| Nicolai Haehnle | dd059c1 | 2017-11-22 12:25:21 +0000 | [diff] [blame] | 1 | # RUN: llc -march=amdgcn -verify-machineinstrs -run-pass si-load-store-opt -o - %s | FileCheck %s |
| 2 | |
| 3 | # Check that SILoadStoreOptimizer honors memory dependencies between moved |
| 4 | # instructions. |
| 5 | # |
| 6 | # The following IR snippet would usually be optimized by the peephole optimizer. |
| 7 | # However, an equivalent situation can occur with buffer instructions as well. |
| 8 | |
| 9 | # CHECK-LABEL: name: mem_dependency |
| 10 | # CHECK: DS_READ2_B32 %0, 0, 1, |
| 11 | # CHECK: DS_WRITE_B32 %0, killed %1, 64, |
| 12 | # CHECK: DS_READ2_B32 %0, 16, 17, |
| 13 | # CHECK: DS_WRITE_B32 killed %0, %5, 0 |
| 14 | |
| 15 | --- | |
| 16 | define amdgpu_kernel void @mem_dependency(i32 addrspace(3)* %ptr.0) nounwind { |
| 17 | %ptr.4 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 1 |
| 18 | %ptr.64 = getelementptr i32, i32 addrspace(3)* %ptr.0, i32 16 |
| 19 | %1 = load i32, i32 addrspace(3)* %ptr.0 |
| 20 | store i32 %1, i32 addrspace(3)* %ptr.64 |
| 21 | %2 = load i32, i32 addrspace(3)* %ptr.64 |
| 22 | %3 = load i32, i32 addrspace(3)* %ptr.4 |
| 23 | %4 = add i32 %2, %3 |
| 24 | store i32 %4, i32 addrspace(3)* %ptr.0 |
| 25 | ret void |
| 26 | } |
| 27 | ... |
| 28 | --- |
| 29 | name: mem_dependency |
| 30 | alignment: 0 |
| 31 | exposesReturnsTwice: false |
| 32 | legalized: false |
| 33 | regBankSelected: false |
| 34 | selected: false |
| 35 | tracksRegLiveness: true |
| 36 | liveins: |
| 37 | - { reg: '%vgpr0', virtual-reg: '%1' } |
| 38 | frameInfo: |
| 39 | isFrameAddressTaken: false |
| 40 | isReturnAddressTaken: false |
| 41 | hasStackMap: false |
| 42 | hasPatchPoint: false |
| 43 | stackSize: 0 |
| 44 | offsetAdjustment: 0 |
| 45 | maxAlignment: 0 |
| 46 | adjustsStack: false |
| 47 | hasCalls: false |
| 48 | maxCallFrameSize: 0 |
| 49 | hasOpaqueSPAdjustment: false |
| 50 | hasVAStart: false |
| 51 | hasMustTailInVarArgFunc: false |
| 52 | body: | |
| 53 | bb.0: |
| 54 | liveins: %vgpr0 |
| 55 | |
| 56 | %1:vgpr_32 = COPY %vgpr0 |
| 57 | %m0 = S_MOV_B32 -1 |
| 58 | %2:vgpr_32 = DS_READ_B32 %1, 0, 0, implicit %m0, implicit %exec :: (load 4 from %ir.ptr.0) |
| 59 | DS_WRITE_B32 %1, killed %2, 64, 0, implicit %m0, implicit %exec :: (store 4 into %ir.ptr.64) |
| 60 | |
| 61 | ; Make this load unmergeable, to tempt SILoadStoreOptimizer into merging the |
| 62 | ; other two loads. |
| 63 | %6:vreg_64 = DS_READ2_B32 %1, 16, 17, 0, implicit %m0, implicit %exec :: (load 8 from %ir.ptr.64, align 4) |
| 64 | %3:vgpr_32 = COPY %6.sub0 |
| 65 | %4:vgpr_32 = DS_READ_B32 %1, 4, 0, implicit %m0, implicit %exec :: (load 4 from %ir.ptr.4) |
| 66 | %5:vgpr_32 = V_ADD_I32_e32 killed %3, killed %4, implicit-def %vcc, implicit %exec |
| 67 | DS_WRITE_B32 killed %1, %5, 0, 0, implicit killed %m0, implicit %exec :: (store 4 into %ir.ptr.0) |
| 68 | S_ENDPGM |
| 69 | |
| 70 | ... |