1
0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-23 13:02:52 +02:00
llvm-mirror/test/CodeGen/AMDGPU/imm.ll

688 lines
26 KiB
LLVM
Raw Normal View History

; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=verde -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=SI %s
; RUN: llc -amdgpu-scalarize-global-loads=false -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefix=GCN -check-prefix=VI %s
; Use a 64-bit value with lo bits that can be represented as an inline constant
; GCN-LABEL: {{^}}i64_imm_inline_lo:
; GCN: v_mov_b32_e32 v[[LO_VGPR:[0-9]+]], 5
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VGPR]]:
define amdgpu_kernel void @i64_imm_inline_lo(i64 addrspace(1) *%out) {
entry:
store i64 1311768464867721221, i64 addrspace(1) *%out ; 0x1234567800000005
ret void
}
; Use a 64-bit value with hi bits that can be represented as an inline constant
; GCN-LABEL: {{^}}i64_imm_inline_hi:
; GCN: v_mov_b32_e32 v[[HI_VGPR:[0-9]+]], 5
; GCN: buffer_store_dwordx2 v{{\[[0-9]+:}}[[HI_VGPR]]
define amdgpu_kernel void @i64_imm_inline_hi(i64 addrspace(1) *%out) {
entry:
store i64 21780256376, i64 addrspace(1) *%out ; 0x0000000512345678
ret void
}
; GCN-LABEL: {{^}}store_imm_neg_0.0_i64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_bfrev_b32_e32 v[[HI_VREG:[0-9]+]], 1{{$}}
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_imm_neg_0.0_i64(i64 addrspace(1) *%out) {
store i64 -9223372036854775808, i64 addrspace(1) *%out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_neg_0.0_i32:
; GCN: v_bfrev_b32_e32 [[REG:v[0-9]+]], 1{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_neg_0.0_i32(i32 addrspace(1)* %out) {
store i32 -2147483648, i32 addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_0.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_0.0_f32(float addrspace(1)* %out) {
store float 0.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_imm_neg_0.0_f32:
; GCN: v_bfrev_b32_e32 [[REG:v[0-9]+]], 1{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_imm_neg_0.0_f32(float addrspace(1)* %out) {
store float -0.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_0.5_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0.5{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_0.5_f32(float addrspace(1)* %out) {
store float 0.5, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_0.5_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], -0.5{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_m_0.5_f32(float addrspace(1)* %out) {
store float -0.5, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_1.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 1.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_1.0_f32(float addrspace(1)* %out) {
store float 1.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_1.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], -1.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_m_1.0_f32(float addrspace(1)* %out) {
store float -1.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_2.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 2.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_2.0_f32(float addrspace(1)* %out) {
store float 2.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_2.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], -2.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_m_2.0_f32(float addrspace(1)* %out) {
store float -2.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_4.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 4.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_4.0_f32(float addrspace(1)* %out) {
store float 4.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_4.0_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], -4.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_m_4.0_f32(float addrspace(1)* %out) {
store float -4.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_inv_2pi_f32:
; SI: v_mov_b32_e32 [[REG:v[0-9]+]], 0x3e22f983{{$}}
; VI: v_mov_b32_e32 [[REG:v[0-9]+]], 0.15915494{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_inv_2pi_f32(float addrspace(1)* %out) {
store float 0x3FC45F3060000000, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_inv_2pi_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0xbe22f983{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_inline_imm_m_inv_2pi_f32(float addrspace(1)* %out) {
store float 0xBFC45F3060000000, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_literal_imm_f32:
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], 0x45800000
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @store_literal_imm_f32(float addrspace(1)* %out) {
store float 4096.0, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_0.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_0.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_0.5_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 0.5{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_0.5_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0.5
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_0.5_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], -0.5{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_0.5_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, -0.5
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_1.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 1.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_1.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 1.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_1.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], -1.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_1.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, -1.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_2.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 2.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_2.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 2.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_2.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], -2.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_2.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, -2.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_4.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 4.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_4.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 4.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_4.0_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], -4.0{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_4.0_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, -4.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}commute_add_inline_imm_0.5_f32:
; GCN: buffer_load_dword [[VAL:v[0-9]+]]
; GCN: v_add_f32_e32 [[REG:v[0-9]+]], 0.5, [[VAL]]
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @commute_add_inline_imm_0.5_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
%x = load float, float addrspace(1)* %in
%y = fadd float %x, 0.5
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}commute_add_literal_f32:
; GCN: buffer_load_dword [[VAL:v[0-9]+]]
; GCN: v_add_f32_e32 [[REG:v[0-9]+]], 0x44800000, [[VAL]]
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @commute_add_literal_f32(float addrspace(1)* %out, float addrspace(1)* %in) {
%x = load float, float addrspace(1)* %in
%y = fadd float %x, 1024.0
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_1_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 1{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_1_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0x36a0000000000000
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_2_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 2{{$}}
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_2_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0x36b0000000000000
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_16_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 16
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_16_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0x36e0000000000000
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_1_f32:
; GCN: s_add_i32 [[VAL:s[0-9]+]], s0, -1
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], [[VAL]]
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_1_f32(float addrspace(1)* %out, float %x) {
%xbc = bitcast float %x to i32
%y = add i32 %xbc, -1
%ybc = bitcast i32 %y to float
store float %ybc, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_2_f32:
; GCN: s_add_i32 [[VAL:s[0-9]+]], s0, -2
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], [[VAL]]
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_2_f32(float addrspace(1)* %out, float %x) {
%xbc = bitcast float %x to i32
%y = add i32 %xbc, -2
%ybc = bitcast i32 %y to float
store float %ybc, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_16_f32:
; GCN: s_add_i32 [[VAL:s[0-9]+]], s0, -16
; GCN: v_mov_b32_e32 [[REG:v[0-9]+]], [[VAL]]
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_neg_16_f32(float addrspace(1)* %out, float %x) {
%xbc = bitcast float %x to i32
%y = add i32 %xbc, -16
%ybc = bitcast i32 %y to float
store float %ybc, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_63_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 63
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_63_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0x36ff800000000000
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_64_f32:
; GCN: s_load_dword [[VAL:s[0-9]+]]
; GCN: v_add_f32_e64 [[REG:v[0-9]+]], [[VAL]], 64
; GCN: buffer_store_dword [[REG]]
define amdgpu_kernel void @add_inline_imm_64_f32(float addrspace(1)* %out, float %x) {
%y = fadd float %x, 0x3700000000000000
store float %y, float addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_0.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 0{{$}}
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_0.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_0.5_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 0.5
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_0.5_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0.5
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_0.5_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], -0.5
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_0.5_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, -0.5
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_1.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 1.0
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_1.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 1.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_1.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], -1.0
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_1.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, -1.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_2.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 2.0
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_2.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 2.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_2.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], -2.0
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_2.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, -2.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_4.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 4.0
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_4.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 4.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_4.0_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], -4.0
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_4.0_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, -4.0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_inv_2pi_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; SI-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0x6dc9c882
; SI-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0x3fc45f30
; SI: v_add_f64 v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; VI: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 0.15915494{{$}}
; VI: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_inv_2pi_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0x3fc45f306dc9c882
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_m_inv_2pi_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0x6dc9c882
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0xbfc45f30
; GCN: v_add_f64 v{{\[[0-9]+:[0-9]+\]}}, s{{\[[0-9]+:[0-9]+\]}}, v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_m_inv_2pi_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0xbfc45f306dc9c882
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_1_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 1{{$}}
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_1_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0x0000000000000001
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_2_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 2{{$}}
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_2_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0x0000000000000002
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_16_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 16
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_16_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0x0000000000000010
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_1_f64:
; GCN: v_mov_b32_e32 v0, -1
; GCN: v_mov_b32_e32 v1, v0
; GCN: buffer_store_dwordx2 v[0:1]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_1_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0xffffffffffffffff
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_2_f64:
; GCN: v_mov_b32_e32 v0, -2
; GCN: v_mov_b32_e32 v1, -1
; GCN: buffer_store_dwordx2 v[0:1]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_2_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0xfffffffffffffffe
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_neg_16_f64:
; GCN: v_mov_b32_e32 v0, -16
; GCN: v_mov_b32_e32 v1, -1
; GCN: buffer_store_dwordx2 v[0:1]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_neg_16_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0xfffffffffffffff0
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_63_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 63
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_63_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0x000000000000003F
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}add_inline_imm_64_f64:
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
; SI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x13
; VI: s_load_dwordx2 [[VAL:s\[[0-9]+:[0-9]+\]]], {{s\[[0-9]+:[0-9]+\]}}, 0x4c
; GCN: v_add_f64 [[REG:v\[[0-9]+:[0-9]+\]]], [[VAL]], 64
; GCN: buffer_store_dwordx2 [[REG]]
AMDGPU: Add pass to lower kernel arguments to loads This replaces most argument uses with loads, but for now not all. The code in SelectionDAG for calling convention lowering is actively harmful for amdgpu_kernel. It attempts to split the argument types into register legal types, which results in low quality code for arbitary types. Since all kernel arguments are passed in memory, we just want the raw types. I've tried a couple of methods of mitigating this in SelectionDAG, but it's easier to just bypass this problem alltogether. It's possible to hack around the problem in the initial lowering, but the real problem is the DAG then expects to be able to use CopyToReg/CopyFromReg for uses of the arguments outside the block. Exposing the argument loads in the IR also has the advantage that the LoadStoreVectorizer can merge them. I'm not sure the best approach to dealing with the IR argument list is. The patch as-is just leaves the IR arguments in place, so all the existing code will still compute the same kernarg size and pointlessly lowers the arguments. Arguably the frontend should emit kernels with an empty argument list in the first place. Alternatively a dummy array could be inserted as a single argument just to reserve space. This does have some disadvantages. Local pointer kernel arguments can no longer have AssertZext placed on them as the equivalent !range metadata is not valid on pointer typed loads. This is mostly bad for SI which needs to know about the known bits in order to use the DS instruction offset, so in this case this is not done. More importantly, this skips noalias arguments since this pass does not yet convert this to the equivalent !alias.scope and !noalias metadata. Producing this metadata correctly seems to be tricky, although this logically is the same as inlining into a function which doesn't exist. Additionally, exposing these loads to the vectorizer may result in degraded aliasing information if a pointer load is merged with another argument load. I'm also not entirely sure this is preserving the current clover ABI, although I would greatly prefer if it would stop widening arguments and match the HSA ABI. As-is I think it is extending < 4-byte arguments to 4-bytes but doesn't align them to 4-bytes. llvm-svn: 335650
2018-06-26 21:10:00 +02:00
define amdgpu_kernel void @add_inline_imm_64_f64(double addrspace(1)* %out, [8 x i32], double %x) {
%y = fadd double %x, 0x0000000000000040
store double %y, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_0.0_f64:
; GCN: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0
; GCN: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], v[[LO_VREG]]{{$}}
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_0.0_f64(double addrspace(1)* %out) {
store double 0.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_literal_imm_neg_0.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_bfrev_b32_e32 v[[HI_VREG:[0-9]+]], 1{{$}}
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_literal_imm_neg_0.0_f64(double addrspace(1)* %out) {
store double -0.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_0.5_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0x3fe00000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_0.5_f64(double addrspace(1)* %out) {
store double 0.5, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_0.5_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0xbfe00000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_m_0.5_f64(double addrspace(1)* %out) {
store double -0.5, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_1.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0x3ff00000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_1.0_f64(double addrspace(1)* %out) {
store double 1.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_1.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0xbff00000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_m_1.0_f64(double addrspace(1)* %out) {
store double -1.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_2.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 2.0
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_2.0_f64(double addrspace(1)* %out) {
store double 2.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_2.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], -2.0
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_m_2.0_f64(double addrspace(1)* %out) {
store double -2.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_4.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0x40100000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_4.0_f64(double addrspace(1)* %out) {
store double 4.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_4.0_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0xc0100000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_m_4.0_f64(double addrspace(1)* %out) {
store double -4.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inv_2pi_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0x6dc9c882
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0x3fc45f30
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inv_2pi_f64(double addrspace(1)* %out) {
store double 0x3fc45f306dc9c882, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_inline_imm_m_inv_2pi_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0x6dc9c882
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0xbfc45f30
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_inline_imm_m_inv_2pi_f64(double addrspace(1)* %out) {
store double 0xbfc45f306dc9c882, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}store_literal_imm_f64:
; GCN-DAG: v_mov_b32_e32 v[[LO_VREG:[0-9]+]], 0{{$}}
; GCN-DAG: v_mov_b32_e32 v[[HI_VREG:[0-9]+]], 0x40b00000
; GCN: buffer_store_dwordx2 v{{\[}}[[LO_VREG]]:[[HI_VREG]]{{\]}}
define amdgpu_kernel void @store_literal_imm_f64(double addrspace(1)* %out) {
store double 4096.0, double addrspace(1)* %out
ret void
}
; GCN-LABEL: {{^}}literal_folding:
; GCN: v_mul_f32_e32 v{{[0-9]+}}, 0x3f4353f8, v{{[0-9]+}}
; GCN: v_mul_f32_e32 v{{[0-9]+}}, 0xbf4353f8, v{{[0-9]+}}
define amdgpu_vs void @literal_folding(float %arg) {
main_body:
%tmp = fmul float %arg, 0x3FE86A7F00000000
%tmp1 = fmul float %arg, 0xBFE86A7F00000000
call void @llvm.amdgcn.exp.f32(i32 12, i32 15, float %tmp, float %tmp, float %tmp1, float %tmp1, i1 true, i1 false) #0
ret void
}
declare void @llvm.amdgcn.exp.f32(i32, i32, float, float, float, float, i1, i1) #0
attributes #0 = { nounwind }