEXT:    popl %esi
133; X86-NEXT:    popl %edi
134; X86-NEXT:    retl
135  %r = call i64 @llvm.smin.i64(i64 %a, i64 %b)
136  ret i64 %r
137}
138
139define i128 @test_i128(i128 %a, i128 %b) nounwind {
140; X64-LABEL: test_i128:
141; X64:       # %bb.0:
142; X64-NEXT:    movq %rdx, %rax
143; X64-NEXT:    cmpq %rdx, %rdi
144; X64-NEXT:    movq %rsi, %rdx
145; X64-NEXT:    sbbq %rcx, %rdx
146; X64-NEXT:    cmovlq %rdi, %rax
147; X64-NEXT:    cmovlq %rsi, %rcx
148; X64-NEXT:    movq %rcx, %rdx
149; X64-NEXT:    retq
150;
151; X86-LABEL: test_i128:
152; X86:       # %bb.0:
153; X86-NEXT:    pushl %ebp
154; X86-NEXT:    pushl %ebx
155; X86-NEXT:    pushl %edi
156; X86-NEXT:    pushl %esi
157; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
158; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
159; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
160; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
161; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
162; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
163; X86-NEXT:    cmpl %edx, %ebx
164; X86-NEXT:    sbbl %esi, %ebp
165; X86-NEXT:    movl %eax, %ebp
166; X86-NEXT:    sbbl %ecx, %ebp
167; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
168; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
169; X86-NEXT:    movl %edi, %eax
170; X86-NEXT:    sbbl %ebp, %eax
171; X86-NEXT:    cmovll %ebx, %edx
172; X86-NEXT:    cmovll {{[0-9]+}}(%esp), %esi
173; X86-NEXT:    cmovll {{[0-9]+}}(%esp), %ecx
174; X86-NEXT:    cmovll %edi, %ebp
175; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
176; X86-NEXT:    movl %ebp, 12(%eax)
177; X86-NEXT:    movl %ecx, 8(%eax)
178; X86-NEXT:    movl %esi, 4(%eax)
179; X86-NEXT:    movl %edx, (%eax)
180; X86-NEXT:    popl %esi
181; X86-NEXT:    popl %edi
182; X86-NEXT:    popl %ebx
183; X86-NEXT:    popl %ebp
184; X86-NEXT:    retl $4
185  %r = call i128 @llvm.smin.i128(i128 %a, i128 %b)
186  ret i128 %r
187}
188
189define <1 x i32> @test_v1i32(<1 x i32> %a, <1 x i32> %b) nounwind {
190; X64-LABEL: test_v1i32:
191; X64:       # %bb.0:
192; X64-NEXT:    movl %esi, %eax
193; X64-NEXT:    cmpl %esi, %edi
194; X64-NEXT:    cmovll %edi, %eax
195; X64-NEXT:    retq
196;
197; X86-LABEL: test_v1i32:
198; X86:       # %bb.0:
199; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
200; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
201; X86-NEXT:    cmpl %eax, %ecx
202; X86-NEXT:    cmovll %ecx, %eax
203; X86-NEXT:    retl
204  %r = call <1 x i32> @llvm.smin.v1i32(<1 x i32> %a, <1 x i32> %b)
205  ret <1 x i32> %r
206}
207
208define <2 x i32> @test_v2i32(<2 x i32> %a, <2 x i32> %b) nounwind {
209; SSE-LABEL: test_v2i32:
210; SSE:       # %bb.0:
211; SSE-NEXT:    movdqa %xmm1, %xmm2
212; SSE-NEXT:    pcmpgtd %xmm0, %xmm2
213; SSE-NEXT:    pand %xmm2, %xmm0
214; SSE-NEXT:    pandn %xmm1, %xmm2
215; SSE-NEXT:    por %xmm2, %xmm0
216; SSE-NEXT:    retq
217;
218; AVX-LABEL: test_v2i32:
219; AVX:       # %bb.0:
220; AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0
221; AVX-NEXT:    retq
222;
223; X86-LABEL: test_v2i32:
224; X86:       # %bb.0:
225; X86-NEXT:    pushl %esi
226; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
227; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
228; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
229; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
230; X86-NEXT:    cmpl %eax, %esi
231; X86-NEXT:    cmovll %esi, %eax
232; X86-NEXT:    cmpl %edx, %ecx
233; X86-NEXT:    cmovll %ecx, %edx
234; X86-NEXT:    popl %esi
235; X86-NEXT:    retl
236  %r = call <2 x i32> @llvm.smin.v2i32(<2 x i32> %a, <2 x i32> %b)
237  ret <2 x i32> %r
238}
239
240define <3 x i32> @test_v3i32(<3 x i32> %a, <3 x i32> %b) nounwind {
241; SSE-LABEL: test_v3i32:
242; SSE:       # %bb.0:
243; SSE-NEXT:    movdqa %xmm1, %xmm2
244; SSE-NEXT:    pcmpgtd %xmm0, %xmm2
245; SSE-NEXT:    pand %xmm2, %xmm0
246; SSE-NEXT:    pandn %xmm1, %xmm2
247; SSE-NEXT:    por %xmm2, %xmm0
248; SSE-NEXT:    retq
249;
250; AVX-LABEL: test_v3i32:
251; AVX:       # %bb.0:
252; AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0
253; AVX-NEXT:    retq
254;
255; X86-LABEL: test_v3i32:
256; X86:       # %bb.0:
257; X86-NEXT:    pushl %ebx
258; X86-NEXT:    pushl %edi
259; X86-NEXT:    pushl %esi
260; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
261; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
262; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
263; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
264; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
265; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
266; X86-NEXT:    cmpl %eax, %ebx
267; X86-NEXT:    cmovll %ebx, %eax
268; X86-NEXT:    cmpl %edx, %edi
269; X86-NEXT:    cmovll %edi, %edx
270; X86-NEXT:    cmpl %ecx, %esi
271; X86-NEXT:    cmovll %esi, %ecx
272; X86-NEXT:    popl %esi
273; X86-NEXT:    popl %edi
274; X86-NEXT:    popl %ebx
275; X86-NEXT:    retl
276  %r = call <3 x i32> @llvm.smin.v3i32(<3 x i32> %a, <3 x i32> %b)
277  ret <3 x i32> %r
278}
279
280define <4 x i32> @test_v4i32(<4 x i32> %a, <4 x i32> %b) nounwind {
281; SSE-LABEL: test_v4i32:
282; SSE:       # %bb.0:
283; SSE-NEXT:    movdqa %xmm1, %xmm2
284; SSE-NEXT:    pcmpgtd %xmm0, %xmm2
285; SSE-NEXT:    pand %xmm2, %xmm0
286; SSE-NEXT:    pandn %xmm1, %xmm2
287; SSE-NEXT:    por %xmm2, %xmm0
288; SSE-NEXT:    retq
289;
290; AVX-LABEL: test_v4i32:
291; AVX:       # %bb.0:
292; AVX-NEXT:    vpminsd %xmm1, %xmm0, %xmm0
293; AVX-NEXT:    retq
294;
295; X86-LABEL: test_v4i32:
296; X86:       # %bb.0:
297; X86-NEXT:    pushl %edi
298; X86-NEXT:    pushl %esi
299; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
300; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
301; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
302; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
303; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
304; X86-NEXT:    cmpl %edi, %eax
305; X86-NEXT:    cmovll %eax, %edi
306; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
307; X86-NEXT:    cmpl %esi, %eax
308; X86-NEXT:    cmovll %eax, %esi
309; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
310; X86-NEXT:    cmpl %edx, %eax
311; X86-NEXT:    cmovll %eax, %edx
312; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
313; X86-NEXT:    cmpl %ecx, %eax
314; X86-NEXT:    cmovll %eax, %ecx
315; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
316; X86-NEXT:    movl %ecx, 12(%eax)
317; X86-NEXT:    movl %edx, 8(%eax)
318; X86-NEXT:    movl %esi, 4(%eax)
319; X86-NEXT:    movl %edi, (%eax)
320; X86-NEXT:    popl %esi
321; X86-NEXT:    popl %edi
322; X86-NEXT:    retl $4
323  %r = call <4 x i32> @llvm.smin.v4i32(<4 x i32> %a, <4 x i32> %b)
324  ret <4 x i32> %r
325}
326
327define <8 x i32> @test_v8i32(<8 x i32> %a, <8 x i32> %b) nounwind {
328; SSE-LABEL: test_v8i32:
329; SSE:       # %bb.0:
330; SSE-NEXT:    movdqa %xmm2, %xmm4
331; SSE-NEXT:    pcmpgtd %xmm0, %xmm4
332; SSE-NEXT:    pand %xmm4, %xmm0
333; SSE-NEXT:    pandn %xmm2, %xmm4
334; SSE-NEXT:    por %xmm4, %xmm0
335; SSE-NEXT:    movdqa %xmm3, %xmm2
336; SSE-NEXT:    pcmpgtd %xmm1, %xmm2
337; SSE-NEXT:    pand %xmm2, %xmm1
338; SSE-NEXT:    pandn %xmm3, %xmm2
339; SSE-NEXT:    por %xmm2, %xmm1
340; SSE-NEXT:    retq
341;
342; AVX1-LABEL: test_v8i32:
343; AVX1:       # %bb.0:
344; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm2
345; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
346; AVX1-NEXT:    vpminsd %xmm2, %xmm3, %xmm2
347; AVX1-NEXT:    vpminsd %xmm1, %xmm0, %xmm0
348; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
349; AVX1-NEXT:    retq
350;
351; AVX2-LABEL: test_v8i32:
352; AVX2:       # %bb.0:
353; AVX2-NEXT:    vpminsd %ymm1, %ymm0, %ymm0
354; AVX2-NEXT:    retq
355;
356; X86-LABEL: test_v8i32:
357; X86:       # %bb.0:
358; X86-NEXT:    pushl %ebp
359; X86-NEXT:    pushl %ebx
360; X86-NEXT:    pushl %edi
361; X86-NEXT:    pushl %esi
362; X86-NEXT:    subl $8, %esp
363; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
364; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
365; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
366; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
367; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
368; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
369; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
370; X86-NEXT:    cmpl %ebp, %eax
371; X86-NEXT:    cmovll %eax, %ebp
372; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
373; X86-NEXT:    cmpl %ebx, %eax
374; X86-NEXT:    cmovll %eax, %ebx
375; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
376; X86-NEXT:    cmpl %edi, %eax
377; X86-NEXT:    cmovll %eax, %edi
378; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
379; X86-NEXT:    cmpl %esi, %eax
380; X86-NEXT:    cmovll %eax, %esi
381; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
382; X86-NEXT:    cmpl %edx, %eax
383; X86-NEXT:    cmovll %eax, %edx
384; X86-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
385; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
386; X86-NEXT:    cmpl %ecx, %eax
387; X86-NEXT:    cmovll %eax, %ecx
388; X86-NEXT:    movl %ecx, (%esp) # 4-byte Spill
389; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
390; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
391; X86-NEXT:    cmpl %edx, %eax
392; X86-NEXT:    cmovll %eax, %edx
393; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
394; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
395; X86-NEXT:    cmpl %eax, %ecx
396; X86-NEXT:    cmovll %ecx, %eax
397; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
398; X86-NEXT:    movl %eax, 28(%ecx)
399; X86-NEXT:    movl %edx, 24(%ecx)
400; X86-NEXT:    movl (%esp), %eax # 4-byte Reload
401; X86-NEXT:    movl %eax, 20(%ecx)
402; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
403; X86-NEXT:    movl %eax, 16(%ecx)
404; X86-NEXT:    movl %esi, 12(%ecx)
405; X86-NEXT:    movl %edi, 8(%ecx)
406; X86-NEXT:    movl %ebx, 4(%ecx)
407; X86-NEXT:    movl %ebp, (%ecx)
408; X86-NEXT:    movl %ecx, %eax
409; X86-NEXT:    addl $8, %esp
410; X86-NEXT:    popl %esi
411; X86-NEXT:    popl %edi
412; X86-NEXT:    popl %ebx
413; X86-NEXT:    popl %ebp
414; X86-NEXT:    retl $4
415  %r = call <8 x i32> @llvm.smin.v8i32(<8 x i32> %a, <8 x i32> %b)
416  ret <8 x i32> %r
417}
418
419define <8 x i16> @test_v8i16(<8 x i16> %a, <8 x i16> %b) nounwind {
420; SSE-LABEL: test_v8i16:
421; SSE:       # %bb.0:
422; SSE-NEXT:    pminsw %xmm1, %xmm0
423; SSE-NEXT:    retq
424;
425; AVX-LABEL: test_v8i16:
426; AVX:       # %bb.0:
427; AVX-NEXT:    vpminsw %xmm1, %xmm0, %xmm0
428; AVX-NEXT:    retq
429;
430; X86-LABEL: test_v8i16:
431; X86:       # %bb.0:
432; X86-NEXT:    pushl %ebp
433; X86-NEXT:    pushl %ebx
434; X86-NEXT:    pushl %edi
435; X86-NEXT:    pushl %esi
436; X86-NEXT:    subl $8, %esp
437; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
438; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
439; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
440; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
441; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
442; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebp
443; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
444; X86-NEXT:    cmpw %bp, %ax
445; X86-NEXT:    cmovll %eax, %ebp
446; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
447; X86-NEXT:    cmpw %bx, %ax
448; X86-NEXT:    cmovll %eax, %ebx
449; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
450; X86-NEXT:    cmpw %di, %ax
451; X86-NEXT:    cmovll %eax, %edi
452; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
453; X86-NEXT:    cmpw %si, %ax
454; X86-NEXT:    cmovll %eax, %esi
455; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
456; X86-NEXT:    cmpw %dx, %ax
457; X86-NEXT:    cmovll %eax, %edx
458; X86-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
459; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
460; X86-NEXT:    cmpw %cx, %ax
461; X86-NEXT:    cmovll %eax, %ecx
462; X86-NEXT:    movl %ecx, (%esp) # 4-byte Spill
463; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
464; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
465; X86-NEXT:    cmpw %dx, %ax
466; X86-NEXT:    cmovll %eax, %edx
467; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
468; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
469; X86-NEXT:    cmpw %ax, %cx
470; X86-NEXT:    cmovll %ecx, %eax
471; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
472; X86-NEXT:    movw %ax, 14(%ecx)
473; X86-NEXT:    movw %dx, 12(%ecx)
474; X86-NEXT:    movl (%esp), %eax # 4-byte Reload
475; X86-NEXT:    movw %ax, 10(%ecx)
476; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %eax # 4-byte Reload
477; X86-NEXT:    movw %ax, 8(%ecx)
478; X86-NEXT:    movw %si, 6(%ecx)
479; X86-NEXT:    movw %di, 4(%ecx)
480; X86-NEXT:    movw %bx, 2(%ecx)
481; X86-NEXT:    movw %bp, (%ecx)
482; X86-NEXT:    movl %ecx, %eax
483; X86-NEXT:    addl $8, %esp
484; X86-NEXT:    popl %esi
485; X86-NEXT:    popl %edi
486; X86-NEXT:    popl %ebx
487; X86-NEXT:    popl %ebp
488; X86-NEXT:    retl $4
489  %r = call <8 x i16> @llvm.smin.v8i16(<8 x i16> %a, <8 x i16> %b)
490  ret <8 x i16> %r
491}
492
493define <16 x i8> @test_v16i8(<16 x i8> %a, <16 x i8> %b) nounwind {
494; SSE-LABEL: test_v16i8:
495; SSE:       # %bb.0:
496; SSE-NEXT:    movdqa %xmm1, %xmm2
497; SSE-NEXT:    pcmpgtb %xmm0, %xmm2
498; SSE-NEXT:    pand %xmm2, %xmm0
499; SSE-NEXT:    pandn %xmm1, %xmm2
500; SSE-NEXT:    por %xmm2, %xmm0
501; SSE-NEXT:    retq
502;
503; AVX-LABEL: test_v16i8:
504; AVX:       # %bb.0:
505; AVX-NEXT:    vpminsb %xmm1, %xmm0, %xmm0
506; AVX-NEXT:    retq
507;
508; X86-LABEL: test_v16i8:
509; X86:       # %bb.0:
510; X86-NEXT:    pushl %ebp
511; X86-NEXT:    pushl %ebx
512; X86-NEXT:    pushl %edi
513; X86-NEXT:    pushl %esi
514; X86-NEXT:    subl $40, %esp
515; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
516; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
517; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
518; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
519; X86-NEXT:    cmpb %bl, %al
520; X86-NEXT:    cmovll %eax, %ebx
521; X86-NEXT:    movl %ebx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
522; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
523; X86-NEXT:    cmpb %dl, %al
524; X86-NEXT:    cmovll %eax, %edx
525; X86-NEXT:    movl %edx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
526; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
527; X86-NEXT:    cmpb %cl, %al
528; X86-NEXT:    cmovll %eax, %ecx
529; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
530; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
531; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
532; X86-NEXT:    cmpb %cl, %al
533; X86-NEXT:    cmovll %eax, %ecx
534; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
535; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
536; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
537; X86-NEXT:    cmpb %cl, %al
538; X86-NEXT:    cmovll %eax, %ecx
539; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
540; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
541; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
542; X86-NEXT:    cmpb %cl, %al
543; X86-NEXT:    cmovll %eax, %ecx
544; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
545; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
546; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
547; X86-NEXT:    cmpb %cl, %al
548; X86-NEXT:    cmovll %eax, %ecx
549; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
550; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
551; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
552; X86-NEXT:    cmpb %cl, %al
553; X86-NEXT:    cmovll %eax, %ecx
554; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
555; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
556; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
557; X86-NEXT:    cmpb %cl, %al
558; X86-NEXT:    cmovll %eax, %ecx
559; X86-NEXT:    movl %ecx, {{[-0-9]+}}(%e{{[sb]}}p) # 4-byte Spill
560; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
561; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
562; X86-NEXT:    cmpb %cl, %al
563; X86-NEXT:    cmovll %eax, %ecx
564; X86-NEXT:    movl %ecx, (%esp) # 4-byte Spill
565; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
566; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
567; X86-NEXT:    cmpb %cl, %al
568; X86-NEXT:    cmovll %eax, %ecx
569; X86-NEXT:    movl %ecx, %ebp
570; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
571; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
572; X86-NEXT:    cmpb %cl, %al
573; X86-NEXT:    cmovll %eax, %ecx
574; X86-NEXT:    movl %ecx, %edi
575; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
576; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
577; X86-NEXT:    cmpb %cl, %al
578; X86-NEXT:    cmovll %eax, %ecx
579; X86-NEXT:    movl %ecx, %esi
580; X86-NEXT:    movl {{[0-9]+}}(%esp), %ebx
581; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
582; X86-NEXT:    cmpb %bl, %al
583; X86-NEXT:    cmovll %eax, %ebx
584; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
585; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
586; X86-NEXT:    cmpb %dl, %al
587; X86-NEXT:    cmovll %eax, %edx
588; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
589; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
590; X86-NEXT:    cmpb %cl, %al
591; X86-NEXT:    cmovll %eax, %ecx
592; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
593; X86-NEXT:    movb %cl, 15(%eax)
594; X86-NEXT:    movb %dl, 14(%eax)
595; X86-NEXT:    movb %bl, 13(%eax)
596; X86-NEXT:    movl %esi, %ecx
597; X86-NEXT:    movb %cl, 12(%eax)
598; X86-NEXT:    movl %edi, %ecx
599; X86-NEXT:    movb %cl, 11(%eax)
600; X86-NEXT:    movl %ebp, %ecx
601; X86-NEXT:    movb %cl, 10(%eax)
602; X86-NEXT:    movl (%esp), %ecx # 4-byte Reload
603; X86-NEXT:    movb %cl, 9(%eax)
604; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
605; X86-NEXT:    movb %cl, 8(%eax)
606; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
607; X86-NEXT:    movb %cl, 7(%eax)
608; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
609; X86-NEXT:    movb %cl, 6(%eax)
610; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
611; X86-NEXT:    movb %cl, 5(%eax)
612; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
613; X86-NEXT:    movb %cl, 4(%eax)
614; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
615; X86-NEXT:    movb %cl, 3(%eax)
616; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
617; X86-NEXT:    movb %cl, 2(%eax)
618; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
619; X86-NEXT:    movb %cl, 1(%eax)
620; X86-NEXT:    movl {{[-0-9]+}}(%e{{[sb]}}p), %ecx # 4-byte Reload
621; X86-NEXT:    movb %cl, (%eax)
622; X86-NEXT:    addl $40, %esp
623; X86-NEXT:    popl %esi
624; X86-NEXT:    popl %edi
625; X86-NEXT:    popl %ebx
626; X86-NEXT:    popl %ebp
627; X86-NEXT:    retl $4
628  %r = call <16 x i8> @llvm.smin.v16i8(<16 x i8> %a, <16 x i8> %b)
629  ret <16 x i8> %r
630}
631
632define i16 @test_signbits_i16(i16 %a, i16 %b) nounwind {
633; X64-LABEL: test_signbits_i16:
634; X64:       # %bb.0:
635; X64-NEXT:    movswl %si, %eax
636; X64-NEXT:    movswl %di, %ecx
637; X64-NEXT:    shrl $15, %ecx
638; X64-NEXT:    shrl $8, %eax
639; X64-NEXT:    cmpw %ax, %cx
640; X64-NEXT:    cmovll %ecx, %eax
641; X64-NEXT:    # kill: def $ax killed $ax killed $eax
642; X64-NEXT:    retq
643;
644; X86-LABEL: test_signbits_i16:
645; X86:       # %bb.0:
646; X86-NEXT:    movsbl {{[0-9]+}}(%esp), %ecx
647; X86-NEXT:    movswl {{[0-9]+}}(%esp), %eax
648; X86-NEXT:    shrl $15, %eax
649; X86-NEXT:    cmpw %cx, %ax
650; X86-NEXT:    cmovgel %ecx, %eax
651; X86-NEXT:    # kill: def $ax killed $ax killed $eax
652; X86-NEXT:    retl
653  %ax = ashr i16 %a, 15
654  %bx = ashr i16 %b, 8
655  %r = call i16 @llvm.smin.i16(i16 %ax, i16 %bx)
656  ret i16 %r
657}
658
659define i32 @test_signbits_i32(i32 %a, i32 %b) nounwind {
660; X64-LABEL: test_signbits_i32:
661; X64:       # %bb.0:
662; X64-NEXT:    movl %esi, %eax
663; X64-NEXT:    sarl $16, %edi
664; X64-NEXT:    sarl $17, %eax
665; X64-NEXT:    cmpl %eax, %edi
666; X64-NEXT:    cmovll %edi, %eax
667; X64-NEXT:    retq
668;
669; X86-LABEL: test_signbits_i32:
670; X86:       # %bb.0:
671; X86-NEXT:    movswl {{[0-9]+}}(%esp), %ecx
672; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
673; X86-NEXT:    sarl $17, %eax
674; X86-NEXT:    cmpl %eax, %ecx
675; X86-NEXT:    cmovll %ecx, %eax
676; X86-NEXT:    retl
677  %ax = ashr i32 %a, 16
678  %bx = ashr i32 %b, 17
679  %r = call i32 @llvm.smin.i32(i32 %ax, i32 %bx)
680  ret i32 %r
681}
682
683define i64 @test_signbits_i64(i64 %a, i64 %b) nounwind {
684; X64-LABEL: test_signbits_i64:
685; X64:       # %bb.0:
686; X64-NEXT:    movq %rsi, %rax
687; X64-NEXT:    sarq $32, %rdi
688; X64-NEXT:    sarq $32, %rax
689; X64-NEXT:    cmpq %rax, %rdi
690; X64-NEXT:    cmovlq %rdi, %rax
691; X64-NEXT:    retq
692;
693; X86-LABEL: test_signbits_i64:
694; X86:       # %bb.0:
695; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
696; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
697; X86-NEXT:    cmpl %eax, %ecx
698; X86-NEXT:    cmovll %ecx, %eax
699; X86-NEXT:    movl %eax, %edx
700; X86-NEXT:    sarl $31, %edx
701; X86-NEXT:    retl
702  %ax = ashr i64 %a, 32
703  %bx = ashr i64 %b, 32
704  %r = call i64 @llvm.smin.i64(i64 %ax, i64 %bx)
705  ret i64 %r
706}
707
708define i128 @test_signbits_i128(i128 %a, i128 %b) nounwind {
709; X64-LABEL: test_signbits_i128:
710; X64:       # %bb.0:
711; X64-NEXT:    movq %rcx, %rax
712; X64-NEXT:    sarq $28, %rax
713; X64-NEXT:    cmpq %rax, %rsi
714; X64-NEXT:    cmovlq %rsi, %rax
715; X64-NEXT:    movq %rax, %rdx
716; X64-NEXT:    sarq $63, %rdx
717; X64-NEXT:    retq
718;
719; X86-LABEL: test_signbits_i128:
720; X86:       # %bb.0:
721; X86-NEXT:    pushl %ebx
722; X86-NEXT:    pushl %edi
723; X86-NEXT:    pushl %esi
724; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
725; X86-NEXT:    movl {{[0-9]+}}(%esp), %esi
726; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
727; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
728; X86-NEXT:    movl {{[0-9]+}}(%esp), %edi
729; X86-NEXT:    shrdl $28, %edi, %ecx
730; X86-NEXT:    sarl $28, %edi
731; X86-NEXT:    cmpl %ecx, %esi
732; X86-NEXT:    movl %edx, %ebx
733; X86-NEXT:    sbbl %edi, %ebx
734; X86-NEXT:    cmovll %esi, %ecx
735; X86-NEXT:    cmovll %edx, %edi
736; X86-NEXT:    movl %edi, 4(%eax)
737; X86-NEXT:    sarl $31, %edi
738; X86-NEXT:    movl %ecx, (%eax)
739; X86-NEXT:    movl %edi, 12(%eax)
740; X86-NEXT:    movl %edi, 8(%eax)
741; X86-NEXT:    popl %esi
742; X86-NEXT:    popl %edi
743; X86-NEXT:    popl %ebx
744; X86-NEXT:    retl $4
745  %ax = ashr i128 %a, 64
746  %bx = ashr i128 %b, 92
747  %r = call i128 @llvm.smin.i128(i128 %ax, i128 %bx)
748  ret i128 %r
749}
750