maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity
llvm-project/llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-load-local.mir

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

791 lines
30 KiB
Plaintext
Raw Normal View History

AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -march=amdgcn -mcpu=tahiti -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX6 %s
# RUN: llc -march=amdgcn -mcpu=hawaii -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX7 %s
# RUN: llc -march=amdgcn -mcpu=fiji -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX7 %s
# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX9 %s
# RUN: llc -march=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=0 -o - %s | FileCheck -check-prefix=GFX9 %s
---
name: load_local_s32_from_4
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_4
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s32), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s32_from_4
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s32), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_B32_]]
; GFX9-LABEL: name: load_local_s32_from_4
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_B32_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (s32), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_B32_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(s32) = G_LOAD %0 :: (load (s32), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %1
...
---
name: load_local_s32_from_2
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_2
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_U16_:%[0-9]+]]:vgpr_32 = DS_READ_U16 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s16), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_U16_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s32_from_2
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_U16_:%[0-9]+]]:vgpr_32 = DS_READ_U16 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s16), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_U16_]]
; GFX9-LABEL: name: load_local_s32_from_2
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_U16_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_U16_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (s16), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_U16_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(s32) = G_LOAD %0 :: (load (s16), align 2, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %1
...
---
name: load_local_s32_from_1
legalized: true
regBankSelected: true
tracksRegLiveness: true
machineFunctionInfo:
scratchRSrcReg: $sgpr0_sgpr1_sgpr2_sgpr3
stackPtrOffsetReg: $sgpr32
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_1
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_U8_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s32_from_1
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_U8_]]
; GFX9-LABEL: name: load_local_s32_from_1
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_U8_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_U8_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_U8_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(s32) = G_LOAD %0 :: (load (s8), align 1, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %1
...
---
name: load_local_v2s32
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_v2s32
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (<2 x s32>), addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_v2s32
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (<2 x s32>), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
; GFX9-LABEL: name: load_local_v2s32
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_READ_B64_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (<2 x s32>), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ_B64_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(<2 x s32>) = G_LOAD %0 :: (load (<2 x s32>), align 8, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_v2s32_align4
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_v2s32_align4
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(<2 x s32>) = G_LOAD [[COPY]](p3) :: (load (<2 x s32>), align 4, addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x s32>)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_v2s32_align4
; GFX7: liveins: $vgpr0
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ2_B32_:%[0-9]+]]:vreg_64 = DS_READ2_B32 [[COPY]], 0, 1, 0, implicit $m0, implicit $exec :: (load (<2 x s32>), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_v2s32_align4
; GFX9: liveins: $vgpr0
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ2_B32_gfx9_:%[0-9]+]]:vreg_64 = DS_READ2_B32_gfx9 [[COPY]], 0, 1, 0, implicit $exec :: (load (<2 x s32>), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(<2 x s32>) = G_LOAD %0 :: (load (<2 x s32>), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_s64
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s64
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s64), addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s64
; GFX7: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (s64), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_s64
; GFX9: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_READ_B64_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (s64), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ_B64_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(s64) = G_LOAD %0 :: (load (s64), align 8, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_s64_align4
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s64_align4
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(s64) = G_LOAD [[COPY]](p3) :: (load (s64), align 4, addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s64_align4
; GFX7: liveins: $vgpr0
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ2_B32_:%[0-9]+]]:vreg_64 = DS_READ2_B32 [[COPY]], 0, 1, 0, implicit $m0, implicit $exec :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_s64_align4
; GFX9: liveins: $vgpr0
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ2_B32_gfx9_:%[0-9]+]]:vreg_64 = DS_READ2_B32_gfx9 [[COPY]], 0, 1, 0, implicit $exec :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(s64) = G_LOAD %0 :: (load (s64), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_p3_from_4
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_p3_from_4
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (p3), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_p3_from_4
; GFX7: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (p3), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_p3_from_4
; GFX9: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_B32_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (p3), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_B32_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(p3) = G_LOAD %0 :: (load (p3), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %1
...
---
name: load_local_p5_from_4
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_p5_from_4
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (p5), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_p5_from_4
; GFX7: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (p5), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_p5_from_4
; GFX9: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_B32_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (p5), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_B32_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(p5) = G_LOAD %0 :: (load (p5), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %1
...
---
name: load_local_p1_align8
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_p1_align8
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (p1), addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_p1_align8
; GFX7: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (p1), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_p1_align8
; GFX9: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_READ_B64_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (p1), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ_B64_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(p1) = G_LOAD %0 :: (load (p1), align 8, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_p1_align4
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_p1_align4
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(p1) = G_LOAD [[COPY]](p3) :: (load (p1), align 4, addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](p1)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_p1_align4
; GFX7: liveins: $vgpr0
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ2_B32_:%[0-9]+]]:vreg_64 = DS_READ2_B32 [[COPY]], 0, 1, 0, implicit $m0, implicit $exec :: (load (p1), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_p1_align4
; GFX9: liveins: $vgpr0
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ2_B32_gfx9_:%[0-9]+]]:vreg_64 = DS_READ2_B32_gfx9 [[COPY]], 0, 1, 0, implicit $exec :: (load (p1), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(p1) = G_LOAD %0 :: (load (p1), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_p999_from_8
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_p999_from_8
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(p999) = G_LOAD [[COPY]](p3) :: (load (p999), addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](p999)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_p999_from_8
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[LOAD:%[0-9]+]]:vreg_64(p999) = G_LOAD [[COPY]](p3) :: (load (p999), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[LOAD]](p999)
; GFX9-LABEL: name: load_local_p999_from_8
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[LOAD:%[0-9]+]]:vreg_64(p999) = G_LOAD [[COPY]](p3) :: (load (p999), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](p999)
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(p999) = G_LOAD %0 :: (load (p999), align 8, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_v2p3
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_v2p3
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(<2 x p3>) = G_LOAD [[COPY]](p3) :: (load (<2 x p3>), addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_v2p3
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[LOAD:%[0-9]+]]:vreg_64(<2 x p3>) = G_LOAD [[COPY]](p3) :: (load (<2 x p3>), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
; GFX9-LABEL: name: load_local_v2p3
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[LOAD:%[0-9]+]]:vreg_64(<2 x p3>) = G_LOAD [[COPY]](p3) :: (load (<2 x p3>), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[LOAD]](<2 x p3>)
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(<2 x p3>) = G_LOAD %0 :: (load (<2 x p3>), align 8, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
---
name: load_local_v2s16
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_v2s16
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (<2 x s16>), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_v2s16
; GFX7: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B32_:%[0-9]+]]:vgpr_32 = DS_READ_B32 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (<2 x s16>), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_B32_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_v2s16
; GFX9: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B32_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_B32_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (<2 x s16>), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_B32_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(<2 x s16>) = G_LOAD %0 :: (load (<2 x s16>), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %1
...
---
name: load_local_v4s16
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_v4s16
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (<4 x s16>), addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_v4s16
; GFX7: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_B64_:%[0-9]+]]:vreg_64 = DS_READ_B64 [[COPY]], 0, 0, implicit $m0, implicit $exec :: (load (<4 x s16>), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ_B64_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9-LABEL: name: load_local_v4s16
; GFX9: liveins: $vgpr0
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_B64_gfx9_:%[0-9]+]]:vreg_64 = DS_READ_B64_gfx9 [[COPY]], 0, 0, implicit $exec :: (load (<4 x s16>), addrspace 3)
AMDGPU/GlobalISel: Fix load/store of types in other address spaces There should probably be a size only matcher. llvm-svn: 371155
2019-09-06 08:36:06 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ_B64_gfx9_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
%0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%1:vgpr(<4 x s16>) = G_LOAD %0 :: (load (<4 x s16>), align 8, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0_vgpr1 = COPY %1
...
# ---
# name: load_local_v6s16
# legalized: true
# regBankSelected: true
# tracksRegLiveness: true
# machineFunctionInfo:
# scratchRSrcReg: $sgpr0_sgpr1_sgpr2_sgpr3
# stackPtrOffsetReg: $sgpr32
# body: |
# bb.0:
# liveins: $vgpr0
# %0:vgpr(p3) = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
# %1:vgpr(<6 x s16>) = G_LOAD %0 :: (load (<6 x s16>), align 4, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
# $vgpr0_vgpr1_vgpr2 = COPY %1
# ...
################################################################################
### Stress addressing modes
################################################################################
---
name: load_local_s32_from_1_gep_65535
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_1_gep_65535
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65535, implicit $exec
; GFX6: %2:vgpr_32, dead %4:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 %2, 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_U8_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s32_from_1_gep_65535
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 [[COPY]], 65535, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_U8_]]
; GFX9-LABEL: name: load_local_s32_from_1_gep_65535
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_U8_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_U8_gfx9 [[COPY]], 65535, 0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_U8_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
%1:vgpr(s32) = G_CONSTANT i32 65535
[globalisel] Rename G_GEP to G_PTR_ADD Summary: G_GEP is rather poorly named. It's a simple pointer+scalar addition and doesn't support any of the complexities of getelementptr. I therefore propose that we rename it. There's a G_PTR_MASK so let's follow that convention and go with G_PTR_ADD Reviewers: volkan, aditya_nandakumar, bogner, rovka, arsenm Subscribers: sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, arphaman, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D69734
2019-11-02 04:18:00 +08:00
%2:vgpr(p3) = G_PTR_ADD %0, %1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%3:vgpr(s32) = G_LOAD %2 :: (load (s8), align 1, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %3
...
---
AMDGPU/GlobalISel: Use known bits for selection llvm-svn: 371409
2019-09-09 23:39:32 +08:00
name: load_local_s32_from_1_gep_65535_known_bits_base_address
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_1_gep_65535_known_bits_base_address
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 2147483647, implicit $exec
; GFX6: [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY]], [[V_MOV_B32_e32_]], implicit $exec
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 [[V_AND_B32_e64_]], 65535, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_U8_]]
AMDGPU/GlobalISel: Use known bits for selection llvm-svn: 371409
2019-09-09 23:39:32 +08:00
; GFX7-LABEL: name: load_local_s32_from_1_gep_65535_known_bits_base_address
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 2147483647, implicit $exec
; GFX7: [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY]], [[V_MOV_B32_e32_]], implicit $exec
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 [[V_AND_B32_e64_]], 65535, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Use known bits for selection llvm-svn: 371409
2019-09-09 23:39:32 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_U8_]]
; GFX9-LABEL: name: load_local_s32_from_1_gep_65535_known_bits_base_address
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX9: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 2147483647, implicit $exec
; GFX9: [[V_AND_B32_e64_:%[0-9]+]]:vgpr_32 = V_AND_B32_e64 [[COPY]], [[V_MOV_B32_e32_]], implicit $exec
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_U8_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_U8_gfx9 [[V_AND_B32_e64_]], 65535, 0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Use known bits for selection llvm-svn: 371409
2019-09-09 23:39:32 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_U8_gfx9_]]
%0:vgpr(s32) = COPY $vgpr0
%1:vgpr(s32) = G_CONSTANT i32 2147483647
%2:vgpr(s32) = G_AND %0, %1
%3:vgpr(p3) = G_INTTOPTR %2
%4:vgpr(s32) = G_CONSTANT i32 65535
[globalisel] Rename G_GEP to G_PTR_ADD Summary: G_GEP is rather poorly named. It's a simple pointer+scalar addition and doesn't support any of the complexities of getelementptr. I therefore propose that we rename it. There's a G_PTR_MASK so let's follow that convention and go with G_PTR_ADD Reviewers: volkan, aditya_nandakumar, bogner, rovka, arsenm Subscribers: sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, arphaman, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D69734
2019-11-02 04:18:00 +08:00
%5:vgpr(p3) = G_PTR_ADD %3, %4
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%6:vgpr(s32) = G_LOAD %5 :: (load (s8), align 1, addrspace 3)
AMDGPU/GlobalISel: Use known bits for selection llvm-svn: 371409
2019-09-09 23:39:32 +08:00
$vgpr0 = COPY %6
...
---
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
name: load_local_s32_from_1_gep_65536
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_1_gep_65536
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65536, implicit $exec
; GFX6: %2:vgpr_32, dead %4:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 %2, 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_U8_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s32_from_1_gep_65536
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65536, implicit $exec
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; GFX7: %2:vgpr_32, dead %4:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 %2, 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_U8_]]
; GFX9-LABEL: name: load_local_s32_from_1_gep_65536
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX9: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 65536, implicit $exec
; GFX9: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_U8_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_U8_gfx9 [[V_ADD_U32_e64_]], 0, 0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_U8_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
%1:vgpr(s32) = G_CONSTANT i32 65536
[globalisel] Rename G_GEP to G_PTR_ADD Summary: G_GEP is rather poorly named. It's a simple pointer+scalar addition and doesn't support any of the complexities of getelementptr. I therefore propose that we rename it. There's a G_PTR_MASK so let's follow that convention and go with G_PTR_ADD Reviewers: volkan, aditya_nandakumar, bogner, rovka, arsenm Subscribers: sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, arphaman, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D69734
2019-11-02 04:18:00 +08:00
%2:vgpr(p3) = G_PTR_ADD %0, %1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%3:vgpr(s32) = G_LOAD %2 :: (load (s8), align 1, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %3
...
---
name: load_local_s32_from_1_gep_m1
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s32_from_1_gep_m1
; GFX6: liveins: $vgpr0
; GFX6: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX6: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
; GFX6: %2:vgpr_32, dead %4:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 %2, 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
; GFX6: $vgpr0 = COPY [[DS_READ_U8_]]
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7-LABEL: name: load_local_s32_from_1_gep_m1
; GFX7: liveins: $vgpr0
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Sign extend integer constants This matches the DAG behavior and fixes immediate folding
2020-07-26 02:37:29 +08:00
; GFX7: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; GFX7: %2:vgpr_32, dead %4:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ_U8_:%[0-9]+]]:vgpr_32 = DS_READ_U8 %2, 0, 0, implicit $m0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX7: $vgpr0 = COPY [[DS_READ_U8_]]
; GFX9-LABEL: name: load_local_s32_from_1_gep_m1
; GFX9: liveins: $vgpr0
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
AMDGPU/GlobalISel: Sign extend integer constants This matches the DAG behavior and fixes immediate folding
2020-07-26 02:37:29 +08:00
; GFX9: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 -1, implicit $exec
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ_U8_gfx9_:%[0-9]+]]:vgpr_32 = DS_READ_U8_gfx9 [[V_ADD_U32_e64_]], 0, 0, implicit $exec :: (load (s8), addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
; GFX9: $vgpr0 = COPY [[DS_READ_U8_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
%1:vgpr(s32) = G_CONSTANT i32 -1
[globalisel] Rename G_GEP to G_PTR_ADD Summary: G_GEP is rather poorly named. It's a simple pointer+scalar addition and doesn't support any of the complexities of getelementptr. I therefore propose that we rename it. There's a G_PTR_MASK so let's follow that convention and go with G_PTR_ADD Reviewers: volkan, aditya_nandakumar, bogner, rovka, arsenm Subscribers: sdardis, jvesely, wdng, nhaehnle, hiraditya, jrtc27, atanasyan, arphaman, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D69734
2019-11-02 04:18:00 +08:00
%2:vgpr(p3) = G_PTR_ADD %0, %1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%3:vgpr(s32) = G_LOAD %2 :: (load (s8), align 1, addrspace 3)
AMDGPU/GlobalISel: Select local loads llvm-svn: 367498
2019-08-01 08:53:38 +08:00
$vgpr0 = COPY %3
...
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
---
name: load_local_s64_align4_from_1_gep_1016
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s64_align4_from_1_gep_1016
; GFX6: liveins: $vgpr0_vgpr1
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: [[C:%[0-9]+]]:vgpr(s32) = G_CONSTANT i32 1016
; GFX6: [[PTR_ADD:%[0-9]+]]:vgpr(p3) = G_PTR_ADD [[COPY]], [[C]](s32)
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(s64) = G_LOAD [[PTR_ADD]](p3) :: (load (s64), align 4, addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7-LABEL: name: load_local_s64_align4_from_1_gep_1016
; GFX7: liveins: $vgpr0_vgpr1
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ2_B32_:%[0-9]+]]:vreg_64 = DS_READ2_B32 [[COPY]], 254, 255, 0, implicit $m0, implicit $exec :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_]]
; GFX9-LABEL: name: load_local_s64_align4_from_1_gep_1016
; GFX9: liveins: $vgpr0_vgpr1
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ2_B32_gfx9_:%[0-9]+]]:vreg_64 = DS_READ2_B32_gfx9 [[COPY]], 254, 255, 0, implicit $exec :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
%1:vgpr(s32) = G_CONSTANT i32 1016
%2:vgpr(p3) = G_PTR_ADD %0, %1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%3:vgpr(s64) = G_LOAD %2 :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
$vgpr0_vgpr1 = COPY %3
...
---
name: load_local_s64_align4_from_1_gep_1020
legalized: true
regBankSelected: true
tracksRegLiveness: true
body: |
bb.0:
liveins: $vgpr0_vgpr1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX6-LABEL: name: load_local_s64_align4_from_1_gep_1020
; GFX6: liveins: $vgpr0_vgpr1
; GFX6: [[COPY:%[0-9]+]]:vgpr(p3) = COPY $vgpr0
; GFX6: [[C:%[0-9]+]]:vgpr(s32) = G_CONSTANT i32 1020
; GFX6: [[PTR_ADD:%[0-9]+]]:vgpr(p3) = G_PTR_ADD [[COPY]], [[C]](s32)
; GFX6: $m0 = S_MOV_B32 -1
; GFX6: [[LOAD:%[0-9]+]]:vreg_64(s64) = G_LOAD [[PTR_ADD]](p3) :: (load (s64), align 4, addrspace 3)
; GFX6: $vgpr0_vgpr1 = COPY [[LOAD]](s64)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7-LABEL: name: load_local_s64_align4_from_1_gep_1020
; GFX7: liveins: $vgpr0_vgpr1
; GFX7: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX7: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1020, implicit $exec
AMDGPU: Rename add/sub with carry out instructions The hardware has created a real mess in the naming for add/sub, which have been renamed basically every generation. Switch the carry out pseudos to have the gfx9/gfx10 names. We were using the original SI/CI v_add_i32/v_sub_i32 names. Later targets reintroduced these names as carryless instructions with a saturating clamp bit, which we do not define. Do this rename so we can unambiguously add these missing instructions. The carry-in versions should also be renamed, but at least those had a consistent _u32 name to begin with. The 16-bit instructions were also renamed, but aren't ambiguous. This does regress assembler error message quality in some cases. In mismatched wave32/wave64 situations, this will switch from "unsupported instruction" to "invalid operand", with the error pointing at the wrong position. I couldn't quite follow how the assembler selects these, but the previous behavior seemed accidental to me. It looked like there was a partial attempt to handle this which was never completed (i.e. there is an AMDGPUOperand::isBoolReg but it isn't used for anything).
2020-07-14 21:18:36 +08:00
; GFX7: %2:vgpr_32, dead %4:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: $m0 = S_MOV_B32 -1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX7: [[DS_READ2_B32_:%[0-9]+]]:vreg_64 = DS_READ2_B32 %2, 0, 1, 0, implicit $m0, implicit $exec :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX7: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_]]
; GFX9-LABEL: name: load_local_s64_align4_from_1_gep_1020
; GFX9: liveins: $vgpr0_vgpr1
; GFX9: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
; GFX9: [[V_MOV_B32_e32_:%[0-9]+]]:vgpr_32 = V_MOV_B32_e32 1020, implicit $exec
; GFX9: [[V_ADD_U32_e64_:%[0-9]+]]:vgpr_32 = V_ADD_U32_e64 [[COPY]], [[V_MOV_B32_e32_]], 0, implicit $exec
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
; GFX9: [[DS_READ2_B32_gfx9_:%[0-9]+]]:vreg_64 = DS_READ2_B32_gfx9 [[V_ADD_U32_e64_]], 0, 1, 0, implicit $exec :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
; GFX9: $vgpr0_vgpr1 = COPY [[DS_READ2_B32_gfx9_]]
%0:vgpr(p3) = COPY $vgpr0
%1:vgpr(s32) = G_CONSTANT i32 1020
%2:vgpr(p3) = G_PTR_ADD %0, %1
CodeGen: Print/parse LLTs in MachineMemOperands This will currently accept the old number of bytes syntax, and convert it to a scalar. This should be removed in the near future (I think I converted all of the tests already, but likely missed a few). Not sure what the exact syntax and policy should be. We can continue printing the number of bytes for non-generic instructions to avoid test churn and only allow non-scalar types for generic instructions. This will currently print the LLT in parentheses, but accept parsing the existing integers and implicitly converting to scalar. The parentheses are a bit ugly, but the parser logic seems unable to deal without either parentheses or some keyword to indicate the start of a type.
2021-05-20 10:25:51 +08:00
%3:vgpr(s64) = G_LOAD %2 :: (load (s64), align 4, addrspace 3)
[AMDGPU][GlobalISel] Select 8-byte LDS Ops with 4-byte alignment Reviewers: arsenm Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits Tags: #llvm Differential Revision: https://reviews.llvm.org/D73585
2020-01-08 05:13:05 +08:00
$vgpr0_vgpr1 = COPY %3
...