| |
| # RUN: llc -O0 -mattr=-fullfp16 -mtriple=aarch64-- \ |
| # RUN: -run-pass=instruction-select -verify-machineinstrs %s -o - | FileCheck %s |
| |
| --- | |
| define <2 x double> @test_v2s64_unmerge(<2 x double> %a) { |
| ret <2 x double> %a |
| } |
| |
| define <4 x float> @test_v4s32_unmerge(<4 x float> %a) { |
| ret <4 x float> %a |
| } |
| |
| define <4 x half> @test_v4s16_unmerge(<4 x half> %a) { |
| ret <4 x half> %a |
| } |
| |
| define <8 x half> @test_v8s16_unmerge(<8 x half> %a) { |
| ret <8 x half> %a |
| } |
| |
| ... |
| --- |
| name: test_v2s64_unmerge |
| alignment: 2 |
| legalized: true |
| regBankSelected: true |
| tracksRegLiveness: true |
| registers: |
| - { id: 0, class: fpr } |
| - { id: 1, class: fpr } |
| - { id: 2, class: fpr } |
| - { id: 3, class: fpr } |
| body: | |
| bb.1 (%ir-block.0): |
| liveins: $q0 |
| ; CHECK-LABEL: name: test_v2s64_unmerge |
| %0:fpr(<2 x s64>) = COPY $q0 |
| |
| ; Since 2 * 64 = 128, we can just directly copy. |
| ; CHECK: %2:fpr64 = COPY %0.dsub |
| ; CHECK: %3:fpr64 = CPYi64 %0, 1 |
| %2:fpr(s64), %3:fpr(s64) = G_UNMERGE_VALUES %0(<2 x s64>) |
| |
| %1:fpr(<2 x s64>) = G_BUILD_VECTOR %2(s64), %3(s64) |
| $q0 = COPY %1(<2 x s64>) |
| RET_ReallyLR implicit $q0 |
| ... |
| --- |
| name: test_v4s32_unmerge |
| alignment: 2 |
| legalized: true |
| regBankSelected: true |
| tracksRegLiveness: true |
| registers: |
| - { id: 0, class: fpr } |
| - { id: 1, class: fpr } |
| - { id: 2, class: fpr } |
| - { id: 3, class: fpr } |
| - { id: 4, class: fpr } |
| - { id: 5, class: fpr } |
| body: | |
| bb.1 (%ir-block.0): |
| liveins: $q0 |
| ; CHECK-LABEL: name: test_v4s32_unmerge |
| %0:fpr(<4 x s32>) = COPY $q0 |
| |
| ; Since 4 * 32 = 128, we can just directly copy. |
| ; CHECK: %2:fpr32 = COPY %0.ssub |
| ; CHECK: %3:fpr32 = CPYi32 %0, 1 |
| ; CHECK: %4:fpr32 = CPYi32 %0, 2 |
| ; CHECK: %5:fpr32 = CPYi32 %0, 3 |
| %2:fpr(s32), %3:fpr(s32), %4:fpr(s32), %5:fpr(s32) = G_UNMERGE_VALUES %0(<4 x s32>) |
| |
| %1:fpr(<4 x s32>) = G_BUILD_VECTOR %2(s32), %3(s32), %4(s32), %5(s32) |
| $q0 = COPY %1(<4 x s32>) |
| RET_ReallyLR implicit $q0 |
| ... |
| --- |
| name: test_v4s16_unmerge |
| alignment: 2 |
| legalized: true |
| regBankSelected: true |
| tracksRegLiveness: true |
| registers: |
| - { id: 0, class: fpr } |
| - { id: 1, class: fpr } |
| - { id: 2, class: fpr } |
| - { id: 3, class: fpr } |
| - { id: 4, class: fpr } |
| - { id: 5, class: fpr } |
| body: | |
| bb.1 (%ir-block.0): |
| liveins: $d0 |
| ; CHECK-LABEL: name: test_v4s16_unmerge |
| %0:fpr(<4 x s16>) = COPY $d0 |
| |
| ; Since 4 * 16 != 128, we need to widen using implicit defs. |
| ; Note that we expect to reuse one of the INSERT_SUBREG results, as CPYi16 |
| ; expects a lane > 0. |
| ; CHECK-DAG: [[IMPDEF1:%[0-9]+]]:fpr128 = IMPLICIT_DEF |
| ; CHECK-NEXT: [[INS_SHARED:%[0-9]+]]:fpr128 = INSERT_SUBREG [[IMPDEF1]], %0, %subreg.dsub |
| ; CHECK: [[IMPDEF2:%[0-9]+]]:fpr128 = IMPLICIT_DEF |
| ; CHECK-NEXT: [[INS2:%[0-9]+]]:fpr128 = INSERT_SUBREG [[IMPDEF2]], %0, %subreg.dsub |
| ; CHECK: [[IMPDEF3:%[0-9]+]]:fpr128 = IMPLICIT_DEF |
| ; CHECK-NEXT: [[INS3:%[0-9]+]]:fpr128 = INSERT_SUBREG [[IMPDEF3]], %0, %subreg.dsub |
| ; CHECK: %2:fpr16 = COPY [[INS_SHARED]].hsub |
| ; CHECK: %3:fpr16 = CPYi16 [[INS_SHARED]], 1 |
| ; CHECK: %4:fpr16 = CPYi16 [[INS2]], 2 |
| ; CHECK: %5:fpr16 = CPYi16 [[INS3]], 3 |
| %2:fpr(s16), %3:fpr(s16), %4:fpr(s16), %5:fpr(s16) = G_UNMERGE_VALUES %0(<4 x s16>) |
| |
| %1:fpr(<4 x s16>) = G_BUILD_VECTOR %2(s16), %3(s16), %4(s16), %5(s16) |
| $d0 = COPY %1(<4 x s16>) |
| RET_ReallyLR implicit $d0 |
| ... |
| --- |
| name: test_v8s16_unmerge |
| alignment: 2 |
| legalized: true |
| regBankSelected: true |
| tracksRegLiveness: true |
| registers: |
| - { id: 0, class: fpr } |
| - { id: 1, class: fpr } |
| - { id: 2, class: fpr } |
| - { id: 3, class: fpr } |
| - { id: 4, class: fpr } |
| - { id: 5, class: fpr } |
| - { id: 6, class: fpr } |
| - { id: 7, class: fpr } |
| - { id: 8, class: fpr } |
| - { id: 9, class: fpr } |
| body: | |
| bb.1 (%ir-block.0): |
| liveins: $q0 |
| ; CHECK-LABEL: name: test_v8s16_unmerge |
| %0:fpr(<8 x s16>) = COPY $q0 |
| |
| ; Since 8 * 16 = 128, we can just directly copy. |
| ; CHECK: %2:fpr16 = COPY %0.hsub |
| ; CHECK: %3:fpr16 = CPYi16 %0, 1 |
| ; CHECK: %4:fpr16 = CPYi16 %0, 2 |
| ; CHECK: %5:fpr16 = CPYi16 %0, 3 |
| ; CHECK: %6:fpr16 = CPYi16 %0, 4 |
| ; CHECK: %7:fpr16 = CPYi16 %0, 5 |
| ; CHECK: %8:fpr16 = CPYi16 %0, 6 |
| ; CHECK: %9:fpr16 = CPYi16 %0, 7 |
| %2:fpr(s16), %3:fpr(s16), %4:fpr(s16), %5:fpr(s16), %6:fpr(s16), %7:fpr(s16), %8:fpr(s16), %9:fpr(s16) = G_UNMERGE_VALUES %0(<8 x s16>) |
| |
| %1:fpr(<8 x s16>) = G_BUILD_VECTOR %2:fpr(s16), %3:fpr(s16), %4:fpr(s16), %5:fpr(s16), %6:fpr(s16), %7:fpr(s16), %8:fpr(s16), %9:fpr(s16) |
| $q0 = COPY %1(<8 x s16>) |
| RET_ReallyLR implicit $q0 |
| ... |