• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt < %s -instcombine -S | FileCheck %s
3
4; If we have some pattern that leaves only some low bits set, and then performs
5; left-shift of those bits, if none of the bits that are left after the final
6; shift are modified by the mask, we can omit the mask.
7
8; There are many variants to this pattern:
9;   f)  ((x << maskNbits) a>> maskNbits) << shiftNbits
10; simplify to:
11;   x << shiftNbits
12; iff (shiftNbits-maskNbits) s>= 0 (i.e. shiftNbits u>= maskNbits)
13
14; Simple tests. We don't care about extra uses.
15
16declare void @use32(i32)
17
18define i32 @t0_basic(i32 %x, i32 %nbits) {
19; CHECK-LABEL: @t0_basic(
20; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS:%.*]]
21; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS]]
22; CHECK-NEXT:    call void @use32(i32 [[T0]])
23; CHECK-NEXT:    call void @use32(i32 [[T1]])
24; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[X]], [[NBITS]]
25; CHECK-NEXT:    ret i32 [[T2]]
26;
27  %t0 = shl i32 %x, %nbits
28  %t1 = ashr i32 %t0, %nbits
29  call void @use32(i32 %t0)
30  call void @use32(i32 %t1)
31  %t2 = shl i32 %t1, %nbits
32  ret i32 %t2
33}
34
35define i32 @t1_bigger_shift(i32 %x, i32 %nbits) {
36; CHECK-LABEL: @t1_bigger_shift(
37; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS:%.*]]
38; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS]]
39; CHECK-NEXT:    [[T2:%.*]] = add i32 [[NBITS]], 1
40; CHECK-NEXT:    call void @use32(i32 [[T0]])
41; CHECK-NEXT:    call void @use32(i32 [[T1]])
42; CHECK-NEXT:    call void @use32(i32 [[T2]])
43; CHECK-NEXT:    [[T3:%.*]] = shl i32 [[X]], [[T2]]
44; CHECK-NEXT:    ret i32 [[T3]]
45;
46  %t0 = shl i32 %x, %nbits
47  %t1 = ashr i32 %t0, %nbits
48  %t2 = add i32 %nbits, 1
49  call void @use32(i32 %t0)
50  call void @use32(i32 %t1)
51  call void @use32(i32 %t2)
52  %t3 = shl i32 %t1, %t2
53  ret i32 %t3
54}
55
56; Vectors
57
58declare void @use3xi32(<3 x i32>)
59
60define <3 x i32> @t2_vec_splat(<3 x i32> %x, <3 x i32> %nbits) {
61; CHECK-LABEL: @t2_vec_splat(
62; CHECK-NEXT:    [[T0:%.*]] = shl <3 x i32> [[X:%.*]], [[NBITS:%.*]]
63; CHECK-NEXT:    [[T1:%.*]] = ashr <3 x i32> [[T0]], [[NBITS]]
64; CHECK-NEXT:    [[T2:%.*]] = add <3 x i32> [[NBITS]], <i32 1, i32 1, i32 1>
65; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T0]])
66; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T1]])
67; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T2]])
68; CHECK-NEXT:    [[T3:%.*]] = shl <3 x i32> [[X]], [[T2]]
69; CHECK-NEXT:    ret <3 x i32> [[T3]]
70;
71  %t0 = shl <3 x i32> %x, %nbits
72  %t1 = ashr <3 x i32> %t0, %nbits
73  %t2 = add <3 x i32> %nbits, <i32 1, i32 1, i32 1>
74  call void @use3xi32(<3 x i32> %t0)
75  call void @use3xi32(<3 x i32> %t1)
76  call void @use3xi32(<3 x i32> %t2)
77  %t3 = shl <3 x i32> %t1, %t2
78  ret <3 x i32> %t3
79}
80
81define <3 x i32> @t3_vec_nonsplat(<3 x i32> %x, <3 x i32> %nbits) {
82; CHECK-LABEL: @t3_vec_nonsplat(
83; CHECK-NEXT:    [[T0:%.*]] = shl <3 x i32> [[X:%.*]], [[NBITS:%.*]]
84; CHECK-NEXT:    [[T1:%.*]] = ashr <3 x i32> [[T0]], [[NBITS]]
85; CHECK-NEXT:    [[T2:%.*]] = add <3 x i32> [[NBITS]], <i32 1, i32 0, i32 2>
86; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T0]])
87; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T1]])
88; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T2]])
89; CHECK-NEXT:    [[T3:%.*]] = shl <3 x i32> [[X]], [[T2]]
90; CHECK-NEXT:    ret <3 x i32> [[T3]]
91;
92  %t0 = shl <3 x i32> %x, %nbits
93  %t1 = ashr <3 x i32> %t0, %nbits
94  %t2 = add <3 x i32> %nbits, <i32 1, i32 0, i32 2>
95  call void @use3xi32(<3 x i32> %t0)
96  call void @use3xi32(<3 x i32> %t1)
97  call void @use3xi32(<3 x i32> %t2)
98  %t3 = shl <3 x i32> %t1, %t2
99  ret <3 x i32> %t3
100}
101
102define <3 x i32> @t4_vec_undef(<3 x i32> %x, <3 x i32> %nbits) {
103; CHECK-LABEL: @t4_vec_undef(
104; CHECK-NEXT:    [[T0:%.*]] = shl <3 x i32> [[X:%.*]], [[NBITS:%.*]]
105; CHECK-NEXT:    [[T1:%.*]] = ashr <3 x i32> [[T0]], [[NBITS]]
106; CHECK-NEXT:    [[T2:%.*]] = add <3 x i32> [[NBITS]], <i32 1, i32 undef, i32 1>
107; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T0]])
108; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T1]])
109; CHECK-NEXT:    call void @use3xi32(<3 x i32> [[T2]])
110; CHECK-NEXT:    [[T3:%.*]] = shl <3 x i32> [[X]], [[T2]]
111; CHECK-NEXT:    ret <3 x i32> [[T3]]
112;
113  %t0 = shl <3 x i32> %x, %nbits
114  %t1 = ashr <3 x i32> %t0, %nbits
115  %t2 = add <3 x i32> %nbits, <i32 1, i32 undef, i32 1>
116  call void @use3xi32(<3 x i32> %t0)
117  call void @use3xi32(<3 x i32> %t1)
118  call void @use3xi32(<3 x i32> %t2)
119  %t3 = shl <3 x i32> %t1, %t2
120  ret <3 x i32> %t3
121}
122
123; Fast-math flags. We must not preserve them!
124
125define i32 @t5_nuw(i32 %x, i32 %nbits) {
126; CHECK-LABEL: @t5_nuw(
127; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS:%.*]]
128; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS]]
129; CHECK-NEXT:    call void @use32(i32 [[T0]])
130; CHECK-NEXT:    call void @use32(i32 [[T1]])
131; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[X]], [[NBITS]]
132; CHECK-NEXT:    ret i32 [[T2]]
133;
134  %t0 = shl i32 %x, %nbits
135  %t1 = ashr i32 %t0, %nbits
136  call void @use32(i32 %t0)
137  call void @use32(i32 %t1)
138  %t2 = shl nuw i32 %t1, %nbits
139  ret i32 %t2
140}
141
142define i32 @t6_nsw(i32 %x, i32 %nbits) {
143; CHECK-LABEL: @t6_nsw(
144; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS:%.*]]
145; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS]]
146; CHECK-NEXT:    call void @use32(i32 [[T0]])
147; CHECK-NEXT:    call void @use32(i32 [[T1]])
148; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[X]], [[NBITS]]
149; CHECK-NEXT:    ret i32 [[T2]]
150;
151  %t0 = shl i32 %x, %nbits
152  %t1 = ashr i32 %t0, %nbits
153  call void @use32(i32 %t0)
154  call void @use32(i32 %t1)
155  %t2 = shl nsw i32 %t1, %nbits
156  ret i32 %t2
157}
158
159define i32 @t7_nuw_nsw(i32 %x, i32 %nbits) {
160; CHECK-LABEL: @t7_nuw_nsw(
161; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS:%.*]]
162; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS]]
163; CHECK-NEXT:    call void @use32(i32 [[T0]])
164; CHECK-NEXT:    call void @use32(i32 [[T1]])
165; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[X]], [[NBITS]]
166; CHECK-NEXT:    ret i32 [[T2]]
167;
168  %t0 = shl i32 %x, %nbits
169  %t1 = ashr i32 %t0, %nbits
170  call void @use32(i32 %t0)
171  call void @use32(i32 %t1)
172  %t2 = shl nuw nsw i32 %t1, %nbits
173  ret i32 %t2
174}
175
176; Special test
177
178declare void @llvm.assume(i1 %cond)
179
180; We can't simplify (%shiftnbits-%masknbits) but we have an assumption.
181define i32 @t8_assume_uge(i32 %x, i32 %masknbits, i32 %shiftnbits) {
182; CHECK-LABEL: @t8_assume_uge(
183; CHECK-NEXT:    [[CMP:%.*]] = icmp uge i32 [[SHIFTNBITS:%.*]], [[MASKNBITS:%.*]]
184; CHECK-NEXT:    call void @llvm.assume(i1 [[CMP]])
185; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[MASKNBITS]]
186; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[MASKNBITS]]
187; CHECK-NEXT:    call void @use32(i32 [[T0]])
188; CHECK-NEXT:    call void @use32(i32 [[T1]])
189; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[T1]], [[SHIFTNBITS]]
190; CHECK-NEXT:    ret i32 [[T2]]
191;
192  %cmp = icmp uge i32 %shiftnbits, %masknbits
193  call void @llvm.assume(i1 %cmp)
194  %t0 = shl i32 %x, %masknbits
195  %t1 = ashr i32 %t0, %masknbits
196  call void @use32(i32 %t0)
197  call void @use32(i32 %t1)
198  %t2 = shl i32 %t1, %shiftnbits
199  ret i32 %t2
200}
201
202; Negative tests
203
204define i32 @n9_different_shamts0(i32 %x, i32 %nbits0, i32 %nbits1) {
205; CHECK-LABEL: @n9_different_shamts0(
206; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS0:%.*]]
207; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS1:%.*]]
208; CHECK-NEXT:    call void @use32(i32 [[T0]])
209; CHECK-NEXT:    call void @use32(i32 [[T1]])
210; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[T1]], [[NBITS0]]
211; CHECK-NEXT:    ret i32 [[T2]]
212;
213  %t0 = shl i32 %x, %nbits0 ; different shift amts
214  %t1 = ashr i32 %t0, %nbits1 ; different shift amts
215  call void @use32(i32 %t0)
216  call void @use32(i32 %t1)
217  %t2 = shl i32 %t1, %nbits0
218  ret i32 %t2
219}
220
221define i32 @n10_different_shamts1(i32 %x, i32 %nbits0, i32 %nbits1) {
222; CHECK-LABEL: @n10_different_shamts1(
223; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS0:%.*]]
224; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS1:%.*]]
225; CHECK-NEXT:    call void @use32(i32 [[T0]])
226; CHECK-NEXT:    call void @use32(i32 [[T1]])
227; CHECK-NEXT:    [[T2:%.*]] = shl i32 [[T1]], [[NBITS1]]
228; CHECK-NEXT:    ret i32 [[T2]]
229;
230  %t0 = shl i32 %x, %nbits0 ; different shift amts
231  %t1 = ashr i32 %t0, %nbits1 ; different shift amts
232  call void @use32(i32 %t0)
233  call void @use32(i32 %t1)
234  %t2 = shl i32 %t1, %nbits1
235  ret i32 %t2
236}
237
238define i32 @n11_shamt_is_smaller(i32 %x, i32 %nbits) {
239; CHECK-LABEL: @n11_shamt_is_smaller(
240; CHECK-NEXT:    [[T0:%.*]] = shl i32 [[X:%.*]], [[NBITS:%.*]]
241; CHECK-NEXT:    [[T1:%.*]] = ashr i32 [[T0]], [[NBITS]]
242; CHECK-NEXT:    [[T2:%.*]] = add i32 [[NBITS]], -1
243; CHECK-NEXT:    call void @use32(i32 [[T0]])
244; CHECK-NEXT:    call void @use32(i32 [[T2]])
245; CHECK-NEXT:    [[T3:%.*]] = shl i32 [[T1]], [[T2]]
246; CHECK-NEXT:    ret i32 [[T3]]
247;
248  %t0 = shl i32 %x, %nbits
249  %t1 = ashr i32 %t0, %nbits
250  %t2 = add i32 %nbits, -1
251  call void @use32(i32 %t0)
252  call void @use32(i32 %t2)
253  %t3 = shl i32 %t1, %t2 ; shift is smaller than mask
254  ret i32 %t3
255}
256