llvm-project/llvm/test/CodeGen/AMDGPU/imm16.ll

320 lines
11 KiB
LLVM
Raw Normal View History

; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=tonga -mattr=-flat-for-global -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
; RUN: llc -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
; FIXME: Merge into imm.ll
; GCN-LABEL: {{^}}store_inline_imm_neg_0.0_i16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0x8000{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff8000{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_neg_0.0_i16(i16 addrspace(1)* %out) {
store volatile i16 -32768, i16 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_0.0_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_0.0_f16(half addrspace(1)* %out) {
store half 0.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_imm_neg_0.0_f16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0x8000{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffff8000{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_imm_neg_0.0_f16(half addrspace(1)* %out) {
store half -0.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_0.5_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3800{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_0.5_f16(half addrspace(1)* %out) {
store half 0.5, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_0.5_f16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xb800{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffb800{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_m_0.5_f16(half addrspace(1)* %out) {
store half -0.5, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_1.0_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3c00{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_1.0_f16(half addrspace(1)* %out) {
store half 1.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_1.0_f16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xbc00{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffbc00{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_m_1.0_f16(half addrspace(1)* %out) {
store half -1.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_2.0_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4000{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_2.0_f16(half addrspace(1)* %out) {
store half 2.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_2.0_f16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xc000{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffc000{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_m_2.0_f16(half addrspace(1)* %out) {
store half -2.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_4.0_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x4400{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_4.0_f16(half addrspace(1)* %out) {
store half 4.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_4.0_f16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xc400{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffc400{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_m_4.0_f16(half addrspace(1)* %out) {
store half -4.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_inv_2pi_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3118{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_inv_2pi_f16(half addrspace(1)* %out) {
store half 0xH3118, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_inv_2pi_f16:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xb118{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0xffffb118{{$}}
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_inline_imm_m_inv_2pi_f16(half addrspace(1)* %out) {
store half 0xHB118, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_literal_imm_f16:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x6c00
; GCN: buffer_store_short [[REG]]
define amdgpu_kernel void @store_literal_imm_f16(half addrspace(1)* %out) {
store half 4096.0, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_0.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_0.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_0.5_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 0.5{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_0.5_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0.5
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_0.5_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], -0.5{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_0.5_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, -0.5
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_1.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 1.0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_1.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 1.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_1.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], -1.0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_1.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, -1.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_2.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 2.0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_2.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 2.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_2.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], -2.0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_2.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, -2.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_4.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 4.0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_4.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 4.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_4.0_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], -4.0{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_4.0_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, -4.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}commute_add_inline_imm_0.5_f16:
; VI: buffer_load_ushort [[VAL:v[0-9]+]]
; VI: v_add_f16_e32 [[REG:v[0-9]+]], 0.5, [[VAL]]
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @commute_add_inline_imm_0.5_f16(half addrspace(1)* %out, half addrspace(1)* %in) {
%x = load half, half addrspace(1)* %in
%y = fadd half %x, 0.5
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}commute_add_literal_f16:
; VI: buffer_load_ushort [[VAL:v[0-9]+]]
; VI: v_add_f16_e32 [[REG:v[0-9]+]], 0x6400, [[VAL]]
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @commute_add_literal_f16(half addrspace(1)* %out, half addrspace(1)* %in) {
%x = load half, half addrspace(1)* %in
%y = fadd half %x, 1024.0
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_1_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 1{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_1_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0xH0001
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_2_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 2{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_2_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0xH0002
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_16_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 16{{$}}
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_16_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0xH0010
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_1_f16:
; VI: v_add_u16_e32 [[REG:v[0-9]+]], -1, [[REG:v[0-9]+]]
; VI: buffer_store_short [[REG]]
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
define amdgpu_kernel void @add_inline_imm_neg_1_f16(half addrspace(1)* %out, i16 addrspace(1)* %in) {
%x = load i16, i16 addrspace(1)* %in
%y = add i16 %x, -1
%ybc = bitcast i16 %y to half
store half %ybc, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_2_f16:
; VI: v_add_u16_e32 [[REG:v[0-9]+]], -2, [[REG:v[0-9]+]]
; VI: buffer_store_short [[REG]]
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
define amdgpu_kernel void @add_inline_imm_neg_2_f16(half addrspace(1)* %out, i16 addrspace(1)* %in) {
%x = load i16, i16 addrspace(1)* %in
%y = add i16 %x, -2
%ybc = bitcast i16 %y to half
store half %ybc, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_16_f16:
; VI: v_add_u16_e32 [[REG:v[0-9]+]], -16, [[REG:v[0-9]+]]
; VI: buffer_store_short [[REG]]
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
define amdgpu_kernel void @add_inline_imm_neg_16_f16(half addrspace(1)* %out, i16 addrspace(1)* %in) {
%x = load i16, i16 addrspace(1)* %in
%y = add i16 %x, -16
%ybc = bitcast i16 %y to half
store half %ybc, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_63_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 63
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_63_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0xH003F
store half %y, half addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_64_f16:
AMDGPU: Try a lot harder to emit scalar loads This has two main components. First, widen widen short constant loads in DAG when they have the correct alignment. This is already done a bit in AMDGPUCodeGenPrepare, since that has access to DivergenceAnalysis. This can't help kernarg loads created in the DAG. Start to use DAG divergence analysis to help this case. The second part is to avoid kernel argument lowering breaking the alignment of short vector elements because calling convention lowering wants to split everything into legal register types. When loading a split type, load the nearest 4-byte aligned segment and shift to get the desired bits. This extra load of the earlier argument piece ends up merging, and the bit extract hopefully folds out. There are a number of improvements and regressions with this, but I think as-is this is a better compromise between several of the worst parts of SelectionDAG. Particularly when i16 is legal, this produces worse code for i8 and i16 element vector kernel arguments. This is partially due to the very weak load merging the DAG does. It only looks for fairly specific combines between pairs of loads which no longer appear. In particular this causes v4i16 loads to be split into 2 components when previously the two halves were merged. Worse, because of the newly introduced shifts, there is a lot more unnecessary vector packing and unpacking code emitted. At least some of this is due to reporting false for isTypeDesirableForOp for i16 as a workaround for the lack of divergence information in the DAG. The cases where this happens it doesn't actually matter, but the relevant code in SimplifyDemandedBits doens't have the context to know to ignore this. The use of the scalar cache is probably more important than the mess of mostly scalar instructions doing this packing and unpacking. Future work can fix this, possibly by making better use of the new DAG divergence information for controlling promotion decisions, or adding another version of shift + trunc + shift combines that doesn't only know about the used types. llvm-svn: 334180
2018-06-07 17:54:49 +08:00
; VI: s_load_dword [[VAL:s[0-9]+]]
; VI: v_add_f16_e64 [[REG:v[0-9]+]], [[VAL]], 64
; VI: buffer_store_short [[REG]]
define amdgpu_kernel void @add_inline_imm_64_f16(half addrspace(1)* %out, half %x) {
%y = fadd half %x, 0xH0040
store half %y, half addrspace(1)* %out
ret void
}