Adam Nemet | c520822 | 2016-09-06 16:08:33 +0000 | [diff] [blame] | 1 | ; RUN: opt -S -jump-threading < %s | FileCheck %s |
| 2 | |
| 3 | ; Check that based solely on static profile estimation we don't update the |
| 4 | ; branch-weight metadata. Even if the function has an entry frequency, a |
| 5 | ; completely cold part of the CFG may be statically estimated. |
| 6 | |
| 7 | ; For example in the loop below, jump threading would update the weight of the |
| 8 | ; loop-exiting branch to 0, drastically inflating the frequency of the loop |
| 9 | ; (in the range of billions). |
| 10 | ; |
| 11 | ; This is the CFG of the loop. There is no run-time profile info for edges |
| 12 | ; inside the loop, so branch and block frequencies are estimated as shown: |
| 13 | ; |
| 14 | ; check_1 (16) |
| 15 | ; (8) / | |
| 16 | ; eq_1 | (8) |
| 17 | ; \ | |
| 18 | ; check_2 (16) |
| 19 | ; (8) / | |
| 20 | ; eq_2 | (8) |
| 21 | ; \ | |
| 22 | ; check_3 (16) |
| 23 | ; (1) / | |
| 24 | ; (loop exit) | (15) |
| 25 | ; | |
| 26 | ; (back edge) |
| 27 | ; |
| 28 | ; First we thread eq_1->check_2 to check_3. Frequencies are updated to remove |
| 29 | ; the frequency of eq_1 from check_2 and then the false edge leaving check_2 |
| 30 | ; (changed frequencies are highlighted with * *): |
| 31 | ; |
| 32 | ; check_1 (16) |
| 33 | ; (8) / | |
| 34 | ; eq_1~ | (8) |
| 35 | ; / | |
| 36 | ; / check_2 (*8*) |
| 37 | ; / (8) / | |
| 38 | ; \ eq_2 | (*0*) |
| 39 | ; \ \ | |
| 40 | ; ` --- check_3 (16) |
| 41 | ; (1) / | |
| 42 | ; (loop exit) | (15) |
| 43 | ; | |
| 44 | ; (back edge) |
| 45 | ; |
| 46 | ; Next we thread eq_1->check_3 and eq_2->check_3 to check_1 as new back edges. |
| 47 | ; Frequencies are updated to remove the frequency of eq_1 and eq_3 from |
| 48 | ; check_3 and then the false edge leaving check_3 (changed frequencies are |
| 49 | ; highlighted with * *): |
| 50 | ; |
| 51 | ; check_1 (16) |
| 52 | ; (8) / | |
| 53 | ; eq_1~ | (8) |
| 54 | ; / | |
| 55 | ; / check_2 (*8*) |
| 56 | ; / (8) / | |
| 57 | ; /-- eq_2~ | (*0*) |
| 58 | ; (back edge) | |
| 59 | ; check_3 (*0*) |
| 60 | ; (*0*) / | |
| 61 | ; (loop exit) | (*0*) |
| 62 | ; | |
| 63 | ; (back edge) |
| 64 | ; |
| 65 | ; As a result, the loop exit edge ends up with 0 frequency which in turn makes |
| 66 | ; the loop header to have maximum frequency. |
| 67 | |
| 68 | declare void @bar() |
| 69 | |
| 70 | define void @foo(i32 *%p, i32 %n) !prof !0 { |
| 71 | entry: |
| 72 | %enter_loop = icmp eq i32 %n, 0 |
| 73 | br i1 %enter_loop, label %exit, label %check_1, !prof !1 |
| 74 | ; CHECK: br i1 %enter_loop, label %exit, label %check_1, !prof !1 |
| 75 | |
| 76 | check_1: |
| 77 | %v = load i32, i32* %p |
| 78 | %cond1 = icmp eq i32 %v, 1 |
| 79 | br i1 %cond1, label %eq_1, label %check_2 |
| 80 | ; No metadata: |
| 81 | ; CHECK: br i1 %cond1, label %check_2.thread, label %check_2{{$}} |
| 82 | |
| 83 | eq_1: |
| 84 | call void @bar() |
| 85 | br label %check_2 |
| 86 | ; Verify the new backedge: |
| 87 | ; CHECK: check_2.thread: |
| 88 | ; CHECK-NEXT: call void @bar() |
| 89 | ; CHECK-NEXT: br label %check_1 |
| 90 | |
| 91 | check_2: |
| 92 | %cond2 = icmp eq i32 %v, 2 |
| 93 | br i1 %cond2, label %eq_2, label %check_3 |
| 94 | ; No metadata: |
| 95 | ; CHECK: br i1 %cond2, label %eq_2, label %check_3{{$}} |
| 96 | |
| 97 | eq_2: |
| 98 | call void @bar() |
| 99 | br label %check_3 |
| 100 | ; Verify the new backedge: |
| 101 | ; CHECK: eq_2: |
| 102 | ; CHECK-NEXT: call void @bar() |
| 103 | ; CHECK-NEXT: br label %check_1 |
| 104 | |
| 105 | check_3: |
| 106 | %condE = icmp eq i32 %v, 3 |
| 107 | br i1 %condE, label %exit, label %check_1 |
| 108 | ; No metadata: |
| 109 | ; CHECK: br i1 %condE, label %exit, label %check_1{{$}} |
| 110 | |
| 111 | exit: |
| 112 | ret void |
| 113 | } |
| 114 | |
| 115 | !0 = !{!"function_entry_count", i64 120} |
| 116 | ; CHECK-NOT: branch_weights |
| 117 | !1 = !{!"branch_weights", i32 119, i32 1} |
| 118 | ; CHECK: !1 = !{!"branch_weights", i32 119, i32 1} |
| 119 | ; CHECK-NOT: branch_weights |