llvm-project/llvm/lib/Target/AMDGPU/AMDGPUGISel.td

332 lines
12 KiB
TableGen

//===-- AMDGPUGIsel.td - AMDGPU GlobalISel Patterns---------*- tablegen -*-===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
// This files contains patterns that should only be used by GlobalISel. For
// example patterns for V_* instructions that have S_* equivalents.
// SelectionDAG does not support selecting V_* instructions.
//===----------------------------------------------------------------------===//
include "AMDGPU.td"
include "AMDGPUCombine.td"
def sd_vsrc0 : ComplexPattern<i32, 1, "">;
def gi_vsrc0 :
GIComplexOperandMatcher<s32, "selectVSRC0">,
GIComplexPatternEquiv<sd_vsrc0>;
def sd_vcsrc : ComplexPattern<i32, 1, "">;
def gi_vcsrc :
GIComplexOperandMatcher<s32, "selectVCSRC">,
GIComplexPatternEquiv<sd_vcsrc>;
def gi_vop3mods0 :
GIComplexOperandMatcher<s32, "selectVOP3Mods0">,
GIComplexPatternEquiv<VOP3Mods0>;
def gi_vop3mods :
GIComplexOperandMatcher<s32, "selectVOP3Mods">,
GIComplexPatternEquiv<VOP3Mods>;
def gi_vop3_no_mods :
GIComplexOperandMatcher<s32, "selectVOP3NoMods">,
GIComplexPatternEquiv<VOP3NoMods>;
def gi_vop3mods_nnan :
GIComplexOperandMatcher<s32, "selectVOP3Mods_nnan">,
GIComplexPatternEquiv<VOP3Mods_nnan>;
def gi_vop3omods :
GIComplexOperandMatcher<s32, "selectVOP3OMods">,
GIComplexPatternEquiv<VOP3OMods>;
def gi_vop3pmods :
GIComplexOperandMatcher<s32, "selectVOP3PMods">,
GIComplexPatternEquiv<VOP3PMods>;
def gi_vop3opselmods :
GIComplexOperandMatcher<s32, "selectVOP3OpSelMods">,
GIComplexPatternEquiv<VOP3OpSelMods>;
// FIXME: Why do we have both VOP3OpSel and VOP3OpSelMods?
def gi_vop3opsel :
GIComplexOperandMatcher<s32, "selectVOP3OpSelMods">,
GIComplexPatternEquiv<VOP3OpSel>;
def gi_smrd_imm :
GIComplexOperandMatcher<s64, "selectSmrdImm">,
GIComplexPatternEquiv<SMRDImm>;
def gi_smrd_imm32 :
GIComplexOperandMatcher<s64, "selectSmrdImm32">,
GIComplexPatternEquiv<SMRDImm32>;
def gi_smrd_sgpr :
GIComplexOperandMatcher<s64, "selectSmrdSgpr">,
GIComplexPatternEquiv<SMRDSgpr>;
def gi_flat_offset :
GIComplexOperandMatcher<s64, "selectFlatOffset">,
GIComplexPatternEquiv<FLATOffset>;
def gi_flat_offset_signed :
GIComplexOperandMatcher<s64, "selectFlatOffsetSigned">,
GIComplexPatternEquiv<FLATOffsetSigned>;
def gi_global_saddr :
GIComplexOperandMatcher<s64, "selectGlobalSAddr">,
GIComplexPatternEquiv<GlobalSAddr>;
def gi_mubuf_scratch_offset :
GIComplexOperandMatcher<s32, "selectMUBUFScratchOffset">,
GIComplexPatternEquiv<MUBUFScratchOffset>;
def gi_mubuf_scratch_offen :
GIComplexOperandMatcher<s32, "selectMUBUFScratchOffen">,
GIComplexPatternEquiv<MUBUFScratchOffen>;
def gi_ds_1addr_1offset :
GIComplexOperandMatcher<s32, "selectDS1Addr1Offset">,
GIComplexPatternEquiv<DS1Addr1Offset>;
def gi_ds_64bit_4byte_aligned :
GIComplexOperandMatcher<s64, "selectDS64Bit4ByteAligned">,
GIComplexPatternEquiv<DS64Bit4ByteAligned>;
def gi_ds_128bit_8byte_aligned :
GIComplexOperandMatcher<s64, "selectDS128Bit8ByteAligned">,
GIComplexPatternEquiv<DS128Bit8ByteAligned>;
def gi_mubuf_addr64 :
GIComplexOperandMatcher<s64, "selectMUBUFAddr64">,
GIComplexPatternEquiv<MUBUFAddr64>;
def gi_mubuf_offset :
GIComplexOperandMatcher<s64, "selectMUBUFOffset">,
GIComplexPatternEquiv<MUBUFOffset>;
def gi_mubuf_addr64_atomic :
GIComplexOperandMatcher<s64, "selectMUBUFAddr64Atomic">,
GIComplexPatternEquiv<MUBUFAddr64Atomic>;
def gi_mubuf_offset_atomic :
GIComplexOperandMatcher<s64, "selectMUBUFOffsetAtomic">,
GIComplexPatternEquiv<MUBUFOffsetAtomic>;
def gi_smrd_buffer_imm :
GIComplexOperandMatcher<s64, "selectSMRDBufferImm">,
GIComplexPatternEquiv<SMRDBufferImm>;
def gi_smrd_buffer_imm32 :
GIComplexOperandMatcher<s64, "selectSMRDBufferImm32">,
GIComplexPatternEquiv<SMRDBufferImm32>;
// Separate load nodes are defined to glue m0 initialization in
// SelectionDAG. The GISel selector can just insert m0 initialization
// directly before before selecting a glue-less load, so hide this
// distinction.
def : GINodeEquiv<G_LOAD, AMDGPUld_glue> {
let CheckMMOIsNonAtomic = 1;
}
def : GINodeEquiv<G_STORE, AMDGPUst_glue> {
let CheckMMOIsNonAtomic = 1;
}
def : GINodeEquiv<G_LOAD, AMDGPUatomic_ld_glue> {
bit CheckMMOIsAtomic = 1;
}
def : GINodeEquiv<G_STORE, AMDGPUatomic_st_glue> {
bit CheckMMOIsAtomic = 1;
}
def : GINodeEquiv<G_ATOMIC_CMPXCHG, atomic_cmp_swap_glue>;
def : GINodeEquiv<G_ATOMICRMW_XCHG, atomic_swap_glue>;
def : GINodeEquiv<G_ATOMICRMW_ADD, atomic_load_add_glue>;
def : GINodeEquiv<G_ATOMICRMW_SUB, atomic_load_sub_glue>;
def : GINodeEquiv<G_ATOMICRMW_AND, atomic_load_and_glue>;
def : GINodeEquiv<G_ATOMICRMW_OR, atomic_load_or_glue>;
def : GINodeEquiv<G_ATOMICRMW_XOR, atomic_load_xor_glue>;
def : GINodeEquiv<G_ATOMICRMW_MIN, atomic_load_min_glue>;
def : GINodeEquiv<G_ATOMICRMW_MAX, atomic_load_max_glue>;
def : GINodeEquiv<G_ATOMICRMW_UMIN, atomic_load_umin_glue>;
def : GINodeEquiv<G_ATOMICRMW_UMAX, atomic_load_umax_glue>;
def : GINodeEquiv<G_ATOMICRMW_FADD, atomic_load_fadd_glue>;
def : GINodeEquiv<G_AMDGPU_FFBH_U32, AMDGPUffbh_u32_impl>;
def : GINodeEquiv<G_AMDGPU_FMIN_LEGACY, AMDGPUfmin_legacy>;
def : GINodeEquiv<G_AMDGPU_FMAX_LEGACY, AMDGPUfmax_legacy>;
def : GINodeEquiv<G_AMDGPU_RCP_IFLAG, AMDGPUrcp_iflag>;
def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE0, AMDGPUcvt_f32_ubyte0>;
def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE1, AMDGPUcvt_f32_ubyte1>;
def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE2, AMDGPUcvt_f32_ubyte2>;
def : GINodeEquiv<G_AMDGPU_CVT_F32_UBYTE3, AMDGPUcvt_f32_ubyte3>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_CMPXCHG, AMDGPUatomic_cmp_swap>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD, SIbuffer_load>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_USHORT, SIbuffer_load_ushort>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_UBYTE, SIbuffer_load_ubyte>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_SSHORT, SIbuffer_load_short>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_SBYTE, SIbuffer_load_byte>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_FORMAT, SIbuffer_load_format>;
def : GINodeEquiv<G_AMDGPU_BUFFER_LOAD_FORMAT_D16, SIbuffer_load_format_d16>;
def : GINodeEquiv<G_AMDGPU_TBUFFER_LOAD_FORMAT, SItbuffer_load>;
def : GINodeEquiv<G_AMDGPU_TBUFFER_LOAD_FORMAT_D16, SItbuffer_load_d16>;
def : GINodeEquiv<G_AMDGPU_BUFFER_STORE, SIbuffer_store>;
def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_SHORT, SIbuffer_store_short>;
def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_BYTE, SIbuffer_store_byte>;
def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_FORMAT, SIbuffer_store_format>;
def : GINodeEquiv<G_AMDGPU_BUFFER_STORE_FORMAT_D16, SIbuffer_store_format_d16>;
def : GINodeEquiv<G_AMDGPU_TBUFFER_STORE_FORMAT, SItbuffer_store>;
def : GINodeEquiv<G_AMDGPU_TBUFFER_STORE_FORMAT_D16, SItbuffer_store_d16>;
// FIXME: Check MMO is atomic
def : GINodeEquiv<G_AMDGPU_ATOMIC_INC, SIatomic_inc>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_DEC, SIatomic_dec>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_INC, atomic_inc_glue>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_DEC, atomic_dec_glue>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_FMIN, SIatomic_fmin>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_FMAX, SIatomic_fmax>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_FMIN, atomic_load_fmin_glue>;
def : GINodeEquiv<G_AMDGPU_ATOMIC_FMAX, atomic_load_fmax_glue>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SWAP, SIbuffer_atomic_swap>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_ADD, SIbuffer_atomic_add>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SUB, SIbuffer_atomic_sub>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SMIN, SIbuffer_atomic_smin>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_UMIN, SIbuffer_atomic_umin>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_SMAX, SIbuffer_atomic_smax>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_UMAX, SIbuffer_atomic_umax>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_AND, SIbuffer_atomic_and>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_OR, SIbuffer_atomic_or>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_XOR, SIbuffer_atomic_xor>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_INC, SIbuffer_atomic_inc>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_DEC, SIbuffer_atomic_dec>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FADD, SIbuffer_atomic_fadd>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_CMPSWAP, SIbuffer_atomic_cmpswap>;
def : GINodeEquiv<G_AMDGPU_S_BUFFER_LOAD, SIsbuffer_load>;
class GISelSop2Pat <
SDPatternOperator node,
Instruction inst,
ValueType dst_vt,
ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
(dst_vt (node (src0_vt SReg_32:$src0), (src1_vt SReg_32:$src1))),
(inst src0_vt:$src0, src1_vt:$src1)
>;
class GISelVop2Pat <
SDPatternOperator node,
Instruction inst,
ValueType dst_vt,
ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
(dst_vt (node (src0_vt (sd_vsrc0 src0_vt:$src0)), (src1_vt VGPR_32:$src1))),
(inst src0_vt:$src0, src1_vt:$src1)
>;
class GISelVop2CommutePat <
SDPatternOperator node,
Instruction inst,
ValueType dst_vt,
ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
(dst_vt (node (src1_vt VGPR_32:$src1), (src0_vt (sd_vsrc0 src0_vt:$src0)))),
(inst src0_vt:$src0, src1_vt:$src1)
>;
class GISelVop3Pat2 <
SDPatternOperator node,
Instruction inst,
ValueType dst_vt,
ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
(dst_vt (node (src0_vt (sd_vcsrc src0_vt:$src0)), (src1_vt (sd_vcsrc src1_vt:$src1)))),
(inst src0_vt:$src0, src1_vt:$src1)
>;
class GISelVop3Pat2CommutePat <
SDPatternOperator node,
Instruction inst,
ValueType dst_vt,
ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
(dst_vt (node (src0_vt (sd_vcsrc src0_vt:$src0)), (src1_vt (sd_vcsrc src1_vt:$src1)))),
(inst src0_vt:$src1, src1_vt:$src0)
>;
class GISelVop3Pat2ModsPat <
SDPatternOperator node,
Instruction inst,
ValueType dst_vt,
ValueType src0_vt = dst_vt, ValueType src1_vt = src0_vt> : GCNPat <
(dst_vt (node (src0_vt (VOP3Mods0 src0_vt:$src0, i32:$src0_modifiers, i1:$clamp, i32:$omods)),
(src1_vt (VOP3Mods src1_vt:$src1, i32:$src1_modifiers)))),
(inst i32:$src0_modifiers, src0_vt:$src0,
i32:$src1_modifiers, src1_vt:$src1, $clamp, $omods)
>;
multiclass GISelVop2IntrPat <
SDPatternOperator node, Instruction inst,
ValueType dst_vt, ValueType src_vt = dst_vt> {
def : GISelVop2Pat <node, inst, dst_vt, src_vt>;
// FIXME: Intrinsics aren't marked as commutable, so we need to add an explicit
// pattern to handle commuting. This is another reason why legalizing to a
// generic machine instruction may be better that matching the intrinsic
// directly.
def : GISelVop2CommutePat <node, inst, dst_vt, src_vt>;
}
// Since GlobalISel is more flexible then SelectionDAG, I think we can get
// away with adding patterns for integer types and not legalizing all
// loads and stores to vector types. This should help simplify the load/store
// legalization.
foreach Ty = [i64, p0, p1, p4] in {
defm : SMRD_Pattern <"S_LOAD_DWORDX2", Ty>;
}
def gi_as_i32timm : GICustomOperandRenderer<"renderTruncTImm32">,
GISDNodeXFormEquiv<as_i32timm>;
def gi_as_i16timm : GICustomOperandRenderer<"renderTruncTImm16">,
GISDNodeXFormEquiv<as_i16timm>;
def gi_as_i8timm : GICustomOperandRenderer<"renderTruncTImm8">,
GISDNodeXFormEquiv<as_i8timm>;
def gi_as_i1timm : GICustomOperandRenderer<"renderTruncTImm1">,
GISDNodeXFormEquiv<as_i1timm>;
def gi_NegateImm : GICustomOperandRenderer<"renderNegateImm">,
GISDNodeXFormEquiv<NegateImm>;
def gi_bitcast_fpimm_to_i32 : GICustomOperandRenderer<"renderBitcastImm">,
GISDNodeXFormEquiv<bitcast_fpimm_to_i32>;
def gi_IMMPopCount : GICustomOperandRenderer<"renderPopcntImm">,
GISDNodeXFormEquiv<IMMPopCount>;
def gi_extract_glc : GICustomOperandRenderer<"renderExtractGLC">,
GISDNodeXFormEquiv<extract_glc>;
def gi_extract_slc : GICustomOperandRenderer<"renderExtractSLC">,
GISDNodeXFormEquiv<extract_slc>;
def gi_extract_dlc : GICustomOperandRenderer<"renderExtractDLC">,
GISDNodeXFormEquiv<extract_dlc>;
def gi_extract_swz : GICustomOperandRenderer<"renderExtractSWZ">,
GISDNodeXFormEquiv<extract_swz>;
def gi_frameindex_to_targetframeindex : GICustomOperandRenderer<"renderFrameIndex">,
GISDNodeXFormEquiv<frameindex_to_targetframeindex>;