Skip to content

Commit 088d636

Browse files
adam-bzowskiadam-bzowski
authored
[ValueTracking] Fix a bug for signed min-max clamping (#121206)
Correctly handle the case where the clamp is over the full range. This fixes an issue introduced in #121206.
1 parent 3496e96 commit 088d636

File tree

2 files changed

+161
-38
lines changed

2 files changed

+161
-38
lines changed

‎llvm/lib/Analysis/ValueTracking.cpp‎

Lines changed: 2 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -1119,7 +1119,8 @@ static void unionWithMinMaxIntrinsicClamp(const IntrinsicInst *II,
11191119
KnownBits &Known) {
11201120
const APInt *CLow, *CHigh;
11211121
if (isSignedMinMaxIntrinsicClamp(II, CLow, CHigh))
1122-
Known = Known.unionWith(ConstantRange(*CLow, *CHigh + 1).toKnownBits());
1122+
Known = Known.unionWith(
1123+
ConstantRange::getNonEmpty(*CLow, *CHigh + 1).toKnownBits());
11231124
}
11241125

11251126
static void computeKnownBitsFromOperator(const Operator *I,

‎llvm/test/Analysis/ValueTracking/knownbits-trunc-with-min-max-clamp.ll‎

Lines changed: 159 additions & 37 deletions
Original file line numberDiff line numberDiff line change
@@ -2,7 +2,8 @@
22
; RUN: opt < %s -passes=aggressive-instcombine -S | FileCheck %s
33

44
; The LIT tests rely on i32, i16 and i8 being valid machine types.
5-
target datalayout = "n8:16:32"
5+
; The bounds checking tests require also i64 and i128.
6+
target datalayout = "n8:16:32:64:128"
67

78
; This LIT test checks if TruncInstCombine pass correctly recognizes the
89
; constraints from a signed min-max clamp. The clamp is a sequence of smin and
@@ -12,6 +13,11 @@ target datalayout = "n8:16:32"
1213
; of smin and smax:
1314
; a) y = smax(smin(x, upper_limit), lower_limit)
1415
; b) y = smin(smax(x, lower_limit), upper_limit)
16+
;
17+
; The clamp is used in TruncInstCombine.cpp pass (as part of aggressive-instcombine)
18+
; to optimize extensions and truncations of lshr. This is what is tested here.
19+
; The pass also optimizes extensions and truncations of other binary operators,
20+
; but in such cases the smin-smax clamp may not be used.
1521

1622
define i8 @test_0a(i16 %x) {
1723
; CHECK-LABEL: define i8 @test_0a(
@@ -47,6 +53,8 @@ define i8 @test_0b(i16 %x) {
4753
ret i8 %b.trunc
4854
}
4955

56+
; The following two tests contain add instead of lshr.
57+
; The optimization works here as well.
5058
define i8 @test_1a(i16 %x) {
5159
; CHECK-LABEL: define i8 @test_1a(
5260
; CHECK-SAME: i16 [[X:%.*]]) {
@@ -81,19 +89,23 @@ define i8 @test_1b(i16 %x) {
8189
ret i8 %b.trunc
8290
}
8391

92+
; Tests for clamping with negative min and max.
93+
94+
; With sext no optimization occurs.
8495
define i8 @test_2a(i16 %x) {
8596
; CHECK-LABEL: define i8 @test_2a(
8697
; CHECK-SAME: i16 [[X:%.*]]) {
8798
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 -1)
8899
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -31)
89-
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
90-
; CHECK-NEXT: [[B:%.*]] = add i8 [[A]], 2
91-
; CHECK-NEXT: ret i8 [[B]]
100+
; CHECK-NEXT: [[A:%.*]] = sext i16 [[TMP2]] to i32
101+
; CHECK-NEXT: [[B:%.*]] = lshr i32 [[A]], 2
102+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i32 [[B]] to i8
103+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
92104
;
93105
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 -1)
94106
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -31)
95107
%a = sext i16 %2 to i32
96-
%b = add i32 %a, 2
108+
%b = lshr i32 %a, 2
97109
%b.trunc = trunc i32 %b to i8
98110
ret i8 %b.trunc
99111
}
@@ -103,31 +115,69 @@ define i8 @test_2b(i16 %x) {
103115
; CHECK-SAME: i16 [[X:%.*]]) {
104116
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 -31)
105117
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 -1)
106-
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
107-
; CHECK-NEXT: [[B:%.*]] = add i8 [[A]], 2
108-
; CHECK-NEXT: ret i8 [[B]]
118+
; CHECK-NEXT: [[A:%.*]] = sext i16 [[TMP2]] to i32
119+
; CHECK-NEXT: [[B:%.*]] = lshr i32 [[A]], 2
120+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i32 [[B]] to i8
121+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
109122
;
110123
%1 = tail call i16 @llvm.smax.i16(i16 %x, i16 -31)
111124
%2 = tail call i16 @llvm.smin.i16(i16 %1, i16 -1)
112125
%a = sext i16 %2 to i32
113-
%b = add i32 %a, 2
126+
%b = lshr i32 %a, 2
127+
%b.trunc = trunc i32 %b to i8
128+
ret i8 %b.trunc
129+
}
130+
131+
; With zext the optimization occurs.
132+
define i8 @test_2c(i16 %x) {
133+
; CHECK-LABEL: define i8 @test_2c(
134+
; CHECK-SAME: i16 [[X:%.*]]) {
135+
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 -1)
136+
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -31)
137+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
138+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
139+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
140+
;
141+
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 -1)
142+
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -31)
143+
%a = zext i16 %2 to i32
144+
%b = lshr i32 %a, 2
114145
%b.trunc = trunc i32 %b to i8
115146
ret i8 %b.trunc
116147
}
117148

149+
define i8 @test_2d(i16 %x) {
150+
; CHECK-LABEL: define i8 @test_2d(
151+
; CHECK-SAME: i16 [[X:%.*]]) {
152+
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 -31)
153+
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 -1)
154+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
155+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
156+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
157+
;
158+
%1 = tail call i16 @llvm.smax.i16(i16 %x, i16 -31)
159+
%2 = tail call i16 @llvm.smin.i16(i16 %1, i16 -1)
160+
%a = zext i16 %2 to i32
161+
%b = lshr i32 %a, 2
162+
%b.trunc = trunc i32 %b to i8
163+
ret i8 %b.trunc
164+
}
165+
166+
; Tests for clamping with mixed-signed min and max.
167+
; With zext the optimization occurs.
118168
define i8 @test_3a(i16 %x) {
119169
; CHECK-LABEL: define i8 @test_3a(
120170
; CHECK-SAME: i16 [[X:%.*]]) {
121171
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 31)
122172
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -31)
123-
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
124-
; CHECK-NEXT: [[B:%.*]] = add i8 [[A]], 2
125-
; CHECK-NEXT: ret i8 [[B]]
173+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
174+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
175+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
126176
;
127177
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 31)
128178
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -31)
129-
%a = sext i16 %2 to i32
130-
%b = add i32 %a, 2
179+
%a = zext i16 %2 to i32
180+
%b = lshr i32 %a, 2
131181
%b.trunc = trunc i32 %b to i8
132182
ret i8 %b.trunc
133183
}
@@ -137,31 +187,32 @@ define i8 @test_3b(i16 %x) {
137187
; CHECK-SAME: i16 [[X:%.*]]) {
138188
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smax.i16(i16 [[X]], i16 -31)
139189
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smin.i16(i16 [[TMP1]], i16 31)
140-
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
141-
; CHECK-NEXT: [[B:%.*]] = add i8 [[A]], 2
142-
; CHECK-NEXT: ret i8 [[B]]
190+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
191+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
192+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
143193
;
144194
%1 = tail call i16 @llvm.smax.i16(i16 %x, i16 -31)
145195
%2 = tail call i16 @llvm.smin.i16(i16 %1, i16 31)
146-
%a = sext i16 %2 to i32
147-
%b = add i32 %a, 2
196+
%a = zext i16 %2 to i32
197+
%b = lshr i32 %a, 2
148198
%b.trunc = trunc i32 %b to i8
149199
ret i8 %b.trunc
150200
}
151201

202+
; Optimizations with vector types.
152203
define <16 x i8> @test_vec_1a(<16 x i16> %x) {
153204
; CHECK-LABEL: define <16 x i8> @test_vec_1a(
154205
; CHECK-SAME: <16 x i16> [[X:%.*]]) {
155206
; CHECK-NEXT: [[TMP1:%.*]] = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> [[X]], <16 x i16> splat (i16 127))
156207
; CHECK-NEXT: [[TMP2:%.*]] = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> [[TMP1]], <16 x i16> zeroinitializer)
157208
; CHECK-NEXT: [[A:%.*]] = trunc <16 x i16> [[TMP2]] to <16 x i8>
158-
; CHECK-NEXT: [[B:%.*]] = add <16 x i8> [[A]], splat (i8 2)
209+
; CHECK-NEXT: [[B:%.*]] = lshr <16 x i8> [[A]], splat (i8 2)
159210
; CHECK-NEXT: ret <16 x i8> [[B]]
160211
;
161212
%1 = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> %x, <16 x i16> splat (i16 127))
162213
%2 = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> %1, <16 x i16> zeroinitializer)
163214
%a = sext <16 x i16> %2 to <16 x i32>
164-
%b = add <16 x i32> %a, splat (i32 2)
215+
%b = lshr <16 x i32> %a, splat (i32 2)
165216
%b.trunc = trunc <16 x i32> %b to <16 x i8>
166217
ret <16 x i8> %b.trunc
167218
}
@@ -172,13 +223,13 @@ define <16 x i8> @test_vec_1b(<16 x i16> %x) {
172223
; CHECK-NEXT: [[TMP1:%.*]] = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> [[X]], <16 x i16> zeroinitializer)
173224
; CHECK-NEXT: [[TMP2:%.*]] = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> [[TMP1]], <16 x i16> splat (i16 127))
174225
; CHECK-NEXT: [[A:%.*]] = trunc <16 x i16> [[TMP2]] to <16 x i8>
175-
; CHECK-NEXT: [[B:%.*]] = add <16 x i8> [[A]], splat (i8 2)
226+
; CHECK-NEXT: [[B:%.*]] = lshr <16 x i8> [[A]], splat (i8 2)
176227
; CHECK-NEXT: ret <16 x i8> [[B]]
177228
;
178229
%1 = tail call <16 x i16> @llvm.smax.v16i16(<16 x i16> %x, <16 x i16> zeroinitializer)
179230
%2 = tail call <16 x i16> @llvm.smin.v16i16(<16 x i16> %1, <16 x i16> splat (i16 127))
180231
%a = sext <16 x i16> %2 to <16 x i32>
181-
%b = add <16 x i32> %a, splat (i32 2)
232+
%b = lshr <16 x i32> %a, splat (i32 2)
182233
%b.trunc = trunc <16 x i32> %b to <16 x i8>
183234
ret <16 x i8> %b.trunc
184235
}
@@ -217,14 +268,14 @@ define i8 @test_bounds_1(i16 %x) {
217268
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 127)
218269
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
219270
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
220-
; CHECK-NEXT: [[SHR:%.*]] = ashr i8 [[A]], 7
221-
; CHECK-NEXT: ret i8 [[SHR]]
271+
; CHECK-NEXT: [[B:%.*]] = lshr i8 [[A]], 7
272+
; CHECK-NEXT: ret i8 [[B]]
222273
;
223274
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 127)
224275
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
225276
%a = sext i16 %2 to i32
226-
%shr = ashr i32 %a, 7
227-
%b.trunc = trunc i32 %shr to i8
277+
%b = lshr i32 %a, 7
278+
%b.trunc = trunc i32 %b to i8
228279
ret i8 %b.trunc
229280
}
230281

@@ -234,15 +285,15 @@ define i8 @test_bounds_2(i16 %x) {
234285
; CHECK-SAME: i16 [[X:%.*]]) {
235286
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 128)
236287
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 0)
237-
; CHECK-NEXT: [[SHR:%.*]] = ashr i16 [[TMP2]], 7
238-
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[SHR]] to i8
239-
; CHECK-NEXT: ret i8 [[B_TRUNC]]
288+
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
289+
; CHECK-NEXT: [[B:%.*]] = lshr i8 [[A]], 7
290+
; CHECK-NEXT: ret i8 [[B]]
240291
;
241292
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 128)
242293
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 0)
243294
%a = sext i16 %2 to i32
244-
%shr = ashr i32 %a, 7
245-
%b.trunc = trunc i32 %shr to i8
295+
%b = lshr i32 %a, 7
296+
%b.trunc = trunc i32 %b to i8
246297
ret i8 %b.trunc
247298
}
248299

@@ -253,14 +304,85 @@ define i8 @test_bounds_3(i16 %x) {
253304
; CHECK-SAME: i16 [[X:%.*]]) {
254305
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 32767)
255306
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 32752)
256-
; CHECK-NEXT: [[A:%.*]] = trunc i16 [[TMP2]] to i8
257-
; CHECK-NEXT: [[AND:%.*]] = and i8 [[A]], -1
258-
; CHECK-NEXT: ret i8 [[AND]]
307+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
308+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
309+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
259310
;
260311
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 32767)
261312
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 32752)
262313
%a = sext i16 %2 to i32
263-
%and = and i32 %a, 255
264-
%b.trunc = trunc i32 %and to i8
314+
%b = lshr i32 %a, 2
315+
%b.trunc = trunc i32 %b to i8
316+
ret i8 %b.trunc
317+
}
318+
319+
; Here min = 128 is greater than max = 0.
320+
define i8 @test_bounds_4(i16 %x) {
321+
; CHECK-LABEL: define i8 @test_bounds_4(
322+
; CHECK-SAME: i16 [[X:%.*]]) {
323+
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 0)
324+
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 128)
325+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
326+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
327+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
328+
;
329+
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 0)
330+
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 128)
331+
%a = sext i16 %2 to i32
332+
%b = lshr i32 %a, 2
333+
%b.trunc = trunc i32 %b to i8
334+
ret i8 %b.trunc
335+
}
336+
337+
; The following 3 tests check the situation where min and max are minimal and
338+
; maximal signed values. No transformations should occur here.
339+
define i8 @test_bounds_5(i16 %x) {
340+
; CHECK-LABEL: define i8 @test_bounds_5(
341+
; CHECK-SAME: i16 [[X:%.*]]) {
342+
; CHECK-NEXT: [[TMP1:%.*]] = tail call i16 @llvm.smin.i16(i16 [[X]], i16 32767)
343+
; CHECK-NEXT: [[TMP2:%.*]] = tail call i16 @llvm.smax.i16(i16 [[TMP1]], i16 -32768)
344+
; CHECK-NEXT: [[B:%.*]] = lshr i16 [[TMP2]], 2
345+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i16 [[B]] to i8
346+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
347+
;
348+
%1 = tail call i16 @llvm.smin.i16(i16 %x, i16 32767)
349+
%2 = tail call i16 @llvm.smax.i16(i16 %1, i16 -32768)
350+
%a = zext i16 %2 to i32
351+
%b = lshr i32 %a, 2
352+
%b.trunc = trunc i32 %b to i8
353+
ret i8 %b.trunc
354+
}
355+
356+
define i8 @test_bounds_6(i32 %x) {
357+
; CHECK-LABEL: define i8 @test_bounds_6(
358+
; CHECK-SAME: i32 [[X:%.*]]) {
359+
; CHECK-NEXT: [[TMP1:%.*]] = tail call i32 @llvm.smin.i32(i32 [[X]], i32 2147483647)
360+
; CHECK-NEXT: [[TMP2:%.*]] = tail call i32 @llvm.smax.i32(i32 [[TMP1]], i32 -2147483648)
361+
; CHECK-NEXT: [[B:%.*]] = lshr i32 [[TMP2]], 2
362+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i32 [[B]] to i8
363+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
364+
;
365+
%1 = tail call i32 @llvm.smin.i32(i32 %x, i32 2147483647)
366+
%2 = tail call i32 @llvm.smax.i32(i32 %1, i32 -2147483648)
367+
%a = zext i32 %2 to i64
368+
%b = lshr i64 %a, 2
369+
%b.trunc = trunc i64 %b to i8
370+
ret i8 %b.trunc
371+
}
372+
373+
define i8 @test_bounds_7(i64 %x) {
374+
; CHECK-LABEL: define i8 @test_bounds_7(
375+
; CHECK-SAME: i64 [[X:%.*]]) {
376+
; CHECK-NEXT: [[TMP1:%.*]] = tail call i64 @llvm.smin.i64(i64 [[X]], i64 9223372036854775807)
377+
; CHECK-NEXT: [[TMP2:%.*]] = tail call i64 @llvm.smax.i64(i64 [[TMP1]], i64 -9223372036854775808)
378+
; CHECK-NEXT: [[B:%.*]] = lshr i64 [[TMP2]], 2
379+
; CHECK-NEXT: [[B_TRUNC:%.*]] = trunc i64 [[B]] to i8
380+
; CHECK-NEXT: ret i8 [[B_TRUNC]]
381+
;
382+
%1 = tail call i64 @llvm.smin.i64(i64 %x, i64 9223372036854775807)
383+
%2 = tail call i64 @llvm.smax.i64(i64 %1, i64 -9223372036854775808)
384+
%a = zext i64 %2 to i128
385+
%b = lshr i128 %a, 2
386+
%b.trunc = trunc i128 %b to i8
265387
ret i8 %b.trunc
266388
}

0 commit comments

Comments
 (0)