1 ; RUN: opt < %s -slsr -gvn -dce -S | FileCheck %s
3 target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
5 declare i32 @foo(i32 %a)
7 define i32 @slsr1(i32 %b, i32 %s) {
10 %mul0 = mul i32 %b, %s
13 %v0 = call i32 @foo(i32 %mul0)
15 ; v1 = foo((b + 1) * s);
17 %mul1 = mul i32 %b1, %s
18 %v1 = call i32 @foo(i32 %mul1)
20 ; v2 = foo((b + 2) * s);
22 %mul2 = mul i32 %b2, %s
23 %v2 = call i32 @foo(i32 %mul2)
25 ; return v0 + v1 + v2;
32 ; v1 = foo((a + 1) * b)
33 ; v2 = foo(a * (b + 1))
34 ; v3 = foo((a + 1) * (b + 1))
35 define i32 @slsr2(i32 %a, i32 %b) {
36 ; CHECK-LABEL: @slsr2(
39 %mul0 = mul i32 %a, %b
42 %mul1 = mul i32 %a1, %b
43 %mul2 = mul i32 %a, %b1
44 %mul3 = mul i32 %a1, %b1
46 %v0 = call i32 @foo(i32 %mul0)
47 %v1 = call i32 @foo(i32 %mul1)
48 %v2 = call i32 @foo(i32 %mul2)
49 %v3 = call i32 @foo(i32 %mul3)
57 ; The bump is a multiple of the stride.
60 ; v1 = foo((b + 2) * s);
61 ; v2 = foo((b + 4) * s);
62 ; return v0 + v1 + v2;
69 ; mul1 = mul0 + bump; // GVN ensures mul1 and mul2 use the same bump.
73 ; return v0 + v1 + v2;
74 define i32 @slsr3(i32 %b, i32 %s) {
75 ; CHECK-LABEL: @slsr3(
76 %mul0 = mul i32 %b, %s
78 %v0 = call i32 @foo(i32 %mul0)
81 %mul1 = mul i32 %b1, %s
82 ; CHECK: [[BUMP:%[a-zA-Z0-9]+]] = mul i32 %s, 2
83 ; CHECK: %mul1 = add i32 %mul0, [[BUMP]]
84 %v1 = call i32 @foo(i32 %mul1)
87 %mul2 = mul i32 %b2, %s
88 ; CHECK: %mul2 = add i32 %mul1, [[BUMP]]
89 %v2 = call i32 @foo(i32 %mul2)
96 ; Do not rewrite a candidate if its potential basis does not dominate it.
100 ; v1 = foo((a + 1) * b);
102 define i32 @not_dominate(i1 %cond, i32 %a, i32 %b) {
103 ; CHECK-LABEL: @not_dominate(
106 br i1 %cond, label %then, label %merge
109 %mul0 = mul i32 %a, %b
110 ; CHECK: %mul0 = mul i32 %a, %b
111 %v0 = call i32 @foo(i32 %mul0)
115 %v0.phi = phi i32 [ 0, %entry ], [ %mul0, %then ]
116 %mul1 = mul i32 %a1, %b
117 ; CHECK: %mul1 = mul i32 %a1, %b
118 %v1 = call i32 @foo(i32 %mul1)
119 %sum = add i32 %v0.phi, %v1