710 lines
26 KiB
YAML
710 lines
26 KiB
YAML
# RUN: llc --mtriple=amdgcn--amdhsa -mcpu=fiji -verify-machineinstrs -run-pass si-fold-operands,si-shrink-instructions %s -o - | FileCheck %s
|
|
--- |
|
|
define amdgpu_kernel void @add_f32_1.0_one_f16_use() #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile float, float addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH3C00
|
|
%f32.add = fadd float %f32.val, 1.000000e+00
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile float %f32.add, float addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_f32_1.0_multi_f16_use() #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile float, float addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH3C00
|
|
%f32.add = fadd float %f32.val, 1.000000e+00
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile float %f32.add, float addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_f32_1.0_one_f32_use_one_f16_use () #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile float, float addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH3C00
|
|
%f32.add = fadd float %f32.val, 1.000000e+00
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile float %f32.add, float addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_f32_1.0_one_f32_use_multi_f16_use () #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile float, float addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH3C00
|
|
%f16.add1 = fadd half %f16.val1, 0xH3C00
|
|
%f32.add = fadd float %f32.val, 1.000000e+00
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile half %f16.add1, half addrspace(1)* undef
|
|
store volatile float %f32.add, float addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_i32_1_multi_f16_use() #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH0001
|
|
%f16.add1 = fadd half %f16.val1, 0xH0001
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile half %f16.add1,half addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_i32_m2_one_f32_use_multi_f16_use () #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile float, float addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xHFFFE
|
|
%f16.add1 = fadd half %f16.val1, 0xHFFFE
|
|
%f32.add = fadd float %f32.val, 0xffffffffc0000000
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile half %f16.add1, half addrspace(1)* undef
|
|
store volatile float %f32.add, float addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_f16_1.0_multi_f32_use() #0 {
|
|
%f32.val0 = load volatile float, float addrspace(1)* undef
|
|
%f32.val1 = load volatile float, float addrspace(1)* undef
|
|
%f32.val = load volatile float, float addrspace(1)* undef
|
|
%f32.add0 = fadd float %f32.val0, 1.0
|
|
%f32.add1 = fadd float %f32.val1, 1.0
|
|
store volatile float %f32.add0, float addrspace(1)* undef
|
|
store volatile float %f32.add1, float addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_f16_1.0_other_high_bits_multi_f16_use() #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile half, half addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH3C00
|
|
%f32.add = fadd half %f32.val, 1.000000e+00
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile half %f32.add, half addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
define amdgpu_kernel void @add_f16_1.0_other_high_bits_use_f16_f32() #0 {
|
|
%f16.val0 = load volatile half, half addrspace(1)* undef
|
|
%f16.val1 = load volatile half, half addrspace(1)* undef
|
|
%f32.val = load volatile half, half addrspace(1)* undef
|
|
%f16.add0 = fadd half %f16.val0, 0xH3C00
|
|
%f32.add = fadd half %f32.val, 1.000000e+00
|
|
store volatile half %f16.add0, half addrspace(1)* undef
|
|
store volatile half %f32.add, half addrspace(1)* undef
|
|
ret void
|
|
}
|
|
|
|
attributes #0 = { nounwind }
|
|
|
|
...
|
|
---
|
|
|
|
# f32 1.0 with a single use should be folded as the low 32-bits of a
|
|
# literal constant.
|
|
|
|
# CHECK-LABEL: name: add_f32_1.0_one_f16_use
|
|
# CHECK: %13:vgpr_32 = V_ADD_F16_e32 1065353216, killed %11, implicit $mode, implicit $exec
|
|
|
|
name: add_f32_1.0_one_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = V_MOV_B32_e32 1065353216, implicit $exec
|
|
%13 = V_ADD_F16_e64 0, killed %11, 0, %12, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %13, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
# Materialized f32 inline immediate should not be folded into the f16
|
|
# operands
|
|
|
|
# CHECK-LABEL: name: add_f32_1.0_multi_f16_use
|
|
# CHECK: %13:vgpr_32 = V_MOV_B32_e32 1065353216, implicit $exec
|
|
# CHECK: %14:vgpr_32 = V_ADD_F16_e32 killed %11, %13, implicit $mode, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 killed %12, killed %13, implicit $mode, implicit $exec
|
|
|
|
|
|
name: add_f32_1.0_multi_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%13 = V_MOV_B32_e32 1065353216, implicit $exec
|
|
%14 = V_ADD_F16_e64 0, killed %11, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, killed %12, 0, killed %13, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %14, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
|
|
# f32 1.0 should be folded into the single f32 use as an inline
|
|
# immediate, and folded into the single f16 use as a literal constant
|
|
|
|
# CHECK-LABEL: name: add_f32_1.0_one_f32_use_one_f16_use
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 1065353216, %11, implicit $mode, implicit $exec
|
|
# CHECK: %16:vgpr_32 = V_ADD_F32_e32 1065353216, killed %13, implicit $mode, implicit $exec
|
|
|
|
name: add_f32_1.0_one_f32_use_one_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
- { id: 16, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%13 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%14 = V_MOV_B32_e32 1065353216, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, %11, 0, %14, 0, 0, implicit $mode, implicit $exec
|
|
%16 = V_ADD_F32_e64 0, killed %13, 0, killed %14, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_DWORD_OFFSET killed %16, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `float addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
|
|
# f32 1.0 should be folded for the single f32 use as an inline
|
|
# constant, and not folded as a multi-use literal for the f16 cases
|
|
|
|
# CHECK-LABEL: name: add_f32_1.0_one_f32_use_multi_f16_use
|
|
# CHECK: %14:vgpr_32 = V_MOV_B32_e32 1065353216, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 %11, %14, implicit $mode, implicit $exec
|
|
# CHECK: %16:vgpr_32 = V_ADD_F16_e32 %12, %14, implicit $mode, implicit $exec
|
|
# CHECK: %17:vgpr_32 = V_ADD_F32_e32 1065353216, killed %13, implicit $mode, implicit $exec
|
|
|
|
name: add_f32_1.0_one_f32_use_multi_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
- { id: 16, class: vgpr_32 }
|
|
- { id: 17, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%13 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%14 = V_MOV_B32_e32 1065353216, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, %11, 0, %14, 0, 0, implicit $mode, implicit $exec
|
|
%16 = V_ADD_F16_e64 0, %12, 0, %14, 0, 0, implicit $mode, implicit $exec
|
|
%17 = V_ADD_F32_e64 0, killed %13, 0, killed %14, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_SHORT_OFFSET killed %16, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_DWORD_OFFSET killed %17, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `float addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
# CHECK-LABEL: name: add_i32_1_multi_f16_use
|
|
# CHECK: %13:vgpr_32 = V_MOV_B32_e32 1, implicit $exec
|
|
# CHECK: %14:vgpr_32 = V_ADD_F16_e32 1, killed %11, implicit $mode, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 1, killed %12, implicit $mode, implicit $exec
|
|
|
|
|
|
name: add_i32_1_multi_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%13 = V_MOV_B32_e32 1, implicit $exec
|
|
%14 = V_ADD_F16_e64 0, killed %11, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, killed %12, 0, killed %13, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %14, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
|
|
# CHECK-LABEL: name: add_i32_m2_one_f32_use_multi_f16_use
|
|
# CHECK: %14:vgpr_32 = V_MOV_B32_e32 -2, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 -2, %11, implicit $mode, implicit $exec
|
|
# CHECK: %16:vgpr_32 = V_ADD_F16_e32 -2, %12, implicit $mode, implicit $exec
|
|
# CHECK: %17:vgpr_32 = V_ADD_F32_e32 -2, killed %13, implicit $mode, implicit $exec
|
|
|
|
name: add_i32_m2_one_f32_use_multi_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
- { id: 16, class: vgpr_32 }
|
|
- { id: 17, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%13 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%14 = V_MOV_B32_e32 -2, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, %11, 0, %14, 0, 0, implicit $mode, implicit $exec
|
|
%16 = V_ADD_F16_e64 0, %12, 0, %14, 0, 0, implicit $mode, implicit $exec
|
|
%17 = V_ADD_F32_e64 0, killed %13, 0, killed %14, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_SHORT_OFFSET killed %16, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_DWORD_OFFSET killed %17, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `float addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
|
|
# f32 1.0 should be folded for the single f32 use as an inline
|
|
# constant, and not folded as a multi-use literal for the f16 cases
|
|
|
|
# CHECK-LABEL: name: add_f16_1.0_multi_f32_use
|
|
# CHECK: %13:vgpr_32 = V_MOV_B32_e32 15360, implicit $exec
|
|
# CHECK: %14:vgpr_32 = V_ADD_F32_e32 %11, %13, implicit $mode, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F32_e32 %12, %13, implicit $mode, implicit $exec
|
|
|
|
name: add_f16_1.0_multi_f32_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%13 = V_MOV_B32_e32 15360, implicit $exec
|
|
%14 = V_ADD_F32_e64 0, %11, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
%15 = V_ADD_F32_e64 0, %12, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_DWORD_OFFSET killed %14, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `float addrspace(1)* undef`)
|
|
BUFFER_STORE_DWORD_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `float addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
|
|
# The low 16-bits are an inline immediate, but the high bits are junk
|
|
# FIXME: Should be able to fold this
|
|
|
|
# CHECK-LABEL: name: add_f16_1.0_other_high_bits_multi_f16_use
|
|
# CHECK: %13:vgpr_32 = V_MOV_B32_e32 80886784, implicit $exec
|
|
# CHECK: %14:vgpr_32 = V_ADD_F16_e32 %11, %13, implicit $mode, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 %12, %13, implicit $mode, implicit $exec
|
|
|
|
name: add_f16_1.0_other_high_bits_multi_f16_use
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%13 = V_MOV_B32_e32 80886784, implicit $exec
|
|
%14 = V_ADD_F16_e64 0, %11, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, %12, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_SHORT_OFFSET killed %14, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|
|
---
|
|
|
|
# FIXME: Should fold inline immediate into f16 and literal use into
|
|
# f32 instruction.
|
|
|
|
# CHECK-LABEL: name: add_f16_1.0_other_high_bits_use_f16_f32
|
|
# CHECK: %13:vgpr_32 = V_MOV_B32_e32 305413120, implicit $exec
|
|
# CHECK: %14:vgpr_32 = V_ADD_F32_e32 %11, %13, implicit $mode, implicit $exec
|
|
# CHECK: %15:vgpr_32 = V_ADD_F16_e32 %12, %13, implicit $mode, implicit $exec
|
|
name: add_f16_1.0_other_high_bits_use_f16_f32
|
|
alignment: 1
|
|
exposesReturnsTwice: false
|
|
legalized: false
|
|
regBankSelected: false
|
|
selected: false
|
|
tracksRegLiveness: true
|
|
registers:
|
|
- { id: 0, class: sreg_64 }
|
|
- { id: 1, class: sreg_32 }
|
|
- { id: 2, class: sgpr_32 }
|
|
- { id: 3, class: vgpr_32 }
|
|
- { id: 4, class: sreg_64 }
|
|
- { id: 5, class: sreg_32 }
|
|
- { id: 6, class: sreg_64 }
|
|
- { id: 7, class: sreg_32 }
|
|
- { id: 8, class: sreg_32 }
|
|
- { id: 9, class: sreg_32 }
|
|
- { id: 10, class: sgpr_128 }
|
|
- { id: 11, class: vgpr_32 }
|
|
- { id: 12, class: vgpr_32 }
|
|
- { id: 13, class: vgpr_32 }
|
|
- { id: 14, class: vgpr_32 }
|
|
- { id: 15, class: vgpr_32 }
|
|
frameInfo:
|
|
isFrameAddressTaken: false
|
|
isReturnAddressTaken: false
|
|
hasStackMap: false
|
|
hasPatchPoint: false
|
|
stackSize: 0
|
|
offsetAdjustment: 0
|
|
maxAlignment: 0
|
|
adjustsStack: false
|
|
hasCalls: false
|
|
maxCallFrameSize: 0
|
|
hasOpaqueSPAdjustment: false
|
|
hasVAStart: false
|
|
hasMustTailInVarArgFunc: false
|
|
body: |
|
|
bb.0 (%ir-block.0):
|
|
%4 = IMPLICIT_DEF
|
|
%5 = COPY %4.sub1
|
|
%6 = IMPLICIT_DEF
|
|
%7 = COPY %6.sub0
|
|
%8 = S_MOV_B32 61440
|
|
%9 = S_MOV_B32 -1
|
|
%10 = REG_SEQUENCE killed %7, 1, killed %5, 2, killed %9, 3, killed %8, 4
|
|
%11 = BUFFER_LOAD_DWORD_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 4 from `float addrspace(1)* undef`)
|
|
%12 = BUFFER_LOAD_USHORT_OFFSET %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile load 2 from `half addrspace(1)* undef`)
|
|
%13 = V_MOV_B32_e32 305413120, implicit $exec
|
|
%14 = V_ADD_F32_e64 0, %11, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
%15 = V_ADD_F16_e64 0, %12, 0, %13, 0, 0, implicit $mode, implicit $exec
|
|
BUFFER_STORE_DWORD_OFFSET killed %14, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 4 into `float addrspace(1)* undef`)
|
|
BUFFER_STORE_SHORT_OFFSET killed %15, %10, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (volatile store 2 into `half addrspace(1)* undef`)
|
|
S_ENDPGM 0
|
|
|
|
...
|