1 //===-- AMDGPUInstrInfo.td - AMDGPU DAG nodes --------------*- tablegen -*-===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains DAG node defintions for the AMDGPU target.
12 //===----------------------------------------------------------------------===//
14 //===----------------------------------------------------------------------===//
15 // AMDGPU DAG Profiles
16 //===----------------------------------------------------------------------===//
18 def AMDGPUDTIntTernaryOp : SDTypeProfile<1, 3, [
19 SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisInt<0>, SDTCisInt<3>
22 def AMDGPUTrigPreOp : SDTypeProfile<1, 2,
23 [SDTCisSameAs<0, 1>, SDTCisFP<0>, SDTCisInt<2>]
26 def AMDGPULdExpOp : SDTypeProfile<1, 2,
27 [SDTCisSameAs<0, 1>, SDTCisFP<0>, SDTCisInt<2>]
30 def AMDGPUFPClassOp : SDTypeProfile<1, 2,
31 [SDTCisInt<0>, SDTCisFP<1>, SDTCisInt<2>]
34 def AMDGPUFPPackOp : SDTypeProfile<1, 2,
35 [SDTCisFP<1>, SDTCisSameAs<1, 2>]
38 def AMDGPUDivScaleOp : SDTypeProfile<2, 3,
39 [SDTCisFP<0>, SDTCisInt<1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisSameAs<0, 4>]
42 // float, float, float, vcc
43 def AMDGPUFmasOp : SDTypeProfile<1, 4,
44 [SDTCisFP<0>, SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisInt<4>]
47 def AMDGPUKillSDT : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
49 def AMDGPUIfOp : SDTypeProfile<1, 2,
50 [SDTCisVT<0, i64>, SDTCisVT<1, i1>, SDTCisVT<2, OtherVT>]
53 def AMDGPUElseOp : SDTypeProfile<1, 2,
54 [SDTCisVT<0, i64>, SDTCisVT<1, i64>, SDTCisVT<2, OtherVT>]
57 def AMDGPULoopOp : SDTypeProfile<0, 2,
58 [SDTCisVT<0, i64>, SDTCisVT<1, OtherVT>]
61 def AMDGPUBreakOp : SDTypeProfile<1, 1,
62 [SDTCisVT<0, i64>, SDTCisVT<1, i64>]
65 def AMDGPUIfBreakOp : SDTypeProfile<1, 2,
66 [SDTCisVT<0, i64>, SDTCisVT<1, i1>, SDTCisVT<2, i64>]
69 def AMDGPUElseBreakOp : SDTypeProfile<1, 2,
70 [SDTCisVT<0, i64>, SDTCisVT<1, i64>, SDTCisVT<2, i64>]
73 //===----------------------------------------------------------------------===//
77 def AMDGPUif : SDNode<"AMDGPUISD::IF", AMDGPUIfOp, [SDNPHasChain]>;
78 def AMDGPUelse : SDNode<"AMDGPUISD::ELSE", AMDGPUElseOp, [SDNPHasChain]>;
79 def AMDGPUloop : SDNode<"AMDGPUISD::LOOP", AMDGPULoopOp, [SDNPHasChain]>;
81 def AMDGPUconstdata_ptr : SDNode<
82 "AMDGPUISD::CONST_DATA_PTR", SDTypeProfile <1, 1, [SDTCisVT<0, iPTR>,
86 // This argument to this node is a dword address.
87 def AMDGPUdwordaddr : SDNode<"AMDGPUISD::DWORDADDR", SDTIntUnaryOp>;
89 // Force dependencies for vector trunc stores
90 def R600dummy_chain : SDNode<"AMDGPUISD::DUMMY_CHAIN", SDTNone, [SDNPHasChain]>;
92 def AMDGPUcos : SDNode<"AMDGPUISD::COS_HW", SDTFPUnaryOp>;
93 def AMDGPUsin : SDNode<"AMDGPUISD::SIN_HW", SDTFPUnaryOp>;
96 def AMDGPUfract : SDNode<"AMDGPUISD::FRACT", SDTFPUnaryOp>;
99 def AMDGPUrcp : SDNode<"AMDGPUISD::RCP", SDTFPUnaryOp>;
101 // out = 1.0 / sqrt(a)
102 def AMDGPUrsq : SDNode<"AMDGPUISD::RSQ", SDTFPUnaryOp>;
104 // out = 1.0 / sqrt(a)
105 def AMDGPUrcp_legacy : SDNode<"AMDGPUISD::RCP_LEGACY", SDTFPUnaryOp>;
106 def AMDGPUrsq_legacy : SDNode<"AMDGPUISD::RSQ_LEGACY", SDTFPUnaryOp>;
108 // out = 1.0 / sqrt(a) result clamped to +/- max_float.
109 def AMDGPUrsq_clamp : SDNode<"AMDGPUISD::RSQ_CLAMP", SDTFPUnaryOp>;
111 def AMDGPUldexp : SDNode<"AMDGPUISD::LDEXP", AMDGPULdExpOp>;
113 def AMDGPUpkrtz_f16_f32 : SDNode<"AMDGPUISD::CVT_PKRTZ_F16_F32", AMDGPUFPPackOp>;
114 def AMDGPUfp_to_f16 : SDNode<"AMDGPUISD::FP_TO_FP16" , SDTFPToIntOp>;
115 def AMDGPUfp16_zext : SDNode<"AMDGPUISD::FP16_ZEXT" , SDTFPToIntOp>;
118 def AMDGPUfp_class : SDNode<"AMDGPUISD::FP_CLASS", AMDGPUFPClassOp>;
120 // out = max(a, b) a and b are floats, where a nan comparison fails.
121 // This is not commutative because this gives the second operand:
122 // x < nan ? x : nan -> nan
123 // nan < x ? nan : x -> x
124 def AMDGPUfmax_legacy : SDNode<"AMDGPUISD::FMAX_LEGACY", SDTFPBinOp,
128 def AMDGPUfmul_legacy : SDNode<"AMDGPUISD::FMUL_LEGACY", SDTFPBinOp,
129 [SDNPCommutative, SDNPAssociative]
132 def AMDGPUclamp : SDNode<"AMDGPUISD::CLAMP", SDTFPUnaryOp>;
134 // out = min(a, b) a and b are floats, where a nan comparison fails.
135 def AMDGPUfmin_legacy : SDNode<"AMDGPUISD::FMIN_LEGACY", SDTFPBinOp,
139 // FIXME: TableGen doesn't like commutative instructions with more
141 // out = max(a, b, c) a, b and c are floats
142 def AMDGPUfmax3 : SDNode<"AMDGPUISD::FMAX3", SDTFPTernaryOp,
143 [/*SDNPCommutative, SDNPAssociative*/]
146 // out = max(a, b, c) a, b, and c are signed ints
147 def AMDGPUsmax3 : SDNode<"AMDGPUISD::SMAX3", AMDGPUDTIntTernaryOp,
148 [/*SDNPCommutative, SDNPAssociative*/]
151 // out = max(a, b, c) a, b and c are unsigned ints
152 def AMDGPUumax3 : SDNode<"AMDGPUISD::UMAX3", AMDGPUDTIntTernaryOp,
153 [/*SDNPCommutative, SDNPAssociative*/]
156 // out = min(a, b, c) a, b and c are floats
157 def AMDGPUfmin3 : SDNode<"AMDGPUISD::FMIN3", SDTFPTernaryOp,
158 [/*SDNPCommutative, SDNPAssociative*/]
161 // out = min(a, b, c) a, b and c are signed ints
162 def AMDGPUsmin3 : SDNode<"AMDGPUISD::SMIN3", AMDGPUDTIntTernaryOp,
163 [/*SDNPCommutative, SDNPAssociative*/]
166 // out = min(a, b) a and b are unsigned ints
167 def AMDGPUumin3 : SDNode<"AMDGPUISD::UMIN3", AMDGPUDTIntTernaryOp,
168 [/*SDNPCommutative, SDNPAssociative*/]
171 // out = (src0 + src1 > 0xFFFFFFFF) ? 1 : 0
172 def AMDGPUcarry : SDNode<"AMDGPUISD::CARRY", SDTIntBinOp, []>;
174 // out = (src1 > src0) ? 1 : 0
175 def AMDGPUborrow : SDNode<"AMDGPUISD::BORROW", SDTIntBinOp, []>;
177 def AMDGPUSetCCOp : SDTypeProfile<1, 3, [ // setcc
178 SDTCisVT<0, i64>, SDTCisSameAs<1, 2>, SDTCisVT<3, OtherVT>
181 def AMDGPUsetcc : SDNode<"AMDGPUISD::SETCC", AMDGPUSetCCOp>;
183 def AMDGPUSetRegOp : SDTypeProfile<0, 2, [
184 SDTCisInt<0>, SDTCisInt<1>
187 def AMDGPUsetreg : SDNode<"AMDGPUISD::SETREG", AMDGPUSetRegOp, [
188 SDNPHasChain, SDNPSideEffect, SDNPOptInGlue, SDNPOutGlue]>;
190 def AMDGPUfma : SDNode<"AMDGPUISD::FMA_W_CHAIN", SDTFPTernaryOp, [
191 SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
193 def AMDGPUmul : SDNode<"AMDGPUISD::FMUL_W_CHAIN", SDTFPBinOp, [
194 SDNPHasChain, SDNPOptInGlue, SDNPOutGlue]>;
196 def AMDGPUcvt_f32_ubyte0 : SDNode<"AMDGPUISD::CVT_F32_UBYTE0",
198 def AMDGPUcvt_f32_ubyte1 : SDNode<"AMDGPUISD::CVT_F32_UBYTE1",
200 def AMDGPUcvt_f32_ubyte2 : SDNode<"AMDGPUISD::CVT_F32_UBYTE2",
202 def AMDGPUcvt_f32_ubyte3 : SDNode<"AMDGPUISD::CVT_F32_UBYTE3",
206 // urecip - This operation is a helper for integer division, it returns the
207 // result of 1 / a as a fractional unsigned integer.
208 // out = (2^32 / a) + e
209 // e is rounding error
210 def AMDGPUurecip : SDNode<"AMDGPUISD::URECIP", SDTIntUnaryOp>;
212 // Special case divide preop and flags.
213 def AMDGPUdiv_scale : SDNode<"AMDGPUISD::DIV_SCALE", AMDGPUDivScaleOp>;
215 // Special case divide FMA with scale and flags (src0 = Quotient,
216 // src1 = Denominator, src2 = Numerator).
217 def AMDGPUdiv_fmas : SDNode<"AMDGPUISD::DIV_FMAS", AMDGPUFmasOp>;
219 // Single or double precision division fixup.
220 // Special case divide fixup and flags(src0 = Quotient, src1 =
221 // Denominator, src2 = Numerator).
222 def AMDGPUdiv_fixup : SDNode<"AMDGPUISD::DIV_FIXUP", SDTFPTernaryOp>;
224 def AMDGPUfmad_ftz : SDNode<"AMDGPUISD::FMAD_FTZ", SDTFPTernaryOp>;
226 // Look Up 2.0 / pi src0 with segment select src1[4:0]
227 def AMDGPUtrig_preop : SDNode<"AMDGPUISD::TRIG_PREOP", AMDGPUTrigPreOp>;
229 def AMDGPUregister_load : SDNode<"AMDGPUISD::REGISTER_LOAD",
230 SDTypeProfile<1, 2, [SDTCisPtrTy<1>, SDTCisInt<2>]>,
231 [SDNPHasChain, SDNPMayLoad]>;
233 def AMDGPUregister_store : SDNode<"AMDGPUISD::REGISTER_STORE",
234 SDTypeProfile<0, 3, [SDTCisPtrTy<1>, SDTCisInt<2>]>,
235 [SDNPHasChain, SDNPMayStore]>;
237 // MSKOR instructions are atomic memory instructions used mainly for storing
238 // 8-bit and 16-bit values. The definition is:
240 // MSKOR(dst, mask, src) MEM[dst] = ((MEM[dst] & ~mask) | src)
242 // src0: vec4(src, 0, 0, mask)
243 // src1: dst - rat offset (aka pointer) in dwords
244 def AMDGPUstore_mskor : SDNode<"AMDGPUISD::STORE_MSKOR",
245 SDTypeProfile<0, 2, []>,
246 [SDNPHasChain, SDNPMayStore, SDNPMemOperand]>;
248 def AMDGPUatomic_cmp_swap : SDNode<"AMDGPUISD::ATOMIC_CMP_SWAP",
249 SDTypeProfile<1, 2, [SDTCisPtrTy<1>, SDTCisVec<2>]>,
250 [SDNPHasChain, SDNPMayStore, SDNPMayLoad,
253 def AMDGPUround : SDNode<"ISD::FROUND",
254 SDTypeProfile<1, 1, [SDTCisFP<0>, SDTCisSameAs<0,1>]>>;
256 def AMDGPUbfe_u32 : SDNode<"AMDGPUISD::BFE_U32", AMDGPUDTIntTernaryOp>;
257 def AMDGPUbfe_i32 : SDNode<"AMDGPUISD::BFE_I32", AMDGPUDTIntTernaryOp>;
258 def AMDGPUbfi : SDNode<"AMDGPUISD::BFI", AMDGPUDTIntTernaryOp>;
259 def AMDGPUbfm : SDNode<"AMDGPUISD::BFM", SDTIntBinOp>;
261 def AMDGPUffbh_u32 : SDNode<"AMDGPUISD::FFBH_U32", SDTIntUnaryOp>;
262 def AMDGPUffbh_i32 : SDNode<"AMDGPUISD::FFBH_I32", SDTIntUnaryOp>;
264 // Signed and unsigned 24-bit multiply. The highest 8-bits are ignore
265 // when performing the mulitply. The result is a 32-bit value.
266 def AMDGPUmul_u24 : SDNode<"AMDGPUISD::MUL_U24", SDTIntBinOp,
267 [SDNPCommutative, SDNPAssociative]
269 def AMDGPUmul_i24 : SDNode<"AMDGPUISD::MUL_I24", SDTIntBinOp,
270 [SDNPCommutative, SDNPAssociative]
273 def AMDGPUmulhi_u24 : SDNode<"AMDGPUISD::MULHI_U24", SDTIntBinOp,
274 [SDNPCommutative, SDNPAssociative]
276 def AMDGPUmulhi_i24 : SDNode<"AMDGPUISD::MULHI_I24", SDTIntBinOp,
277 [SDNPCommutative, SDNPAssociative]
280 def AMDGPUmad_u24 : SDNode<"AMDGPUISD::MAD_U24", AMDGPUDTIntTernaryOp,
283 def AMDGPUmad_i24 : SDNode<"AMDGPUISD::MAD_I24", AMDGPUDTIntTernaryOp,
287 def AMDGPUsmed3 : SDNode<"AMDGPUISD::SMED3", AMDGPUDTIntTernaryOp,
291 def AMDGPUumed3 : SDNode<"AMDGPUISD::UMED3", AMDGPUDTIntTernaryOp,
295 def AMDGPUfmed3 : SDNode<"AMDGPUISD::FMED3", SDTFPTernaryOp, []>;
297 def AMDGPUsendmsg : SDNode<"AMDGPUISD::SENDMSG",
298 SDTypeProfile<0, 1, [SDTCisInt<0>]>,
299 [SDNPHasChain, SDNPInGlue]>;
301 def AMDGPUsendmsghalt : SDNode<"AMDGPUISD::SENDMSGHALT",
302 SDTypeProfile<0, 1, [SDTCisInt<0>]>,
303 [SDNPHasChain, SDNPInGlue]>;
305 def AMDGPUinterp_mov : SDNode<"AMDGPUISD::INTERP_MOV",
306 SDTypeProfile<1, 3, [SDTCisFP<0>]>,
309 def AMDGPUinterp_p1 : SDNode<"AMDGPUISD::INTERP_P1",
310 SDTypeProfile<1, 3, [SDTCisFP<0>]>,
311 [SDNPInGlue, SDNPOutGlue]>;
313 def AMDGPUinterp_p2 : SDNode<"AMDGPUISD::INTERP_P2",
314 SDTypeProfile<1, 4, [SDTCisFP<0>]>,
318 def AMDGPUkill : SDNode<"AMDGPUISD::KILL", AMDGPUKillSDT,
319 [SDNPHasChain, SDNPSideEffect]>;
322 def AMDGPUExportOp : SDTypeProfile<0, 8, [
323 SDTCisInt<0>, // i8 tgt
324 SDTCisInt<1>, // i8 en
326 SDTCisSameAs<3, 2>, // f32 src1
327 SDTCisSameAs<4, 2>, // f32 src2
328 SDTCisSameAs<5, 2>, // f32 src3
329 SDTCisInt<6>, // i1 compr
331 SDTCisInt<1> // i1 vm
335 def AMDGPUexport: SDNode<"AMDGPUISD::EXPORT", AMDGPUExportOp,
336 [SDNPHasChain, SDNPMayStore]>;
338 def AMDGPUexport_done: SDNode<"AMDGPUISD::EXPORT_DONE", AMDGPUExportOp,
339 [SDNPHasChain, SDNPMayLoad, SDNPMayStore]>;
342 def R600ExportOp : SDTypeProfile<0, 7, [SDTCisFP<0>, SDTCisInt<1>]>;
344 def R600_EXPORT: SDNode<"AMDGPUISD::R600_EXPORT", R600ExportOp,
345 [SDNPHasChain, SDNPSideEffect]>;
347 //===----------------------------------------------------------------------===//
348 // Flow Control Profile Types
349 //===----------------------------------------------------------------------===//
350 // Branch instruction where second and third are basic blocks
351 def SDTIL_BRCond : SDTypeProfile<0, 2, [
355 //===----------------------------------------------------------------------===//
356 // Flow Control DAG Nodes
357 //===----------------------------------------------------------------------===//
358 def IL_brcond : SDNode<"AMDGPUISD::BRANCH_COND", SDTIL_BRCond, [SDNPHasChain]>;
360 //===----------------------------------------------------------------------===//
361 // Call/Return DAG Nodes
362 //===----------------------------------------------------------------------===//
363 def AMDGPUendpgm : SDNode<"AMDGPUISD::ENDPGM", SDTNone,
364 [SDNPHasChain, SDNPOptInGlue]>;
366 def AMDGPUreturn_to_epilog : SDNode<"AMDGPUISD::RETURN_TO_EPILOG", SDTNone,
367 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]>;
369 def AMDGPUret_flag : SDNode<"AMDGPUISD::RET_FLAG", SDTNone,
370 [SDNPHasChain, SDNPOptInGlue, SDNPVariadic]