| Justin Lebar | 6827de1 | 2016-03-14 23:15:34 +0000 | [diff] [blame] | 1 | ; RUN: opt < %s -loop-unroll -unroll-runtime -unroll-allow-partial -S | FileCheck %s |
| 2 | |
| 3 | declare void @f() convergent |
| 4 | |
| 5 | ; Although this loop contains a convergent instruction, it should be |
| 6 | ; fully unrolled. |
| 7 | ; |
| 8 | ; CHECK-LABEL: @full_unroll( |
| 9 | define i32 @full_unroll() { |
| 10 | entry: |
| 11 | br label %l3 |
| 12 | |
| 13 | l3: |
| 14 | %x.0 = phi i32 [ 0, %entry ], [ %inc, %l3 ] |
| 15 | ; CHECK: call void @f() |
| 16 | ; CHECK: call void @f() |
| 17 | ; CHECK: call void @f() |
| 18 | ; CHECK-NOT: call void @f() |
| 19 | call void @f() ;convergent |
| 20 | %inc = add nsw i32 %x.0, 1 |
| 21 | %exitcond = icmp eq i32 %inc, 3 |
| 22 | br i1 %exitcond, label %exit, label %l3 |
| 23 | |
| 24 | exit: |
| 25 | ret i32 0 |
| 26 | } |
| 27 | |
| 28 | ; This loop contains a convergent instruction, but it should be partially |
| 29 | ; unrolled. The unroll count is the largest power of 2 that divides the |
| 30 | ; multiple -- 4, in this case. |
| 31 | ; |
| 32 | ; CHECK-LABEL: @runtime_unroll( |
| 33 | define i32 @runtime_unroll(i32 %n) { |
| 34 | entry: |
| 35 | %loop_ctl = mul nsw i32 %n, 12 |
| 36 | br label %l3 |
| 37 | |
| 38 | l3: |
| 39 | %x.0 = phi i32 [ 0, %entry ], [ %inc, %l3 ] |
| 40 | ; CHECK: call void @f() |
| 41 | ; CHECK: call void @f() |
| 42 | ; CHECK: call void @f() |
| 43 | ; CHECK: call void @f() |
| 44 | ; CHECK-NOT: call void @f() |
| 45 | call void @f() convergent |
| 46 | %inc = add nsw i32 %x.0, 1 |
| 47 | %exitcond = icmp eq i32 %inc, %loop_ctl |
| 48 | br i1 %exitcond, label %exit, label %l3 |
| 49 | |
| 50 | exit: |
| 51 | ret i32 0 |
| 52 | } |
| 53 | |
| 54 | ; This loop contains a convergent instruction, so its partial unroll |
| 55 | ; count must divide its trip multiple. This overrides its unroll |
| 56 | ; pragma -- we unroll exactly 8 times, even though 16 is requested. |
| 57 | ; CHECK-LABEL: @pragma_unroll |
| 58 | define i32 @pragma_unroll(i32 %n) { |
| 59 | entry: |
| 60 | %loop_ctl = mul nsw i32 %n, 24 |
| 61 | br label %l3, !llvm.loop !0 |
| 62 | |
| 63 | l3: |
| 64 | %x.0 = phi i32 [ 0, %entry ], [ %inc, %l3 ] |
| 65 | ; CHECK: call void @f() |
| 66 | ; CHECK: call void @f() |
| 67 | ; CHECK: call void @f() |
| 68 | ; CHECK: call void @f() |
| 69 | ; CHECK: call void @f() |
| 70 | ; CHECK: call void @f() |
| 71 | ; CHECK: call void @f() |
| 72 | ; CHECK: call void @f() |
| 73 | ; CHECK-NOT: call void @f() |
| 74 | call void @f() convergent |
| 75 | %inc = add nsw i32 %x.0, 1 |
| 76 | %exitcond = icmp eq i32 %inc, %loop_ctl |
| 77 | br i1 %exitcond, label %exit, label %l3, !llvm.loop !0 |
| 78 | |
| 79 | exit: |
| 80 | ret i32 0 |
| 81 | } |
| 82 | |
| 83 | !0 = !{!0, !{!"llvm.loop.unroll.count", i32 16}} |