Skip to content

Commit 90a256e

Browse files
yashnatoryashnator
committed
[Instcombine][Tests] Add test for folding select
to cmp with constants in non-canonical inequalities
1 parent 339797d commit 90a256e

File tree

1 file changed

+326
-0
lines changed

1 file changed

+326
-0
lines changed

‎llvm/test/Transforms/InstCombine/select-to-cmp.ll

Lines changed: 326 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,326 @@
1+
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
2+
; RUN: opt -passes=instcombine -S < %s | FileCheck %s
3+
4+
; Tests for select to scmp
5+
6+
define i32 @scmp_x_0_inverted(i32 %x) {
7+
; CHECK-LABEL: define i32 @scmp_x_0_inverted(
8+
; CHECK-SAME: i32 [[X:%.*]]) {
9+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ne i32 [[X]], 0
10+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i32
11+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt i32 [[X]], -1
12+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
13+
; CHECK-NEXT: ret i32 [[TMP1]]
14+
;
15+
%2 = icmp ne i32 %x, 0
16+
%3 = zext i1 %2 to i32
17+
%4 = icmp sgt i32 %x, -1
18+
%5 = select i1 %4, i32 %3, i32 -1
19+
ret i32 %5
20+
}
21+
22+
; y = -10
23+
define i32 @scmp_x_0_inverted_const_neg10(i32 %x) {
24+
; CHECK-LABEL: define i32 @scmp_x_0_inverted_const_neg10(
25+
; CHECK-SAME: i32 [[X:%.*]]) {
26+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ne i32 [[X]], -10
27+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i32
28+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt i32 [[X]], -11
29+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
30+
; CHECK-NEXT: ret i32 [[TMP1]]
31+
;
32+
%1 = icmp ne i32 %x, -10
33+
%2 = zext i1 %1 to i32
34+
%3 = icmp sgt i32 %x, -11
35+
%4 = select i1 %3, i32 %2, i32 -1
36+
ret i32 %4
37+
}
38+
39+
; y = 7 (i8)
40+
define i8 @scmp_x_0_inverted_i8(i8 %x) {
41+
; CHECK-LABEL: define i8 @scmp_x_0_inverted_i8(
42+
; CHECK-SAME: i8 [[X:%.*]]) {
43+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ne i8 [[X]], 7
44+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i8
45+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt i8 [[X]], 6
46+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i8 [[TMP2]], i8 -1
47+
; CHECK-NEXT: ret i8 [[TMP1]]
48+
;
49+
%1 = icmp ne i8 %x, 7
50+
%2 = zext i1 %1 to i8
51+
%3 = icmp sgt i8 %x, 6
52+
%4 = select i1 %3, i8 %2, i8 -1
53+
ret i8 %4
54+
}
55+
56+
; scmp using ints of two kinds- i32 and i64
57+
define i32 @scmp_x_0_inverted_i64_neq(i32 %x) {
58+
; CHECK-LABEL: define i32 @scmp_x_0_inverted_i64_neq(
59+
; CHECK-SAME: i32 [[X:%.*]]) {
60+
; CHECK-NEXT: [[CMP1:%.*]] = icmp ne i32 [[X]], 0
61+
; CHECK-NEXT: [[CMP2:%.*]] = icmp sgt i32 [[X]], -1
62+
; CHECK-NEXT: [[TMP1:%.*]] = zext i1 [[CMP1]] to i32
63+
; CHECK-NEXT: [[RET:%.*]] = select i1 [[CMP2]], i32 [[TMP1]], i32 -1
64+
; CHECK-NEXT: ret i32 [[RET]]
65+
;
66+
%x64 = sext i32 %x to i64
67+
%cmp1 = icmp ne i64 %x64, 0
68+
%zext = zext i1 %cmp1 to i64
69+
%cmp2 = icmp sgt i64 %x64, -1
70+
%sel = select i1 %cmp2, i64 %zext, i64 -1
71+
%ret = trunc i64 %sel to i32
72+
ret i32 %ret
73+
}
74+
75+
; Same example as previous but with inequality
76+
define i32 @scmp_x_0_inverted_i64_sgt(i32 %x) {
77+
; CHECK-LABEL: define i32 @scmp_x_0_inverted_i64_sgt(
78+
; CHECK-SAME: i32 [[X:%.*]]) {
79+
; CHECK-NEXT: [[CMP1:%.*]] = icmp sgt i32 [[X]], 0
80+
; CHECK-NEXT: [[CMP2:%.*]] = icmp sgt i32 [[X]], -1
81+
; CHECK-NEXT: [[TMP1:%.*]] = zext i1 [[CMP1]] to i32
82+
; CHECK-NEXT: [[RET:%.*]] = select i1 [[CMP2]], i32 [[TMP1]], i32 -1
83+
; CHECK-NEXT: ret i32 [[RET]]
84+
;
85+
%x64 = sext i32 %x to i64
86+
%cmp1 = icmp sgt i64 %x64, 0
87+
%zext = zext i1 %cmp1 to i64
88+
%cmp2 = icmp sgt i64 %x64, -1
89+
%sel = select i1 %cmp2, i64 %zext, i64 -1
90+
%ret = trunc i64 %sel to i32
91+
ret i32 %ret
92+
}
93+
94+
; y = -1000
95+
define i32 @scmp_x_0_inverted_const_neg1000(i32 %x) {
96+
; CHECK-LABEL: define i32 @scmp_x_0_inverted_const_neg1000(
97+
; CHECK-SAME: i32 [[X:%.*]]) {
98+
; CHECK-NEXT: [[TMP4:%.*]] = icmp sgt i32 [[X]], -1000
99+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i32
100+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt i32 [[X]], -1001
101+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
102+
; CHECK-NEXT: ret i32 [[TMP1]]
103+
;
104+
%1 = icmp sgt i32 %x, -1000
105+
%2 = zext i1 %1 to i32
106+
%3 = icmp sgt i32 %x, -1001
107+
%4 = select i1 %3, i32 %2, i32 -1
108+
ret i32 %4
109+
}
110+
111+
; y = 1729
112+
define i32 @scmp_x_0_inverted_const_1729_sgt(i32 %x) {
113+
; CHECK-LABEL: define i32 @scmp_x_0_inverted_const_1729_sgt(
114+
; CHECK-SAME: i32 [[X:%.*]]) {
115+
; CHECK-NEXT: [[TMP4:%.*]] = icmp sgt i32 [[X]], 1729
116+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i32
117+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt i32 [[X]], 1728
118+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
119+
; CHECK-NEXT: ret i32 [[TMP1]]
120+
;
121+
%1 = icmp sgt i32 %x, 1729
122+
%2 = zext i1 %1 to i32
123+
%3 = icmp sgt i32 %x, 1728
124+
%4 = select i1 %3, i32 %2, i32 -1
125+
ret i32 %4
126+
}
127+
128+
; ucmp with 10
129+
define i32 @ucmp_x_10_inverted(i32 %x) {
130+
; CHECK-LABEL: define i32 @ucmp_x_10_inverted(
131+
; CHECK-SAME: i32 [[X:%.*]]) {
132+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ne i32 [[X]], 10
133+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i32
134+
; CHECK-NEXT: [[TMP3:%.*]] = icmp ugt i32 [[X]], 9
135+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
136+
; CHECK-NEXT: ret i32 [[TMP1]]
137+
;
138+
%1 = icmp ne i32 %x, 10
139+
%2 = zext i1 %1 to i32
140+
%3 = icmp ugt i32 %x, 9
141+
%4 = select i1 %3, i32 %2, i32 -1
142+
ret i32 %4
143+
}
144+
145+
; ucmp with -3, wraps around
146+
define i32 @ucmp_x_neg1_inverted(i32 %x) {
147+
; CHECK-LABEL: define i32 @ucmp_x_neg1_inverted(
148+
; CHECK-SAME: i32 [[X:%.*]]) {
149+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ne i32 [[X]], -3
150+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i32
151+
; CHECK-NEXT: [[TMP3:%.*]] = icmp ugt i32 [[X]], -4
152+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
153+
; CHECK-NEXT: ret i32 [[TMP1]]
154+
;
155+
%1 = icmp ne i32 %x, -3
156+
%2 = zext i1 %1 to i32
157+
%3 = icmp ugt i32 %x, -4
158+
%4 = select i1 %3, i32 %2, i32 -1
159+
ret i32 %4
160+
}
161+
162+
; ucmp with -4, wraps around
163+
define i8 @ucmp_x_neg4_i8_ugt(i8 %x) {
164+
; CHECK-LABEL: define i8 @ucmp_x_neg4_i8_ugt(
165+
; CHECK-SAME: i8 [[X:%.*]]) {
166+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ugt i8 [[X]], -4
167+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP4]] to i8
168+
; CHECK-NEXT: [[TMP3:%.*]] = icmp ugt i8 [[X]], -5
169+
; CHECK-NEXT: [[TMP1:%.*]] = select i1 [[TMP3]], i8 [[TMP2]], i8 -1
170+
; CHECK-NEXT: ret i8 [[TMP1]]
171+
;
172+
%1 = icmp ugt i8 %x, -4
173+
%2 = zext i1 %1 to i8
174+
%3 = icmp ugt i8 %x, -5
175+
%4 = select i1 %3, i8 %2, i8 -1
176+
ret i8 %4
177+
}
178+
179+
; Vector tests
180+
181+
; Test with splat vec
182+
define <4 x i32> @scmp_x_0_inverted_splat_vec(<4 x i32> %x) {
183+
; CHECK-LABEL: define <4 x i32> @scmp_x_0_inverted_splat_vec(
184+
; CHECK-SAME: <4 x i32> [[X:%.*]]) {
185+
; CHECK-NEXT: [[TMP4:%.*]] = icmp ne <4 x i32> [[X]], zeroinitializer
186+
; CHECK-NEXT: [[TMP2:%.*]] = zext <4 x i1> [[TMP4]] to <4 x i32>
187+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt <4 x i32> [[X]], splat (i32 -1)
188+
; CHECK-NEXT: [[TMP1:%.*]] = select <4 x i1> [[TMP3]], <4 x i32> [[TMP2]], <4 x i32> splat (i32 -1)
189+
; CHECK-NEXT: ret <4 x i32> [[TMP1]]
190+
;
191+
%2 = icmp ne <4 x i32> %x, zeroinitializer
192+
%3 = zext <4 x i1> %2 to <4 x i32>
193+
%4 = icmp sgt <4 x i32> %x, <i32 -1, i32 -1, i32 -1, i32 -1>
194+
%5 = select <4 x i1> %4, <4 x i32> %3, <4 x i32> <i32 -1, i32 -1, i32 -1, i32 -1>
195+
ret <4 x i32> %5
196+
}
197+
198+
; Test with non-splat vector and different bitwidth
199+
define <4 x i32> @non_splat_vec_scmp_diff_bitwidth(<4 x i32> %x) {
200+
; CHECK-LABEL: define <4 x i32> @non_splat_vec_scmp_diff_bitwidth(
201+
; CHECK-SAME: <4 x i32> [[X:%.*]]) {
202+
; CHECK-NEXT: [[CMP1:%.*]] = icmp slt <4 x i32> [[X]], <i32 0, i32 1, i32 -1, i32 5>
203+
; CHECK-NEXT: [[CMP2:%.*]] = icmp slt <4 x i32> [[X]], <i32 1, i32 2, i32 0, i32 6>
204+
; CHECK-NEXT: [[TMP1:%.*]] = sext <4 x i1> [[CMP1]] to <4 x i32>
205+
; CHECK-NEXT: [[RET:%.*]] = select <4 x i1> [[CMP2]], <4 x i32> [[TMP1]], <4 x i32> splat (i32 1)
206+
; CHECK-NEXT: ret <4 x i32> [[RET]]
207+
;
208+
%x64 = sext <4 x i32> %x to <4 x i64>
209+
%cmp1 = icmp slt <4 x i64> %x64, <i64 0, i64 1, i64 -1, i64 5>
210+
%sext = sext <4 x i1> %cmp1 to <4 x i64>
211+
%cmp2 = icmp slt <4 x i64> %x64, <i64 1, i64 2, i64 0, i64 6>
212+
%sel = select <4 x i1> %cmp2, <4 x i64> %sext, <4 x i64> <i64 1, i64 1, i64 1, i64 1>
213+
%ret = trunc <4 x i64> %sel to <4 x i32>
214+
ret <4 x i32> %ret
215+
}
216+
217+
; Negative examples
218+
219+
; Not scmp due to wrong RHS of the predicate
220+
define i32 @scmp_ne_0(i32 %0) {
221+
; CHECK-LABEL: define i32 @scmp_ne_0(
222+
; CHECK-SAME: i32 [[TMP0:%.*]]) {
223+
; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i32 [[TMP0]], 0
224+
; CHECK-NEXT: [[TMP3:%.*]] = zext i1 [[TMP2]] to i32
225+
; CHECK-NEXT: [[TMP4:%.*]] = icmp sgt i32 [[TMP0]], 1
226+
; CHECK-NEXT: [[TMP5:%.*]] = select i1 [[TMP4]], i32 [[TMP3]], i32 -1
227+
; CHECK-NEXT: ret i32 [[TMP5]]
228+
;
229+
%2 = icmp ne i32 %0, 0
230+
%3 = zext i1 %2 to i32
231+
%4 = icmp sgt i32 %0, 1
232+
%5 = select i1 %4, i32 %3, i32 -1
233+
ret i32 %5
234+
}
235+
236+
; y = 0 with unsigned compare but RHS wraps
237+
define i32 @ucmp_x_0_inverted(i32 %x) {
238+
; CHECK-LABEL: define i32 @ucmp_x_0_inverted(
239+
; CHECK-SAME: i32 [[X:%.*]]) {
240+
; CHECK-NEXT: ret i32 -1
241+
;
242+
%1 = icmp ne i32 %x, 0
243+
%2 = zext i1 %1 to i32
244+
%3 = icmp ugt i32 %x, -1
245+
%4 = select i1 %3, i32 %2, i32 -1
246+
ret i32 %4
247+
}
248+
249+
; Don't fold with INT32_MIN
250+
define i32 @scmp_x_0_inverted_const_min(i32 %x) {
251+
; CHECK-LABEL: define i32 @scmp_x_0_inverted_const_min(
252+
; CHECK-SAME: i32 [[X:%.*]]) {
253+
; CHECK-NEXT: [[TMP1:%.*]] = icmp ne i32 [[X]], -2147483648
254+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP1]] to i32
255+
; CHECK-NEXT: ret i32 [[TMP2]]
256+
;
257+
%1 = icmp ne i32 %x, -2147483648
258+
%2 = zext i1 %1 to i32
259+
%3 = icmp sge i32 %x, -2147483648
260+
%4 = select i1 %3, i32 %2, i32 -1
261+
ret i32 %4
262+
}
263+
264+
; Unsigned cmp of zext of i32 with i64 -1 should always be -1
265+
define i32 @ucmp_x_0_inverted_i64_ugt(i32 %x) {
266+
; CHECK-LABEL: define i32 @ucmp_x_0_inverted_i64_ugt(
267+
; CHECK-SAME: i32 [[X:%.*]]) {
268+
; CHECK-NEXT: ret i32 -1
269+
;
270+
%x64 = zext i32 %x to i64
271+
%cmp1 = icmp ugt i64 %x64, 0
272+
%zext = zext i1 %cmp1 to i64
273+
%cmp2 = icmp ugt i64 %x64, -1
274+
%sel = select i1 %cmp2, i64 %zext, i64 -1
275+
%ret = trunc i64 %sel to i32
276+
ret i32 %ret
277+
}
278+
279+
; y = 4294967295 (UINT32_MAX), simply sign extend neq
280+
define i32 @ucmp_x_const_u32max(i32 %x) {
281+
; CHECK-LABEL: define i32 @ucmp_x_const_u32max(
282+
; CHECK-SAME: i32 [[X:%.*]]) {
283+
; CHECK-NEXT: [[TMP1:%.*]] = icmp ne i32 [[X]], -1
284+
; CHECK-NEXT: [[TMP2:%.*]] = sext i1 [[TMP1]] to i32
285+
; CHECK-NEXT: ret i32 [[TMP2]]
286+
;
287+
%1 = icmp ugt i32 %x, 4294967295
288+
%2 = zext i1 %1 to i32
289+
%3 = icmp ugt i32 %x, 4294967294
290+
%4 = select i1 %3, i32 %2, i32 -1
291+
ret i32 %4
292+
}
293+
294+
; Don't fold with different signedness
295+
define i32 @different_signedness_neg(i32 %x) {
296+
; CHECK-LABEL: define i32 @different_signedness_neg(
297+
; CHECK-SAME: i32 [[X:%.*]]) {
298+
; CHECK-NEXT: [[TMP1:%.*]] = icmp ugt i32 [[X]], -10
299+
; CHECK-NEXT: [[TMP2:%.*]] = zext i1 [[TMP1]] to i32
300+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt i32 [[X]], -11
301+
; CHECK-NEXT: [[TMP4:%.*]] = select i1 [[TMP3]], i32 [[TMP2]], i32 -1
302+
; CHECK-NEXT: ret i32 [[TMP4]]
303+
;
304+
%1 = icmp ugt i32 %x, -10
305+
%2 = zext i1 %1 to i32
306+
%3 = icmp sgt i32 %x, -11
307+
%4 = select i1 %3, i32 %2, i32 -1
308+
ret i32 %4
309+
}
310+
311+
; Test with wrong false value
312+
define <4 x i32> @scmp_x_0_inverted_vec(<4 x i32> %x) {
313+
; CHECK-LABEL: define <4 x i32> @scmp_x_0_inverted_vec(
314+
; CHECK-SAME: <4 x i32> [[X:%.*]]) {
315+
; CHECK-NEXT: [[TMP1:%.*]] = icmp ne <4 x i32> [[X]], zeroinitializer
316+
; CHECK-NEXT: [[TMP2:%.*]] = zext <4 x i1> [[TMP1]] to <4 x i32>
317+
; CHECK-NEXT: [[TMP3:%.*]] = icmp sgt <4 x i32> [[X]], splat (i32 -1)
318+
; CHECK-NEXT: [[TMP4:%.*]] = select <4 x i1> [[TMP3]], <4 x i32> [[TMP2]], <4 x i32> <i32 -1, i32 -2, i32 -1, i32 -1>
319+
; CHECK-NEXT: ret <4 x i32> [[TMP4]]
320+
;
321+
%2 = icmp ne <4 x i32> %x, zeroinitializer
322+
%3 = zext <4 x i1> %2 to <4 x i32>
323+
%4 = icmp sgt <4 x i32> %x, <i32 -1, i32 -1, i32 -1, i32 -1>
324+
%5 = select <4 x i1> %4, <4 x i32> %3, <4 x i32> <i32 -1, i32 -2, i32 -1, i32 -1>
325+
ret <4 x i32> %5
326+
}

0 commit comments

Comments
 (0)