1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
|
; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt < %s -instcombine -S | FileCheck %s
; PR4374
define float @test1(float %x, float %y) {
; CHECK-LABEL: @test1(
; CHECK-NEXT: [[T1:%.*]] = fsub float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fsub float -0.000000e+00, [[T1]]
; CHECK-NEXT: ret float [[T2]]
;
%t1 = fsub float %x, %y
%t2 = fsub float -0.0, %t1
ret float %t2
}
; Can't do anything with the test above because -0.0 - 0.0 = -0.0, but if we have nsz:
; -(X - Y) --> Y - X
define float @neg_sub_nsz(float %x, float %y) {
; CHECK-LABEL: @neg_sub_nsz(
; CHECK-NEXT: [[TMP1:%.*]] = fsub nsz float [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: ret float [[TMP1]]
;
%t1 = fsub float %x, %y
%t2 = fsub nsz float -0.0, %t1
ret float %t2
}
; If the subtract has another use, we don't do the transform (even though it
; doesn't increase the IR instruction count) because we assume that fneg is
; easier to analyze and generally cheaper than generic fsub.
declare void @use(float)
declare void @use2(float, double)
define float @neg_sub_nsz_extra_use(float %x, float %y) {
; CHECK-LABEL: @neg_sub_nsz_extra_use(
; CHECK-NEXT: [[T1:%.*]] = fsub float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fsub nsz float -0.000000e+00, [[T1]]
; CHECK-NEXT: call void @use(float [[T1]])
; CHECK-NEXT: ret float [[T2]]
;
%t1 = fsub float %x, %y
%t2 = fsub nsz float -0.0, %t1
call void @use(float %t1)
ret float %t2
}
; With nsz: Z - (X - Y) --> Z + (Y - X)
define float @sub_sub_nsz(float %x, float %y, float %z) {
; CHECK-LABEL: @sub_sub_nsz(
; CHECK-NEXT: [[TMP1:%.*]] = fsub nsz float [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fadd nsz float [[TMP1]], [[Z:%.*]]
; CHECK-NEXT: ret float [[T2]]
;
%t1 = fsub float %x, %y
%t2 = fsub nsz float %z, %t1
ret float %t2
}
; Same as above: if 'Z' is not -0.0, swap fsub operands and convert to fadd.
define float @sub_sub_known_not_negzero(float %x, float %y) {
; CHECK-LABEL: @sub_sub_known_not_negzero(
; CHECK-NEXT: [[TMP1:%.*]] = fsub float [[Y:%.*]], [[X:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fadd float [[TMP1]], 4.200000e+01
; CHECK-NEXT: ret float [[T2]]
;
%t1 = fsub float %x, %y
%t2 = fsub float 42.0, %t1
ret float %t2
}
; <rdar://problem/7530098>
define double @test2(double %x, double %y) {
; CHECK-LABEL: @test2(
; CHECK-NEXT: [[T1:%.*]] = fadd double [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fsub double [[X]], [[T1]]
; CHECK-NEXT: ret double [[T2]]
;
%t1 = fadd double %x, %y
%t2 = fsub double %x, %t1
ret double %t2
}
; X - C --> X + (-C)
define float @constant_op1(float %x, float %y) {
; CHECK-LABEL: @constant_op1(
; CHECK-NEXT: [[R:%.*]] = fadd float [[X:%.*]], -4.200000e+01
; CHECK-NEXT: ret float [[R]]
;
%r = fsub float %x, 42.0
ret float %r
}
define <2 x float> @constant_op1_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @constant_op1_vec(
; CHECK-NEXT: [[R:%.*]] = fadd <2 x float> [[X:%.*]], <float -4.200000e+01, float 4.200000e+01>
; CHECK-NEXT: ret <2 x float> [[R]]
;
%r = fsub <2 x float> %x, <float 42.0, float -42.0>
ret <2 x float> %r
}
define <2 x float> @constant_op1_vec_undef(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @constant_op1_vec_undef(
; CHECK-NEXT: [[R:%.*]] = fadd <2 x float> [[X:%.*]], <float 0x7FF8000000000000, float 4.200000e+01>
; CHECK-NEXT: ret <2 x float> [[R]]
;
%r = fsub <2 x float> %x, <float undef, float -42.0>
ret <2 x float> %r
}
; X - (-Y) --> X + Y
define float @neg_op1(float %x, float %y) {
; CHECK-LABEL: @neg_op1(
; CHECK-NEXT: [[R:%.*]] = fadd float [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret float [[R]]
;
%negy = fsub float -0.0, %y
%r = fsub float %x, %negy
ret float %r
}
define <2 x float> @neg_op1_vec(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @neg_op1_vec(
; CHECK-NEXT: [[R:%.*]] = fadd <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[R]]
;
%negy = fsub <2 x float> <float -0.0, float -0.0>, %y
%r = fsub <2 x float> %x, %negy
ret <2 x float> %r
}
define <2 x float> @neg_op1_vec_undef(<2 x float> %x, <2 x float> %y) {
; CHECK-LABEL: @neg_op1_vec_undef(
; CHECK-NEXT: [[R:%.*]] = fadd <2 x float> [[X:%.*]], [[Y:%.*]]
; CHECK-NEXT: ret <2 x float> [[R]]
;
%negy = fsub <2 x float> <float -0.0, float undef>, %y
%r = fsub <2 x float> %x, %negy
ret <2 x float> %r
}
; Similar to above - but look through fpext/fptrunc casts to find the fneg.
define double @neg_ext_op1(float %a, double %b) {
; CHECK-LABEL: @neg_ext_op1(
; CHECK-NEXT: [[TMP1:%.*]] = fpext float [[A:%.*]] to double
; CHECK-NEXT: [[T3:%.*]] = fadd double [[TMP1]], [[B:%.*]]
; CHECK-NEXT: ret double [[T3]]
;
%t1 = fsub float -0.0, %a
%t2 = fpext float %t1 to double
%t3 = fsub double %b, %t2
ret double %t3
}
; Verify that vectors work too.
define <2 x float> @neg_trunc_op1(<2 x double> %a, <2 x float> %b) {
; CHECK-LABEL: @neg_trunc_op1(
; CHECK-NEXT: [[TMP1:%.*]] = fptrunc <2 x double> [[A:%.*]] to <2 x float>
; CHECK-NEXT: [[T3:%.*]] = fadd <2 x float> [[TMP1]], [[B:%.*]]
; CHECK-NEXT: ret <2 x float> [[T3]]
;
%t1 = fsub <2 x double> <double -0.0, double -0.0>, %a
%t2 = fptrunc <2 x double> %t1 to <2 x float>
%t3 = fsub <2 x float> %b, %t2
ret <2 x float> %t3
}
; No FMF needed, but they should propagate to the fadd.
define double @neg_ext_op1_fast(float %a, double %b) {
; CHECK-LABEL: @neg_ext_op1_fast(
; CHECK-NEXT: [[TMP1:%.*]] = fpext float [[A:%.*]] to double
; CHECK-NEXT: [[T3:%.*]] = fadd fast double [[TMP1]], [[B:%.*]]
; CHECK-NEXT: ret double [[T3]]
;
%t1 = fsub float -0.0, %a
%t2 = fpext float %t1 to double
%t3 = fsub fast double %b, %t2
ret double %t3
}
; Extra use should prevent the transform.
define float @neg_ext_op1_extra_use(half %a, float %b) {
; CHECK-LABEL: @neg_ext_op1_extra_use(
; CHECK-NEXT: [[T1:%.*]] = fsub half 0xH8000, [[A:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fpext half [[T1]] to float
; CHECK-NEXT: [[T3:%.*]] = fsub float [[B:%.*]], [[T2]]
; CHECK-NEXT: call void @use(float [[T2]])
; CHECK-NEXT: ret float [[T3]]
;
%t1 = fsub half -0.0, %a
%t2 = fpext half %t1 to float
%t3 = fsub float %b, %t2
call void @use(float %t2)
ret float %t3
}
; One-use fptrunc is always hoisted above fneg, so the corresponding
; multi-use bug for fptrunc isn't visible with a fold starting from
; the last fsub.
define float @neg_trunc_op1_extra_use(double %a, float %b) {
; CHECK-LABEL: @neg_trunc_op1_extra_use(
; CHECK-NEXT: [[TMP1:%.*]] = fptrunc double [[A:%.*]] to float
; CHECK-NEXT: [[T2:%.*]] = fsub float -0.000000e+00, [[TMP1]]
; CHECK-NEXT: [[T3:%.*]] = fadd float [[TMP1]], [[B:%.*]]
; CHECK-NEXT: call void @use(float [[T2]])
; CHECK-NEXT: ret float [[T3]]
;
%t1 = fsub double -0.0, %a
%t2 = fptrunc double %t1 to float
%t3 = fsub float %b, %t2
call void @use(float %t2)
ret float %t3
}
; Extra uses should prevent the transform.
define float @neg_trunc_op1_extra_uses(double %a, float %b) {
; CHECK-LABEL: @neg_trunc_op1_extra_uses(
; CHECK-NEXT: [[T1:%.*]] = fsub double -0.000000e+00, [[A:%.*]]
; CHECK-NEXT: [[T2:%.*]] = fptrunc double [[T1]] to float
; CHECK-NEXT: [[T3:%.*]] = fsub float [[B:%.*]], [[T2]]
; CHECK-NEXT: call void @use2(float [[T2]], double [[T1]])
; CHECK-NEXT: ret float [[T3]]
;
%t1 = fsub double -0.0, %a
%t2 = fptrunc double %t1 to float
%t3 = fsub float %b, %t2
call void @use2(float %t2, double %t1)
ret float %t3
}
; Don't negate a constant expression to form fadd and induce infinite looping:
; https://bugs.llvm.org/show_bug.cgi?id=37605
@b = external global i16, align 1
define float @PR37605(float %conv) {
; CHECK-LABEL: @PR37605(
; CHECK-NEXT: [[SUB:%.*]] = fsub float [[CONV:%.*]], bitcast (i32 ptrtoint (i16* @b to i32) to float)
; CHECK-NEXT: ret float [[SUB]]
;
%sub = fsub float %conv, bitcast (i32 ptrtoint (i16* @b to i32) to float)
ret float %sub
}
|
