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
|
;=========================== begin_copyright_notice ============================
;
; Copyright (C) 2025 Intel Corporation
;
; SPDX-License-Identifier: MIT
;
;============================ end_copyright_notice =============================
; REQUIRES: llvm-14-plus, regkeys
; RUN: igc_opt -opaque-pointers -igc-divrem-increment-reduction -regkey SanitizeDivRemIncrementDivisorIsZero=1 -S < %s | FileCheck %s
; RUN: igc_opt -opaque-pointers -igc-divrem-increment-reduction -regkey GuardDivRemIncrementDividendOverflow=1 -S < %s | FileCheck %s --check-prefix=OVF
; ------------------------------------------------
;
; Simple increment reduction, optimization for second udiv/urem pair based on result of first.
; - Test divisor is zero sanitization logic
; - Test nuw flag detection
; Ensure only can only look past Add/Or/Sub for finding true offset
; Do CSE when two DivRemGroup offsets are the same
; - Test nuw flag detection
; Check that DivRemPairs are still matched when udiv/urem are in reverse order or are not back to back
; Simple decrement reduction
; ------------------------------------------------
; CHECK-LABEL: @test_int_divrem_increment_reduction
; OVF-LABEL: @test_int_divrem_increment_reduction
define void @test_int_divrem_increment_reduction(i32 %a, i32 %b, ptr %dest1, ptr %dest2) {
; First DivRemGroup, retain
%quo = udiv i32 %a, %b
%rem = urem i32 %a, %b
%next = add i32 %a, 1
; Next DivRemGroup, DivRemPair optimization
; CHECK: [[PREINC_REM:%.*]] = add i32 %rem, 1
; CHECK-NEXT: [[PREINC_REM_TEST:%.*]] = icmp eq i32 [[PREINC_REM]], %b
; CHECK-NEXT: [[PREINC_QUO:%.*]] = add i32 %quo, 1
; CHECK-NEXT: [[NEW_NEXT_QUO:%.*]] = select i1 [[PREINC_REM_TEST]], i32 [[PREINC_QUO]], i32 %quo
; CHECK-NEXT: [[NEW_NEXT_REM:%.*]] = select i1 [[PREINC_REM_TEST]], i32 0, i32 [[PREINC_REM]]
; Sanitize logic to avoid UB
; CHECK-NEXT: [[DIVISOR_IS_ZERO:%.*]] = icmp eq i32 %b, 0
; CHECK-NEXT: [[FINAL_NEXT_QUO:%.*]] = select i1 [[DIVISOR_IS_ZERO]], i32 -1, i32 [[NEW_NEXT_QUO]]
; CHECK-NEXT: [[FINAL_NEXT_REM:%.*]] = select i1 [[DIVISOR_IS_ZERO]], i32 -1, i32 [[NEW_NEXT_REM]]
; Check that original udiv/urem is deleted
; CHECK-NOT: udiv i32 %next, %b
; CHECK-NOT: urem i32 %next, %b
; nuw flag not present on increment instruction, did not optimize, so original udiv/urem should still be present
; OVF: udiv i32 %next, %b
; OVF: urem i32 %next, %b
%next_quo = udiv i32 %next, %b
%next_rem = urem i32 %next, %b
; Check that original uses are replaced
; CHECK: store i32 [[FINAL_NEXT_QUO]], ptr %dest1
; CHECK: store i32 [[FINAL_NEXT_REM]], ptr %dest2
; Unchanged insts
; OVF: store i32 %next_quo, ptr %dest1
; OVF: store i32 %next_rem, ptr %dest2
store i32 %next_quo, ptr %dest1
store i32 %next_rem, ptr %dest2
ret void
}
; We can't assume what the dividend will look like
; CHECK-LABEL: @test_arbitrary_dividend
; OVF-LABEL: @test_arbitrary_dividend
define i32 @test_arbitrary_dividend(float %a, i32 %b) {
; CHECK: [[BC:%.*]] = bitcast float %a to i32
; CHECK-NEXT: [[UDIV:%.*]] = udiv i32 [[BC]], %b
; CHECK-NEXT: [[UREM:%.*]] = urem i32 [[BC]], %b
%bc = bitcast float %a to i32
%1 = udiv i32 %bc, %b
%2 = urem i32 %bc, %b
ret i32 %2
}
; Perform CSE for indirectly equal DivRemPairs
; CHECK-LABEL: @test_cse
; OVF-LABEL: @test_cse
define void @test_cse(i32 %a, i32 %b, ptr %dest) {
%quo = udiv i32 %a, %b
%rem = urem i32 %a, %b
; Next DivRemGroup, DivRemPair optimization, even when GuardDivRemIncrementDividendOverflow=1, because of add nuw i32 %a, 1
; Without flag it should still be optimized, but similar check to first test, so skip CHECK prefix
; OVF: [[PREINC_REM:%.*]] = add i32 %rem, 1
; OVF-NEXT: [[PREINC_REM_TEST:%.*]] = icmp eq i32 [[PREINC_REM]], %b
; OVF-NEXT: [[PREINC_QUO:%.*]] = add i32 %quo, 1
; OVF-NEXT: [[NEW_NEXT_QUO:%.*]] = select i1 [[PREINC_REM_TEST]], i32 [[PREINC_QUO]], i32 %quo
; OVF-NEXT: [[NEW_NEXT_REM:%.*]] = select i1 [[PREINC_REM_TEST]], i32 0, i32 [[PREINC_REM]]
%next = add nuw i32 %a, 1
%next_quo = udiv i32 %next, %b
%next_rem = urem i32 %next, %b
; sub has no nuw flag, so do not do replacement/CSE
; OVF: udiv i32 %next2, %b
; OVF: urem i32 %next2, %b
%next2 = sub i32 %next, 1
%next2_quo = udiv i32 %next2, %b
%next2_rem = urem i32 %next2, %b
; Check that uses are replaced
; OVF: %sum_quo = add i32 %quo, [[NEW_NEXT_QUO]]
%sum_quo = add i32 %quo, %next_quo
; Check that %next2_quo and %next2_rem are CSE'ed to %quo and %rem, because %a + 1 - 1 = %a
; CHECK: %sum_quo2 = add i32 %sum_quo, %quo
; OVF: %sum_quo2 = add i32 %sum_quo, %next2_quo
%sum_quo2 = add i32 %sum_quo, %next2_quo
; OVF: %sum_rem = add i32 %rem, [[NEW_NEXT_REM]]
%sum_rem = add i32 %rem, %next_rem
; CHECK: %sum_rem2 = add i32 %sum_rem, %rem
; OVF: %sum_rem2 = add i32 %sum_rem, %next2_rem
%sum_rem2 = add i32 %sum_rem, %next2_rem
%res = add i32 %sum_quo2, %sum_rem2
store i32 %res, ptr %dest
ret void
}
; reverse order and non-back-to-back udiv/urem, still do optimization
; CHECK-LABEL: @test_reverse_order_and_non_back_to_back
define void @test_reverse_order_and_non_back_to_back(i32 %a, i32 %b, ptr %dest) {
%rem = urem i32 %a, %b
%dest_i = ptrtoint ptr %dest to i32
%dest1_i = add i32 %dest_i, %a
%quo = udiv i32 %a, %b
%next = add nuw i32 %a, 1
; Next DivRemGroup, DivRemPair optimization
; CHECK: [[PREINC_REM:%.*]] = add i32 %rem, 1
; CHECK-NEXT: [[PREINC_REM_TEST:%.*]] = icmp eq i32 [[PREINC_REM]], %b
; CHECK-NEXT: [[PREINC_QUO:%.*]] = add i32 %quo, 1
; CHECK-NEXT: [[NEW_NEXT_QUO:%.*]] = select i1 [[PREINC_REM_TEST]], i32 [[PREINC_QUO]], i32 %quo
; CHECK-NEXT: [[NEW_NEXT_REM:%.*]] = select i1 [[PREINC_REM_TEST]], i32 0, i32 [[PREINC_REM]]
; Sanitize logic to avoid UB, enabled by default
; CHECK-NEXT: [[DIVISOR_IS_ZERO:%.*]] = icmp eq i32 %b, 0
; CHECK-NEXT: [[FINAL_NEXT_QUO:%.*]] = select i1 [[DIVISOR_IS_ZERO]], i32 -1, i32 [[NEW_NEXT_QUO]]
; CHECK-NEXT: [[FINAL_NEXT_REM:%.*]] = select i1 [[DIVISOR_IS_ZERO]], i32 -1, i32 [[NEW_NEXT_REM]]
; Check that original udiv/urem is deleted
; CHECK-NOT: urem i32 %next, %b
; CHECK-NOT: udiv i32 %next, %b
%next_rem = urem i32 %next, %b
%next_quo = udiv i32 %next, %b
%dest2_i = add i32 %dest_i, %b
%dest1 = inttoptr i32 %dest1_i to ptr
%dest2 = inttoptr i32 %dest2_i to ptr
; CHECK: store i32 [[FINAL_NEXT_QUO]], ptr %dest1
; CHECK: store i32 [[FINAL_NEXT_REM]], ptr %dest2
store i32 %next_quo, ptr %dest1
store i32 %next_rem, ptr %dest2
ret void
}
; CHECK-LABEL: @test_int_divrem_decrement_reduction
define void @test_int_divrem_decrement_reduction(i32 %a, i32 %b, ptr %dest1, ptr %dest2) {
; First DivRemGroup, retain
%quo = udiv i32 %a, %b
%rem = urem i32 %a, %b
%next = sub i32 %a, 1
; Next DivRemGroup, TODO: Negative offset DivRemPair optimization
; Ensure unoptimized result emitted for now
; CHECK: udiv i32 %next, %b
; CHECK: urem i32 %next, %b
%next_quo = udiv i32 %next, %b
%next_rem = urem i32 %next, %b
store i32 %next_quo, ptr %dest1
store i32 %next_rem, ptr %dest2
ret void
}
|
