| Evan Cheng | 342e316 | 2011-08-30 01:34:54 +0000 | [diff] [blame] | 1 | ; RUN: llc %s -mtriple=thumbv7-apple-darwin -mcpu=cortex-a8 -o - |
| 2 | |
| 3 | ; When a i64 sub is expanded to subc + sube. |
| 4 | ; libcall #1 |
| 5 | ; \ |
| 6 | ; \ subc |
| 7 | ; \ / \ |
| 8 | ; \ / \ |
| 9 | ; \ / libcall #2 |
| 10 | ; sube |
| 11 | ; |
| 12 | ; If the libcalls are not serialized (i.e. both have chains which are dag |
| 13 | ; entry), legalizer can serialize them in arbitrary orders. If it's |
| 14 | ; unlucky, it can force libcall #2 before libcall #1 in the above case. |
| 15 | ; |
| 16 | ; subc |
| 17 | ; | |
| 18 | ; libcall #2 |
| 19 | ; | |
| 20 | ; libcall #1 |
| 21 | ; | |
| 22 | ; sube |
| 23 | ; |
| 24 | ; However since subc and sube are "glued" together, this ends up being a |
| 25 | ; cycle when the scheduler combine subc and sube as a single scheduling |
| 26 | ; unit. |
| 27 | ; |
| 28 | ; The right solution is to fix LegalizeType too chains the libcalls together. |
| 29 | ; However, LegalizeType is not processing nodes in order. The fix now is to |
| 30 | ; fix subc / sube (and addc / adde) to use physical register dependency instead. |
| 31 | ; rdar://10019576 |
| 32 | |
| 33 | define void @t() nounwind { |
| 34 | entry: |
| 35 | %tmp = load i64* undef, align 4 |
| 36 | %tmp5 = udiv i64 %tmp, 30 |
| 37 | %tmp13 = and i64 %tmp5, 64739244643450880 |
| 38 | %tmp16 = sub i64 0, %tmp13 |
| 39 | %tmp19 = and i64 %tmp16, 63 |
| 40 | %tmp20 = urem i64 %tmp19, 3 |
| 41 | %tmp22 = and i64 %tmp16, -272346829004752 |
| 42 | store i64 %tmp22, i64* undef, align 4 |
| 43 | store i64 %tmp20, i64* undef, align 4 |
| 44 | ret void |
| 45 | } |