xref: /llvm-project/llvm/test/CodeGen/X86/fshl.ll (revision 7b3bbd83c0c24087072ec5b22a76799ab31f87d5) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc < %s -mtriple=i686-unknown-unknown | FileCheck %s --check-prefixes=X86,X86-FAST
3; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+slow-shld | FileCheck %s --check-prefixes=X86,X86-SLOW
4; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=X64,X64-FAST
5; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+slow-shld | FileCheck %s --check-prefixes=X64,X64-SLOW
6
7declare i8 @llvm.fshl.i8(i8, i8, i8) nounwind readnone
8declare i16 @llvm.fshl.i16(i16, i16, i16) nounwind readnone
9declare i32 @llvm.fshl.i32(i32, i32, i32) nounwind readnone
10declare i64 @llvm.fshl.i64(i64, i64, i64) nounwind readnone
11declare i128 @llvm.fshl.i128(i128, i128, i128) nounwind readnone
12
13;
14; Variable Funnel Shift
15;
16
17define i8 @var_shift_i8(i8 %x, i8 %y, i8 %z) nounwind {
18; X86-LABEL: var_shift_i8:
19; X86:       # %bb.0:
20; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
21; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %edx
22; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
23; X86-NEXT:    shll $8, %eax
24; X86-NEXT:    orl %edx, %eax
25; X86-NEXT:    andb $7, %cl
26; X86-NEXT:    shll %cl, %eax
27; X86-NEXT:    movb %ah, %al
28; X86-NEXT:    retl
29;
30; X64-LABEL: var_shift_i8:
31; X64:       # %bb.0:
32; X64-NEXT:    movl %edx, %ecx
33; X64-NEXT:    shll $8, %edi
34; X64-NEXT:    movzbl %sil, %eax
35; X64-NEXT:    orl %edi, %eax
36; X64-NEXT:    andb $7, %cl
37; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
38; X64-NEXT:    shll %cl, %eax
39; X64-NEXT:    shrl $8, %eax
40; X64-NEXT:    # kill: def $al killed $al killed $eax
41; X64-NEXT:    retq
42  %tmp = tail call i8 @llvm.fshl.i8(i8 %x, i8 %y, i8 %z)
43  ret i8 %tmp
44}
45
46define i16 @var_shift_i16(i16 %x, i16 %y, i16 %z) nounwind {
47; X86-FAST-LABEL: var_shift_i16:
48; X86-FAST:       # %bb.0:
49; X86-FAST-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
50; X86-FAST-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
51; X86-FAST-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
52; X86-FAST-NEXT:    andb $15, %cl
53; X86-FAST-NEXT:    shldw %cl, %dx, %ax
54; X86-FAST-NEXT:    retl
55;
56; X86-SLOW-LABEL: var_shift_i16:
57; X86-SLOW:       # %bb.0:
58; X86-SLOW-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
59; X86-SLOW-NEXT:    movzwl {{[0-9]+}}(%esp), %edx
60; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
61; X86-SLOW-NEXT:    shll $16, %eax
62; X86-SLOW-NEXT:    orl %edx, %eax
63; X86-SLOW-NEXT:    andb $15, %cl
64; X86-SLOW-NEXT:    shll %cl, %eax
65; X86-SLOW-NEXT:    shrl $16, %eax
66; X86-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
67; X86-SLOW-NEXT:    retl
68;
69; X64-FAST-LABEL: var_shift_i16:
70; X64-FAST:       # %bb.0:
71; X64-FAST-NEXT:    movl %edx, %ecx
72; X64-FAST-NEXT:    movl %edi, %eax
73; X64-FAST-NEXT:    andb $15, %cl
74; X64-FAST-NEXT:    # kill: def $cl killed $cl killed $ecx
75; X64-FAST-NEXT:    shldw %cl, %si, %ax
76; X64-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
77; X64-FAST-NEXT:    retq
78;
79; X64-SLOW-LABEL: var_shift_i16:
80; X64-SLOW:       # %bb.0:
81; X64-SLOW-NEXT:    movl %edx, %ecx
82; X64-SLOW-NEXT:    shll $16, %edi
83; X64-SLOW-NEXT:    movzwl %si, %eax
84; X64-SLOW-NEXT:    orl %edi, %eax
85; X64-SLOW-NEXT:    andb $15, %cl
86; X64-SLOW-NEXT:    # kill: def $cl killed $cl killed $ecx
87; X64-SLOW-NEXT:    shll %cl, %eax
88; X64-SLOW-NEXT:    shrl $16, %eax
89; X64-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
90; X64-SLOW-NEXT:    retq
91  %tmp = tail call i16 @llvm.fshl.i16(i16 %x, i16 %y, i16 %z)
92  ret i16 %tmp
93}
94
95define i32 @var_shift_i32(i32 %x, i32 %y, i32 %z) nounwind {
96; X86-FAST-LABEL: var_shift_i32:
97; X86-FAST:       # %bb.0:
98; X86-FAST-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
99; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edx
100; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %eax
101; X86-FAST-NEXT:    shldl %cl, %edx, %eax
102; X86-FAST-NEXT:    retl
103;
104; X86-SLOW-LABEL: var_shift_i32:
105; X86-SLOW:       # %bb.0:
106; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
107; X86-SLOW-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
108; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %edx
109; X86-SLOW-NEXT:    shll %cl, %edx
110; X86-SLOW-NEXT:    notb %cl
111; X86-SLOW-NEXT:    shrl %eax
112; X86-SLOW-NEXT:    shrl %cl, %eax
113; X86-SLOW-NEXT:    orl %edx, %eax
114; X86-SLOW-NEXT:    retl
115;
116; X64-FAST-LABEL: var_shift_i32:
117; X64-FAST:       # %bb.0:
118; X64-FAST-NEXT:    movl %edx, %ecx
119; X64-FAST-NEXT:    movl %edi, %eax
120; X64-FAST-NEXT:    # kill: def $cl killed $cl killed $ecx
121; X64-FAST-NEXT:    shldl %cl, %esi, %eax
122; X64-FAST-NEXT:    retq
123;
124; X64-SLOW-LABEL: var_shift_i32:
125; X64-SLOW:       # %bb.0:
126; X64-SLOW-NEXT:    movl %edx, %ecx
127; X64-SLOW-NEXT:    movl %esi, %eax
128; X64-SLOW-NEXT:    shll %cl, %edi
129; X64-SLOW-NEXT:    shrl %eax
130; X64-SLOW-NEXT:    notb %cl
131; X64-SLOW-NEXT:    # kill: def $cl killed $cl killed $ecx
132; X64-SLOW-NEXT:    shrl %cl, %eax
133; X64-SLOW-NEXT:    orl %edi, %eax
134; X64-SLOW-NEXT:    retq
135  %tmp = tail call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %z)
136  ret i32 %tmp
137}
138
139define i32 @var_shift_i32_optsize(i32 %x, i32 %y, i32 %z) nounwind optsize {
140; X86-LABEL: var_shift_i32_optsize:
141; X86:       # %bb.0:
142; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
143; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
144; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
145; X86-NEXT:    shldl %cl, %edx, %eax
146; X86-NEXT:    retl
147;
148; X64-LABEL: var_shift_i32_optsize:
149; X64:       # %bb.0:
150; X64-NEXT:    movl %edx, %ecx
151; X64-NEXT:    movl %edi, %eax
152; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
153; X64-NEXT:    shldl %cl, %esi, %eax
154; X64-NEXT:    retq
155  %tmp = tail call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %z)
156  ret i32 %tmp
157}
158
159define i32 @var_shift_i32_pgso(i32 %x, i32 %y, i32 %z) nounwind !prof !14 {
160; X86-LABEL: var_shift_i32_pgso:
161; X86:       # %bb.0:
162; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
163; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
164; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
165; X86-NEXT:    shldl %cl, %edx, %eax
166; X86-NEXT:    retl
167;
168; X64-LABEL: var_shift_i32_pgso:
169; X64:       # %bb.0:
170; X64-NEXT:    movl %edx, %ecx
171; X64-NEXT:    movl %edi, %eax
172; X64-NEXT:    # kill: def $cl killed $cl killed $ecx
173; X64-NEXT:    shldl %cl, %esi, %eax
174; X64-NEXT:    retq
175  %tmp = tail call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 %z)
176  ret i32 %tmp
177}
178
179define i64 @var_shift_i64(i64 %x, i64 %y, i64 %z) nounwind {
180; X86-FAST-LABEL: var_shift_i64:
181; X86-FAST:       # %bb.0:
182; X86-FAST-NEXT:    pushl %edi
183; X86-FAST-NEXT:    pushl %esi
184; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %esi
185; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edx
186; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ecx
187; X86-FAST-NEXT:    testb $32, %cl
188; X86-FAST-NEXT:    jne .LBB5_1
189; X86-FAST-NEXT:  # %bb.2:
190; X86-FAST-NEXT:    movl %edx, %edi
191; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edx
192; X86-FAST-NEXT:    jmp .LBB5_3
193; X86-FAST-NEXT:  .LBB5_1:
194; X86-FAST-NEXT:    movl %esi, %edi
195; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %esi
196; X86-FAST-NEXT:  .LBB5_3:
197; X86-FAST-NEXT:    movl %edi, %eax
198; X86-FAST-NEXT:    shldl %cl, %esi, %eax
199; X86-FAST-NEXT:    # kill: def $cl killed $cl killed $ecx
200; X86-FAST-NEXT:    shldl %cl, %edi, %edx
201; X86-FAST-NEXT:    popl %esi
202; X86-FAST-NEXT:    popl %edi
203; X86-FAST-NEXT:    retl
204;
205; X86-SLOW-LABEL: var_shift_i64:
206; X86-SLOW:       # %bb.0:
207; X86-SLOW-NEXT:    pushl %ebx
208; X86-SLOW-NEXT:    pushl %edi
209; X86-SLOW-NEXT:    pushl %esi
210; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
211; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %edx
212; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ebx
213; X86-SLOW-NEXT:    testb $32, %bl
214; X86-SLOW-NEXT:    jne .LBB5_1
215; X86-SLOW-NEXT:  # %bb.2:
216; X86-SLOW-NEXT:    movl %edx, %esi
217; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %edx
218; X86-SLOW-NEXT:    jmp .LBB5_3
219; X86-SLOW-NEXT:  .LBB5_1:
220; X86-SLOW-NEXT:    movl %eax, %esi
221; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
222; X86-SLOW-NEXT:  .LBB5_3:
223; X86-SLOW-NEXT:    movl %esi, %edi
224; X86-SLOW-NEXT:    movl %ebx, %ecx
225; X86-SLOW-NEXT:    shll %cl, %edi
226; X86-SLOW-NEXT:    shrl %eax
227; X86-SLOW-NEXT:    notb %cl
228; X86-SLOW-NEXT:    shrl %cl, %eax
229; X86-SLOW-NEXT:    orl %edi, %eax
230; X86-SLOW-NEXT:    shrl %esi
231; X86-SLOW-NEXT:    shrl %cl, %esi
232; X86-SLOW-NEXT:    movl %ebx, %ecx
233; X86-SLOW-NEXT:    shll %cl, %edx
234; X86-SLOW-NEXT:    orl %esi, %edx
235; X86-SLOW-NEXT:    popl %esi
236; X86-SLOW-NEXT:    popl %edi
237; X86-SLOW-NEXT:    popl %ebx
238; X86-SLOW-NEXT:    retl
239;
240; X64-FAST-LABEL: var_shift_i64:
241; X64-FAST:       # %bb.0:
242; X64-FAST-NEXT:    movq %rdx, %rcx
243; X64-FAST-NEXT:    movq %rdi, %rax
244; X64-FAST-NEXT:    # kill: def $cl killed $cl killed $rcx
245; X64-FAST-NEXT:    shldq %cl, %rsi, %rax
246; X64-FAST-NEXT:    retq
247;
248; X64-SLOW-LABEL: var_shift_i64:
249; X64-SLOW:       # %bb.0:
250; X64-SLOW-NEXT:    movq %rdx, %rcx
251; X64-SLOW-NEXT:    movq %rsi, %rax
252; X64-SLOW-NEXT:    shlq %cl, %rdi
253; X64-SLOW-NEXT:    shrq %rax
254; X64-SLOW-NEXT:    notb %cl
255; X64-SLOW-NEXT:    # kill: def $cl killed $cl killed $rcx
256; X64-SLOW-NEXT:    shrq %cl, %rax
257; X64-SLOW-NEXT:    orq %rdi, %rax
258; X64-SLOW-NEXT:    retq
259  %tmp = tail call i64 @llvm.fshl.i64(i64 %x, i64 %y, i64 %z)
260  ret i64 %tmp
261}
262
263define i128 @var_shift_i128(i128 %x, i128 %y, i128 %z) nounwind {
264; X86-FAST-LABEL: var_shift_i128:
265; X86-FAST:       # %bb.0:
266; X86-FAST-NEXT:    pushl %ebp
267; X86-FAST-NEXT:    pushl %ebx
268; X86-FAST-NEXT:    pushl %edi
269; X86-FAST-NEXT:    pushl %esi
270; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %esi
271; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edx
272; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ebx
273; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edi
274; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ecx
275; X86-FAST-NEXT:    testb $64, %cl
276; X86-FAST-NEXT:    jne .LBB6_1
277; X86-FAST-NEXT:  # %bb.2:
278; X86-FAST-NEXT:    movl %ebx, %ebp
279; X86-FAST-NEXT:    movl %esi, %ebx
280; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %esi
281; X86-FAST-NEXT:    movl %edi, %eax
282; X86-FAST-NEXT:    movl %edx, %edi
283; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edx
284; X86-FAST-NEXT:    testb $32, %cl
285; X86-FAST-NEXT:    je .LBB6_5
286; X86-FAST-NEXT:  .LBB6_4:
287; X86-FAST-NEXT:    movl %esi, %edx
288; X86-FAST-NEXT:    movl %edi, %esi
289; X86-FAST-NEXT:    movl %ebx, %edi
290; X86-FAST-NEXT:    movl %eax, %ebx
291; X86-FAST-NEXT:    jmp .LBB6_6
292; X86-FAST-NEXT:  .LBB6_1:
293; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ebp
294; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %eax
295; X86-FAST-NEXT:    testb $32, %cl
296; X86-FAST-NEXT:    jne .LBB6_4
297; X86-FAST-NEXT:  .LBB6_5:
298; X86-FAST-NEXT:    movl %eax, %ebp
299; X86-FAST-NEXT:  .LBB6_6:
300; X86-FAST-NEXT:    movl %ebx, %eax
301; X86-FAST-NEXT:    shldl %cl, %ebp, %eax
302; X86-FAST-NEXT:    movl %edi, %ebp
303; X86-FAST-NEXT:    shldl %cl, %ebx, %ebp
304; X86-FAST-NEXT:    movl %esi, %ebx
305; X86-FAST-NEXT:    shldl %cl, %edi, %ebx
306; X86-FAST-NEXT:    # kill: def $cl killed $cl killed $ecx
307; X86-FAST-NEXT:    shldl %cl, %esi, %edx
308; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ecx
309; X86-FAST-NEXT:    movl %edx, 12(%ecx)
310; X86-FAST-NEXT:    movl %ebx, 8(%ecx)
311; X86-FAST-NEXT:    movl %ebp, 4(%ecx)
312; X86-FAST-NEXT:    movl %eax, (%ecx)
313; X86-FAST-NEXT:    movl %ecx, %eax
314; X86-FAST-NEXT:    popl %esi
315; X86-FAST-NEXT:    popl %edi
316; X86-FAST-NEXT:    popl %ebx
317; X86-FAST-NEXT:    popl %ebp
318; X86-FAST-NEXT:    retl $4
319;
320; X86-SLOW-LABEL: var_shift_i128:
321; X86-SLOW:       # %bb.0:
322; X86-SLOW-NEXT:    pushl %ebp
323; X86-SLOW-NEXT:    pushl %ebx
324; X86-SLOW-NEXT:    pushl %edi
325; X86-SLOW-NEXT:    pushl %esi
326; X86-SLOW-NEXT:    pushl %eax
327; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ebx
328; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %esi
329; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %edx
330; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %edi
331; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
332; X86-SLOW-NEXT:    testb $64, %al
333; X86-SLOW-NEXT:    jne .LBB6_1
334; X86-SLOW-NEXT:  # %bb.2:
335; X86-SLOW-NEXT:    movl %edx, %ebp
336; X86-SLOW-NEXT:    movl %ebx, %edx
337; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ebx
338; X86-SLOW-NEXT:    movl %edi, %ecx
339; X86-SLOW-NEXT:    movl %esi, %edi
340; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %esi
341; X86-SLOW-NEXT:    testb $32, %al
342; X86-SLOW-NEXT:    je .LBB6_5
343; X86-SLOW-NEXT:  .LBB6_4:
344; X86-SLOW-NEXT:    movl %ebx, (%esp) # 4-byte Spill
345; X86-SLOW-NEXT:    movl %edi, %ebx
346; X86-SLOW-NEXT:    movl %edx, %edi
347; X86-SLOW-NEXT:    movl %ecx, %edx
348; X86-SLOW-NEXT:    jmp .LBB6_6
349; X86-SLOW-NEXT:  .LBB6_1:
350; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ebp
351; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ecx
352; X86-SLOW-NEXT:    testb $32, %al
353; X86-SLOW-NEXT:    jne .LBB6_4
354; X86-SLOW-NEXT:  .LBB6_5:
355; X86-SLOW-NEXT:    movl %ecx, %ebp
356; X86-SLOW-NEXT:    movl %esi, (%esp) # 4-byte Spill
357; X86-SLOW-NEXT:  .LBB6_6:
358; X86-SLOW-NEXT:    movl %edx, %esi
359; X86-SLOW-NEXT:    movl %eax, %ecx
360; X86-SLOW-NEXT:    shll %cl, %esi
361; X86-SLOW-NEXT:    shrl %ebp
362; X86-SLOW-NEXT:    movb %al, %ch
363; X86-SLOW-NEXT:    notb %ch
364; X86-SLOW-NEXT:    movb %ch, %cl
365; X86-SLOW-NEXT:    shrl %cl, %ebp
366; X86-SLOW-NEXT:    orl %esi, %ebp
367; X86-SLOW-NEXT:    movl %edi, %esi
368; X86-SLOW-NEXT:    movb %al, %cl
369; X86-SLOW-NEXT:    shll %cl, %esi
370; X86-SLOW-NEXT:    shrl %edx
371; X86-SLOW-NEXT:    movb %ch, %cl
372; X86-SLOW-NEXT:    shrl %cl, %edx
373; X86-SLOW-NEXT:    orl %esi, %edx
374; X86-SLOW-NEXT:    movl %ebx, %esi
375; X86-SLOW-NEXT:    movb %al, %cl
376; X86-SLOW-NEXT:    shll %cl, %esi
377; X86-SLOW-NEXT:    shrl %edi
378; X86-SLOW-NEXT:    movb %ch, %cl
379; X86-SLOW-NEXT:    shrl %cl, %edi
380; X86-SLOW-NEXT:    orl %esi, %edi
381; X86-SLOW-NEXT:    movb %al, %cl
382; X86-SLOW-NEXT:    movl (%esp), %eax # 4-byte Reload
383; X86-SLOW-NEXT:    shll %cl, %eax
384; X86-SLOW-NEXT:    shrl %ebx
385; X86-SLOW-NEXT:    movb %ch, %cl
386; X86-SLOW-NEXT:    shrl %cl, %ebx
387; X86-SLOW-NEXT:    orl %eax, %ebx
388; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
389; X86-SLOW-NEXT:    movl %ebx, 12(%eax)
390; X86-SLOW-NEXT:    movl %edi, 8(%eax)
391; X86-SLOW-NEXT:    movl %edx, 4(%eax)
392; X86-SLOW-NEXT:    movl %ebp, (%eax)
393; X86-SLOW-NEXT:    addl $4, %esp
394; X86-SLOW-NEXT:    popl %esi
395; X86-SLOW-NEXT:    popl %edi
396; X86-SLOW-NEXT:    popl %ebx
397; X86-SLOW-NEXT:    popl %ebp
398; X86-SLOW-NEXT:    retl $4
399;
400; X64-FAST-LABEL: var_shift_i128:
401; X64-FAST:       # %bb.0:
402; X64-FAST-NEXT:    testb $64, %r8b
403; X64-FAST-NEXT:    cmovneq %rdi, %rsi
404; X64-FAST-NEXT:    cmoveq %rcx, %rdx
405; X64-FAST-NEXT:    cmovneq %rcx, %rdi
406; X64-FAST-NEXT:    movq %rdi, %rax
407; X64-FAST-NEXT:    movl %r8d, %ecx
408; X64-FAST-NEXT:    shldq %cl, %rdx, %rax
409; X64-FAST-NEXT:    shldq %cl, %rdi, %rsi
410; X64-FAST-NEXT:    movq %rsi, %rdx
411; X64-FAST-NEXT:    retq
412;
413; X64-SLOW-LABEL: var_shift_i128:
414; X64-SLOW:       # %bb.0:
415; X64-SLOW-NEXT:    testb $64, %r8b
416; X64-SLOW-NEXT:    cmovneq %rdi, %rsi
417; X64-SLOW-NEXT:    cmoveq %rcx, %rdx
418; X64-SLOW-NEXT:    cmovneq %rcx, %rdi
419; X64-SLOW-NEXT:    movq %rdi, %rax
420; X64-SLOW-NEXT:    movl %r8d, %ecx
421; X64-SLOW-NEXT:    shlq %cl, %rax
422; X64-SLOW-NEXT:    shrq %rdx
423; X64-SLOW-NEXT:    movl %r8d, %r9d
424; X64-SLOW-NEXT:    notb %r9b
425; X64-SLOW-NEXT:    movl %r9d, %ecx
426; X64-SLOW-NEXT:    shrq %cl, %rdx
427; X64-SLOW-NEXT:    orq %rdx, %rax
428; X64-SLOW-NEXT:    movl %r8d, %ecx
429; X64-SLOW-NEXT:    shlq %cl, %rsi
430; X64-SLOW-NEXT:    shrq %rdi
431; X64-SLOW-NEXT:    movl %r9d, %ecx
432; X64-SLOW-NEXT:    shrq %cl, %rdi
433; X64-SLOW-NEXT:    orq %rsi, %rdi
434; X64-SLOW-NEXT:    movq %rdi, %rdx
435; X64-SLOW-NEXT:    retq
436  %tmp = tail call i128 @llvm.fshl.i128(i128 %x, i128 %y, i128 %z)
437  ret i128 %tmp
438}
439
440;
441; Const Funnel Shift
442;
443
444define i8 @const_shift_i8(i8 %x, i8 %y) nounwind {
445; X86-LABEL: const_shift_i8:
446; X86:       # %bb.0:
447; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %eax
448; X86-NEXT:    movzbl {{[0-9]+}}(%esp), %ecx
449; X86-NEXT:    shrb %cl
450; X86-NEXT:    shlb $7, %al
451; X86-NEXT:    orb %cl, %al
452; X86-NEXT:    retl
453;
454; X64-LABEL: const_shift_i8:
455; X64:       # %bb.0:
456; X64-NEXT:    # kill: def $esi killed $esi def $rsi
457; X64-NEXT:    # kill: def $edi killed $edi def $rdi
458; X64-NEXT:    shrb %sil
459; X64-NEXT:    shlb $7, %dil
460; X64-NEXT:    leal (%rdi,%rsi), %eax
461; X64-NEXT:    # kill: def $al killed $al killed $eax
462; X64-NEXT:    retq
463  %tmp = tail call i8 @llvm.fshl.i8(i8 %x, i8 %y, i8 7)
464  ret i8 %tmp
465}
466
467define i16 @const_shift_i16(i16 %x, i16 %y) nounwind {
468; X86-FAST-LABEL: const_shift_i16:
469; X86-FAST:       # %bb.0:
470; X86-FAST-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
471; X86-FAST-NEXT:    movzwl {{[0-9]+}}(%esp), %eax
472; X86-FAST-NEXT:    shldw $7, %cx, %ax
473; X86-FAST-NEXT:    retl
474;
475; X86-SLOW-LABEL: const_shift_i16:
476; X86-SLOW:       # %bb.0:
477; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
478; X86-SLOW-NEXT:    movzwl {{[0-9]+}}(%esp), %ecx
479; X86-SLOW-NEXT:    shrl $9, %ecx
480; X86-SLOW-NEXT:    shll $7, %eax
481; X86-SLOW-NEXT:    orl %ecx, %eax
482; X86-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
483; X86-SLOW-NEXT:    retl
484;
485; X64-FAST-LABEL: const_shift_i16:
486; X64-FAST:       # %bb.0:
487; X64-FAST-NEXT:    movl %edi, %eax
488; X64-FAST-NEXT:    shldw $7, %si, %ax
489; X64-FAST-NEXT:    # kill: def $ax killed $ax killed $eax
490; X64-FAST-NEXT:    retq
491;
492; X64-SLOW-LABEL: const_shift_i16:
493; X64-SLOW:       # %bb.0:
494; X64-SLOW-NEXT:    movzwl %si, %eax
495; X64-SLOW-NEXT:    shll $7, %edi
496; X64-SLOW-NEXT:    shrl $9, %eax
497; X64-SLOW-NEXT:    orl %edi, %eax
498; X64-SLOW-NEXT:    # kill: def $ax killed $ax killed $eax
499; X64-SLOW-NEXT:    retq
500  %tmp = tail call i16 @llvm.fshl.i16(i16 %x, i16 %y, i16 7)
501  ret i16 %tmp
502}
503
504define i32 @const_shift_i32(i32 %x, i32 %y) nounwind {
505; X86-FAST-LABEL: const_shift_i32:
506; X86-FAST:       # %bb.0:
507; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ecx
508; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %eax
509; X86-FAST-NEXT:    shldl $7, %ecx, %eax
510; X86-FAST-NEXT:    retl
511;
512; X86-SLOW-LABEL: const_shift_i32:
513; X86-SLOW:       # %bb.0:
514; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %eax
515; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ecx
516; X86-SLOW-NEXT:    shrl $25, %ecx
517; X86-SLOW-NEXT:    shll $7, %eax
518; X86-SLOW-NEXT:    orl %ecx, %eax
519; X86-SLOW-NEXT:    retl
520;
521; X64-FAST-LABEL: const_shift_i32:
522; X64-FAST:       # %bb.0:
523; X64-FAST-NEXT:    movl %edi, %eax
524; X64-FAST-NEXT:    shldl $7, %esi, %eax
525; X64-FAST-NEXT:    retq
526;
527; X64-SLOW-LABEL: const_shift_i32:
528; X64-SLOW:       # %bb.0:
529; X64-SLOW-NEXT:    # kill: def $esi killed $esi def $rsi
530; X64-SLOW-NEXT:    # kill: def $edi killed $edi def $rdi
531; X64-SLOW-NEXT:    shrl $25, %esi
532; X64-SLOW-NEXT:    shll $7, %edi
533; X64-SLOW-NEXT:    leal (%rdi,%rsi), %eax
534; X64-SLOW-NEXT:    retq
535  %tmp = tail call i32 @llvm.fshl.i32(i32 %x, i32 %y, i32 7)
536  ret i32 %tmp
537}
538
539define i64 @const_shift_i64(i64 %x, i64 %y) nounwind {
540; X86-FAST-LABEL: const_shift_i64:
541; X86-FAST:       # %bb.0:
542; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %eax
543; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %ecx
544; X86-FAST-NEXT:    movl {{[0-9]+}}(%esp), %edx
545; X86-FAST-NEXT:    shrdl $25, %ecx, %eax
546; X86-FAST-NEXT:    shldl $7, %ecx, %edx
547; X86-FAST-NEXT:    retl
548;
549; X86-SLOW-LABEL: const_shift_i64:
550; X86-SLOW:       # %bb.0:
551; X86-SLOW-NEXT:    pushl %esi
552; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %ecx
553; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %edx
554; X86-SLOW-NEXT:    movl {{[0-9]+}}(%esp), %esi
555; X86-SLOW-NEXT:    shrl $25, %esi
556; X86-SLOW-NEXT:    movl %ecx, %eax
557; X86-SLOW-NEXT:    shll $7, %eax
558; X86-SLOW-NEXT:    orl %esi, %eax
559; X86-SLOW-NEXT:    shrl $25, %ecx
560; X86-SLOW-NEXT:    shll $7, %edx
561; X86-SLOW-NEXT:    orl %ecx, %edx
562; X86-SLOW-NEXT:    popl %esi
563; X86-SLOW-NEXT:    retl
564;
565; X64-FAST-LABEL: const_shift_i64:
566; X64-FAST:       # %bb.0:
567; X64-FAST-NEXT:    movq %rdi, %rax
568; X64-FAST-NEXT:    shldq $7, %rsi, %rax
569; X64-FAST-NEXT:    retq
570;
571; X64-SLOW-LABEL: const_shift_i64:
572; X64-SLOW:       # %bb.0:
573; X64-SLOW-NEXT:    shrq $57, %rsi
574; X64-SLOW-NEXT:    shlq $7, %rdi
575; X64-SLOW-NEXT:    leaq (%rdi,%rsi), %rax
576; X64-SLOW-NEXT:    retq
577  %tmp = tail call i64 @llvm.fshl.i64(i64 %x, i64 %y, i64 7)
578  ret i64 %tmp
579}
580
581;
582; Combine Consecutive Loads
583;
584
585define i8 @combine_fshl_load_i8(ptr %p) nounwind {
586; X86-LABEL: combine_fshl_load_i8:
587; X86:       # %bb.0:
588; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
589; X86-NEXT:    movzbl 1(%eax), %eax
590; X86-NEXT:    retl
591;
592; X64-LABEL: combine_fshl_load_i8:
593; X64:       # %bb.0:
594; X64-NEXT:    movzbl 1(%rdi), %eax
595; X64-NEXT:    retq
596  %p1 = getelementptr i8, ptr %p, i32 1
597  %ld0 = load i8, ptr%p
598  %ld1 = load i8, ptr%p1
599  %res = call i8 @llvm.fshl.i8(i8 %ld1, i8 %ld0, i8 8)
600  ret i8 %res
601}
602
603define i16 @combine_fshl_load_i16(ptr %p) nounwind {
604; X86-LABEL: combine_fshl_load_i16:
605; X86:       # %bb.0:
606; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
607; X86-NEXT:    movzwl 1(%eax), %eax
608; X86-NEXT:    retl
609;
610; X64-LABEL: combine_fshl_load_i16:
611; X64:       # %bb.0:
612; X64-NEXT:    movzwl 1(%rdi), %eax
613; X64-NEXT:    retq
614  %p1 = getelementptr i16, ptr %p, i32 1
615  %ld0 = load i16, ptr%p
616  %ld1 = load i16, ptr%p1
617  %res = call i16 @llvm.fshl.i16(i16 %ld1, i16 %ld0, i16 8)
618  ret i16 %res
619}
620
621define i32 @combine_fshl_load_i32(ptr %p) nounwind {
622; X86-LABEL: combine_fshl_load_i32:
623; X86:       # %bb.0:
624; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
625; X86-NEXT:    movl 11(%eax), %eax
626; X86-NEXT:    retl
627;
628; X64-LABEL: combine_fshl_load_i32:
629; X64:       # %bb.0:
630; X64-NEXT:    movl 11(%rdi), %eax
631; X64-NEXT:    retq
632  %p0 = getelementptr i32, ptr %p, i32 2
633  %p1 = getelementptr i32, ptr %p, i32 3
634  %ld0 = load i32, ptr%p0
635  %ld1 = load i32, ptr%p1
636  %res = call i32 @llvm.fshl.i32(i32 %ld1, i32 %ld0, i32 8)
637  ret i32 %res
638}
639
640define i64 @combine_fshl_load_i64(ptr %p) nounwind {
641; X86-LABEL: combine_fshl_load_i64:
642; X86:       # %bb.0:
643; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
644; X86-NEXT:    movl 13(%ecx), %eax
645; X86-NEXT:    movl 17(%ecx), %edx
646; X86-NEXT:    retl
647;
648; X64-LABEL: combine_fshl_load_i64:
649; X64:       # %bb.0:
650; X64-NEXT:    movq 13(%rdi), %rax
651; X64-NEXT:    retq
652  %p0 = getelementptr i64, ptr %p, i64 1
653  %p1 = getelementptr i64, ptr %p, i64 2
654  %ld0 = load i64, ptr%p0
655  %ld1 = load i64, ptr%p1
656  %res = call i64 @llvm.fshl.i64(i64 %ld1, i64 %ld0, i64 24)
657  ret i64 %res
658}
659
660!llvm.module.flags = !{!0}
661!0 = !{i32 1, !"ProfileSummary", !1}
662!1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
663!2 = !{!"ProfileFormat", !"InstrProf"}
664!3 = !{!"TotalCount", i64 10000}
665!4 = !{!"MaxCount", i64 10}
666!5 = !{!"MaxInternalCount", i64 1}
667!6 = !{!"MaxFunctionCount", i64 1000}
668!7 = !{!"NumCounts", i64 3}
669!8 = !{!"NumFunctions", i64 3}
670!9 = !{!"DetailedSummary", !10}
671!10 = !{!11, !12, !13}
672!11 = !{i32 10000, i64 100, i32 1}
673!12 = !{i32 999000, i64 100, i32 1}
674!13 = !{i32 999999, i64 1, i32 2}
675!14 = !{!"function_entry_count", i64 0}
676