xref: /llvm-project/llvm/test/Transforms/InstCombine/shl-demand.ll (revision 38fffa630ee80163dc65e759392ad29798905679)
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2; RUN: opt -passes=instcombine -S < %s | FileCheck %s
3
4
5; If we only want bits that already match the signbit then we don't need to shift.
6; https://alive2.llvm.org/ce/z/WJBPVt
7define i32 @src_srem_shl_demand_max_signbit(i32 %a0) {
8; CHECK-LABEL: @src_srem_shl_demand_max_signbit(
9; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 2
10; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SREM]], -2147483648
11; CHECK-NEXT:    ret i32 [[MASK]]
12;
13  %srem = srem i32 %a0, 2           ; srem  = SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSD
14  %shl = shl i32 %srem, 30          ; shl   = SD000000000000000000000000000000
15  %mask = and i32 %shl, -2147483648 ; mask  = 10000000000000000000000000000000
16  ret i32 %mask
17}
18
19define i32 @src_srem_shl_demand_min_signbit(i32 %a0) {
20; CHECK-LABEL: @src_srem_shl_demand_min_signbit(
21; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 1073741823
22; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SREM]], -2147483648
23; CHECK-NEXT:    ret i32 [[MASK]]
24;
25  %srem = srem i32 %a0, 1073741823  ; srem  = SSDDDDDDDDDDDDDDDDDDDDDDDDDDDDDD
26  %shl = shl i32 %srem, 1           ; shl   = SDDDDDDDDDDDDDDDDDDDDDDDDDDDDDD0
27  %mask = and i32 %shl, -2147483648 ; mask  = 10000000000000000000000000000000
28  ret i32 %mask
29}
30
31define i32 @src_srem_shl_demand_max_mask(i32 %a0) {
32; CHECK-LABEL: @src_srem_shl_demand_max_mask(
33; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 2
34; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SREM]], -4
35; CHECK-NEXT:    ret i32 [[MASK]]
36;
37  %srem = srem i32 %a0, 2           ; srem = SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSSD
38  %shl = shl i32 %srem, 1           ; shl  = SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSD0
39  %mask = and i32 %shl, -4          ; mask = 11111111111111111111111111111100
40  ret i32 %mask
41}
42
43; Negative test - mask demands non-signbit from shift source
44define i32 @src_srem_shl_demand_max_signbit_mask_hit_first_demand(i32 %a0) {
45; CHECK-LABEL: @src_srem_shl_demand_max_signbit_mask_hit_first_demand(
46; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 4
47; CHECK-NEXT:    [[SHL:%.*]] = shl nsw i32 [[SREM]], 29
48; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SHL]], -1073741824
49; CHECK-NEXT:    ret i32 [[MASK]]
50;
51  %srem = srem i32 %a0, 4           ; srem  = SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSDD
52  %shl = shl i32 %srem, 29          ; shl   = SDD00000000000000000000000000000
53  %mask = and i32 %shl, -1073741824 ; mask  = 11000000000000000000000000000000
54  ret i32 %mask
55}
56
57define i32 @src_srem_shl_demand_min_signbit_mask_hit_last_demand(i32 %a0) {
58; CHECK-LABEL: @src_srem_shl_demand_min_signbit_mask_hit_last_demand(
59; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 536870912
60; CHECK-NEXT:    [[SHL:%.*]] = shl nsw i32 [[SREM]], 1
61; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SHL]], -1073741822
62; CHECK-NEXT:    ret i32 [[MASK]]
63;
64  %srem = srem i32 %a0, 536870912   ; srem  = SSSDDDDDDDDDDDDDDDDDDDDDDDDDDDDD
65  %shl = shl i32 %srem, 1           ; shl   = SSDDDDDDDDDDDDDDDDDDDDDDDDDDDDD0
66  %mask = and i32 %shl, -1073741822 ; mask  = 11000000000000000000000000000010
67  ret i32 %mask
68}
69
70define i32 @src_srem_shl_demand_eliminate_signbit(i32 %a0) {
71; CHECK-LABEL: @src_srem_shl_demand_eliminate_signbit(
72; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 1073741824
73; CHECK-NEXT:    [[SHL:%.*]] = shl nsw i32 [[SREM]], 1
74; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SHL]], 2
75; CHECK-NEXT:    ret i32 [[MASK]]
76;
77  %srem = srem i32 %a0, 1073741824  ; srem  = SSDDDDDDDDDDDDDDDDDDDDDDDDDDDDDD
78  %shl = shl i32 %srem, 1           ; shl   = DDDDDDDDDDDDDDDDDDDDDDDDDDDDDDD0
79  %mask = and i32 %shl, 2           ; mask  = 00000000000000000000000000000010
80  ret i32 %mask
81}
82
83define i32 @src_srem_shl_demand_max_mask_hit_demand(i32 %a0) {
84; CHECK-LABEL: @src_srem_shl_demand_max_mask_hit_demand(
85; CHECK-NEXT:    [[SREM:%.*]] = srem i32 [[A0:%.*]], 4
86; CHECK-NEXT:    [[SHL:%.*]] = shl nsw i32 [[SREM]], 1
87; CHECK-NEXT:    [[MASK:%.*]] = and i32 [[SHL]], -4
88; CHECK-NEXT:    ret i32 [[MASK]]
89;
90  %srem = srem i32 %a0, 4           ; srem = SSSSSSSSSSSSSSSSSSSSSSSSSSSSSSDD
91  %shl= shl i32 %srem, 1            ; shl  = SSSSSSSSSSSSSSSSSSSSSSSSSSSSSDD0
92  %mask = and i32 %shl, -4          ; mask = 11111111111111111111111111111100
93  ret i32 %mask
94}
95
96define <2 x i32> @src_srem_shl_mask_vector(<2 x i32> %a0) {
97; CHECK-LABEL: @src_srem_shl_mask_vector(
98; CHECK-NEXT:    [[SREM:%.*]] = srem <2 x i32> [[A0:%.*]], splat (i32 4)
99; CHECK-NEXT:    [[SHL:%.*]] = shl nsw <2 x i32> [[SREM]], splat (i32 29)
100; CHECK-NEXT:    [[MASK:%.*]] = and <2 x i32> [[SHL]], splat (i32 -1073741824)
101; CHECK-NEXT:    ret <2 x i32> [[MASK]]
102;
103  %srem = srem <2 x i32> %a0, <i32 4, i32 4>
104  %shl = shl <2 x i32> %srem, <i32 29, i32 29>
105  %mask = and <2 x i32> %shl, <i32 -1073741824, i32 -1073741824>
106  ret <2 x i32> %mask
107}
108
109define <2 x i32> @src_srem_shl_mask_vector_nonconstant(<2 x i32> %a0, <2 x i32> %a1) {
110; CHECK-LABEL: @src_srem_shl_mask_vector_nonconstant(
111; CHECK-NEXT:    [[SREM:%.*]] = srem <2 x i32> [[A0:%.*]], splat (i32 4)
112; CHECK-NEXT:    [[SHL:%.*]] = shl <2 x i32> [[SREM]], [[A1:%.*]]
113; CHECK-NEXT:    [[MASK:%.*]] = and <2 x i32> [[SHL]], splat (i32 -1073741824)
114; CHECK-NEXT:    ret <2 x i32> [[MASK]]
115;
116  %srem = srem <2 x i32> %a0, <i32 4, i32 4>
117  %shl = shl <2 x i32> %srem, %a1
118  %mask = and <2 x i32> %shl, <i32 -1073741824, i32 -1073741824>
119  ret <2 x i32> %mask
120}
121
122define i16 @sext_shl_trunc_same_size(i16 %x, i32 %y) {
123; CHECK-LABEL: @sext_shl_trunc_same_size(
124; CHECK-NEXT:    [[CONV1:%.*]] = zext i16 [[X:%.*]] to i32
125; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[CONV1]], [[Y:%.*]]
126; CHECK-NEXT:    [[T:%.*]] = trunc i32 [[SHL]] to i16
127; CHECK-NEXT:    ret i16 [[T]]
128;
129  %conv = sext i16 %x to i32
130  %shl = shl i32 %conv, %y
131  %t = trunc i32 %shl to i16
132  ret i16 %t
133}
134
135define i5 @sext_shl_trunc_smaller(i16 %x, i32 %y) {
136; CHECK-LABEL: @sext_shl_trunc_smaller(
137; CHECK-NEXT:    [[CONV1:%.*]] = zext i16 [[X:%.*]] to i32
138; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[CONV1]], [[Y:%.*]]
139; CHECK-NEXT:    [[T:%.*]] = trunc i32 [[SHL]] to i5
140; CHECK-NEXT:    ret i5 [[T]]
141;
142  %conv = sext i16 %x to i32
143  %shl = shl i32 %conv, %y
144  %t = trunc i32 %shl to i5
145  ret i5 %t
146}
147
148; negative test - demanding 1 high-bit too many to change the extend
149
150define i17 @sext_shl_trunc_larger(i16 %x, i32 %y) {
151; CHECK-LABEL: @sext_shl_trunc_larger(
152; CHECK-NEXT:    [[CONV:%.*]] = sext i16 [[X:%.*]] to i32
153; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[CONV]], [[Y:%.*]]
154; CHECK-NEXT:    [[T:%.*]] = trunc i32 [[SHL]] to i17
155; CHECK-NEXT:    ret i17 [[T]]
156;
157  %conv = sext i16 %x to i32
158  %shl = shl i32 %conv, %y
159  %t = trunc i32 %shl to i17
160  ret i17 %t
161}
162
163define i32 @sext_shl_mask(i16 %x, i32 %y) {
164; CHECK-LABEL: @sext_shl_mask(
165; CHECK-NEXT:    [[CONV1:%.*]] = zext i16 [[X:%.*]] to i32
166; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[CONV1]], [[Y:%.*]]
167; CHECK-NEXT:    [[T:%.*]] = and i32 [[SHL]], 65535
168; CHECK-NEXT:    ret i32 [[T]]
169;
170  %conv = sext i16 %x to i32
171  %shl = shl i32 %conv, %y
172  %t = and i32 %shl, 65535
173  ret i32 %t
174}
175
176; negative test - demanding a bit that could change with sext
177
178define i32 @sext_shl_mask_higher(i16 %x, i32 %y) {
179; CHECK-LABEL: @sext_shl_mask_higher(
180; CHECK-NEXT:    [[CONV:%.*]] = sext i16 [[X:%.*]] to i32
181; CHECK-NEXT:    [[SHL:%.*]] = shl i32 [[CONV]], [[Y:%.*]]
182; CHECK-NEXT:    [[T:%.*]] = and i32 [[SHL]], 65536
183; CHECK-NEXT:    ret i32 [[T]]
184;
185  %conv = sext i16 %x to i32
186  %shl = shl i32 %conv, %y
187  %t = and i32 %shl, 65536
188  ret i32 %t
189}
190
191; May need some, but not all of the bits set by the 'or'.
192
193define i32 @set_shl_mask(i32 %x, i32 %y) {
194; CHECK-LABEL: @set_shl_mask(
195; CHECK-NEXT:    [[Z:%.*]] = or i32 [[X:%.*]], 65537
196; CHECK-NEXT:    [[S:%.*]] = shl i32 [[Z]], [[Y:%.*]]
197; CHECK-NEXT:    [[R:%.*]] = and i32 [[S]], 65536
198; CHECK-NEXT:    ret i32 [[R]]
199;
200  %z = or i32 %x, 196609
201  %s = shl i32 %z, %y
202  %r = and i32 %s, 65536
203  ret i32 %r
204}
205
206; PR50341
207
208define i8 @must_drop_poison(i32 %x, i32 %y)  {
209; CHECK-LABEL: @must_drop_poison(
210; CHECK-NEXT:    [[S:%.*]] = shl i32 [[X:%.*]], [[Y:%.*]]
211; CHECK-NEXT:    [[T:%.*]] = trunc i32 [[S]] to i8
212; CHECK-NEXT:    ret i8 [[T]]
213;
214  %a = and i32 %x, 255
215  %s = shl nuw nsw i32 %a, %y
216  %t = trunc i32 %s to i8
217  ret i8 %t
218}
219
220; This would infinite loop with D110170 / bb9333c3504a
221
222define i32 @f_t15_t01_t09(i40 %t2) {
223; CHECK-LABEL: @f_t15_t01_t09(
224; CHECK-NEXT:    [[SH_DIFF:%.*]] = ashr i40 [[T2:%.*]], 15
225; CHECK-NEXT:    [[TR_SH_DIFF:%.*]] = trunc nsw i40 [[SH_DIFF]] to i32
226; CHECK-NEXT:    [[SHL1:%.*]] = and i32 [[TR_SH_DIFF]], -65536
227; CHECK-NEXT:    ret i32 [[SHL1]]
228;
229  %downscale = ashr i40 %t2, 31
230  %resize = trunc i40 %downscale to i32
231  %shl1 = shl i32 %resize, 16
232  %resize1 = ashr i32 %shl1, 16
233  %r = shl i32 %resize1, 31
234  ret i32 %shl1
235}
236