1 | //=- WebAssemblyISelLowering.cpp - WebAssembly DAG Lowering Implementation -==// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | /// |
9 | /// \file |
10 | /// This file implements the WebAssemblyTargetLowering class. |
11 | /// |
12 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "WebAssemblyISelLowering.h" |
15 | #include "MCTargetDesc/WebAssemblyMCTargetDesc.h" |
16 | #include "Utils/WebAssemblyTypeUtilities.h" |
17 | #include "WebAssemblyMachineFunctionInfo.h" |
18 | #include "WebAssemblySubtarget.h" |
19 | #include "WebAssemblyTargetMachine.h" |
20 | #include "WebAssemblyUtilities.h" |
21 | #include "llvm/CodeGen/CallingConvLower.h" |
22 | #include "llvm/CodeGen/MachineFrameInfo.h" |
23 | #include "llvm/CodeGen/MachineFunctionPass.h" |
24 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
25 | #include "llvm/CodeGen/MachineJumpTableInfo.h" |
26 | #include "llvm/CodeGen/MachineModuleInfo.h" |
27 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
28 | #include "llvm/CodeGen/SelectionDAG.h" |
29 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
30 | #include "llvm/IR/DiagnosticInfo.h" |
31 | #include "llvm/IR/DiagnosticPrinter.h" |
32 | #include "llvm/IR/Function.h" |
33 | #include "llvm/IR/Intrinsics.h" |
34 | #include "llvm/IR/IntrinsicsWebAssembly.h" |
35 | #include "llvm/IR/PatternMatch.h" |
36 | #include "llvm/Support/Debug.h" |
37 | #include "llvm/Support/ErrorHandling.h" |
38 | #include "llvm/Support/KnownBits.h" |
39 | #include "llvm/Support/MathExtras.h" |
40 | #include "llvm/Support/raw_ostream.h" |
41 | #include "llvm/Target/TargetOptions.h" |
42 | using namespace llvm; |
43 | |
44 | #define DEBUG_TYPE "wasm-lower" |
45 | |
46 | WebAssemblyTargetLowering::WebAssemblyTargetLowering( |
47 | const TargetMachine &TM, const WebAssemblySubtarget &STI) |
48 | : TargetLowering(TM), Subtarget(&STI) { |
49 | auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32; |
50 | |
51 | // Booleans always contain 0 or 1. |
52 | setBooleanContents(ZeroOrOneBooleanContent); |
53 | // Except in SIMD vectors |
54 | setBooleanVectorContents(ZeroOrNegativeOneBooleanContent); |
55 | // We don't know the microarchitecture here, so just reduce register pressure. |
56 | setSchedulingPreference(Sched::RegPressure); |
57 | // Tell ISel that we have a stack pointer. |
58 | setStackPointerRegisterToSaveRestore( |
59 | Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32); |
60 | // Set up the register classes. |
61 | addRegisterClass(MVT::VT: i32, RC: &WebAssembly::I32RegClass); |
62 | addRegisterClass(MVT::VT: i64, RC: &WebAssembly::I64RegClass); |
63 | addRegisterClass(MVT::VT: f32, RC: &WebAssembly::F32RegClass); |
64 | addRegisterClass(MVT::VT: f64, RC: &WebAssembly::F64RegClass); |
65 | if (Subtarget->hasSIMD128()) { |
66 | addRegisterClass(MVT::VT: v16i8, RC: &WebAssembly::V128RegClass); |
67 | addRegisterClass(MVT::VT: v8i16, RC: &WebAssembly::V128RegClass); |
68 | addRegisterClass(MVT::VT: v4i32, RC: &WebAssembly::V128RegClass); |
69 | addRegisterClass(MVT::VT: v4f32, RC: &WebAssembly::V128RegClass); |
70 | addRegisterClass(MVT::VT: v2i64, RC: &WebAssembly::V128RegClass); |
71 | addRegisterClass(MVT::VT: v2f64, RC: &WebAssembly::V128RegClass); |
72 | } |
73 | if (Subtarget->hasReferenceTypes()) { |
74 | addRegisterClass(MVT::VT: externref, RC: &WebAssembly::EXTERNREFRegClass); |
75 | addRegisterClass(MVT::VT: funcref, RC: &WebAssembly::FUNCREFRegClass); |
76 | } |
77 | // Compute derived properties from the register classes. |
78 | computeRegisterProperties(Subtarget->getRegisterInfo()); |
79 | |
80 | // Transform loads and stores to pointers in address space 1 to loads and |
81 | // stores to WebAssembly global variables, outside linear memory. |
82 | for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) { |
83 | setOperationAction(ISD::LOAD, T, Custom); |
84 | setOperationAction(ISD::STORE, T, Custom); |
85 | } |
86 | if (Subtarget->hasSIMD128()) { |
87 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
88 | MVT::v2f64}) { |
89 | setOperationAction(ISD::LOAD, T, Custom); |
90 | setOperationAction(ISD::STORE, T, Custom); |
91 | } |
92 | } |
93 | if (Subtarget->hasReferenceTypes()) { |
94 | // We need custom load and store lowering for both externref, funcref and |
95 | // Other. The MVT::Other here represents tables of reference types. |
96 | for (auto T : {MVT::externref, MVT::funcref, MVT::Other}) { |
97 | setOperationAction(ISD::LOAD, T, Custom); |
98 | setOperationAction(ISD::STORE, T, Custom); |
99 | } |
100 | } |
101 | |
102 | setOperationAction(ISD::GlobalAddress, MVTPtr, Custom); |
103 | setOperationAction(ISD::GlobalTLSAddress, MVTPtr, Custom); |
104 | setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom); |
105 | setOperationAction(ISD::JumpTable, MVTPtr, Custom); |
106 | setOperationAction(ISD::BlockAddress, MVTPtr, Custom); |
107 | setOperationAction(ISD::BRIND, MVT::Other, Custom); |
108 | |
109 | // Take the default expansion for va_arg, va_copy, and va_end. There is no |
110 | // default action for va_start, so we do that custom. |
111 | setOperationAction(ISD::VASTART, MVT::Other, Custom); |
112 | setOperationAction(ISD::VAARG, MVT::Other, Expand); |
113 | setOperationAction(ISD::VACOPY, MVT::Other, Expand); |
114 | setOperationAction(ISD::VAEND, MVT::Other, Expand); |
115 | |
116 | for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) { |
117 | // Don't expand the floating-point types to constant pools. |
118 | setOperationAction(ISD::ConstantFP, T, Legal); |
119 | // Expand floating-point comparisons. |
120 | for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE, |
121 | ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE}) |
122 | setCondCodeAction(CC, T, Expand); |
123 | // Expand floating-point library function operators. |
124 | for (auto Op : |
125 | {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM, ISD::FMA}) |
126 | setOperationAction(Op, T, Expand); |
127 | // Note supported floating-point library function operators that otherwise |
128 | // default to expand. |
129 | for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT, |
130 | ISD::FRINT, ISD::FROUNDEVEN}) |
131 | setOperationAction(Op, T, Legal); |
132 | // Support minimum and maximum, which otherwise default to expand. |
133 | setOperationAction(ISD::FMINIMUM, T, Legal); |
134 | setOperationAction(ISD::FMAXIMUM, T, Legal); |
135 | // WebAssembly currently has no builtin f16 support. |
136 | setOperationAction(ISD::FP16_TO_FP, T, Expand); |
137 | setOperationAction(ISD::FP_TO_FP16, T, Expand); |
138 | setLoadExtAction(ISD::EXTLOAD, T, MVT::f16, Expand); |
139 | setTruncStoreAction(T, MVT::f16, Expand); |
140 | } |
141 | |
142 | // Expand unavailable integer operations. |
143 | for (auto Op : |
144 | {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU, |
145 | ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS, |
146 | ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) { |
147 | for (auto T : {MVT::i32, MVT::i64}) |
148 | setOperationAction(Op, T, Expand); |
149 | if (Subtarget->hasSIMD128()) |
150 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
151 | setOperationAction(Op, T, Expand); |
152 | } |
153 | |
154 | if (Subtarget->hasNontrappingFPToInt()) |
155 | for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) |
156 | for (auto T : {MVT::i32, MVT::i64}) |
157 | setOperationAction(Op, T, Custom); |
158 | |
159 | // SIMD-specific configuration |
160 | if (Subtarget->hasSIMD128()) { |
161 | // Combine vector mask reductions into alltrue/anytrue |
162 | setTargetDAGCombine(ISD::SETCC); |
163 | |
164 | // Convert vector to integer bitcasts to bitmask |
165 | setTargetDAGCombine(ISD::BITCAST); |
166 | |
167 | // Hoist bitcasts out of shuffles |
168 | setTargetDAGCombine(ISD::VECTOR_SHUFFLE); |
169 | |
170 | // Combine extends of extract_subvectors into widening ops |
171 | setTargetDAGCombine({ISD::SIGN_EXTEND, ISD::ZERO_EXTEND}); |
172 | |
173 | // Combine int_to_fp or fp_extend of extract_vectors and vice versa into |
174 | // conversions ops |
175 | setTargetDAGCombine({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_EXTEND, |
176 | ISD::EXTRACT_SUBVECTOR}); |
177 | |
178 | // Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa |
179 | // into conversion ops |
180 | setTargetDAGCombine({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT, |
181 | ISD::FP_ROUND, ISD::CONCAT_VECTORS}); |
182 | |
183 | setTargetDAGCombine(ISD::TRUNCATE); |
184 | |
185 | // Support saturating add for i8x16 and i16x8 |
186 | for (auto Op : {ISD::SADDSAT, ISD::UADDSAT}) |
187 | for (auto T : {MVT::v16i8, MVT::v8i16}) |
188 | setOperationAction(Op, T, Legal); |
189 | |
190 | // Support integer abs |
191 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
192 | setOperationAction(ISD::ABS, T, Legal); |
193 | |
194 | // Custom lower BUILD_VECTORs to minimize number of replace_lanes |
195 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
196 | MVT::v2f64}) |
197 | setOperationAction(ISD::BUILD_VECTOR, T, Custom); |
198 | |
199 | // We have custom shuffle lowering to expose the shuffle mask |
200 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
201 | MVT::v2f64}) |
202 | setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom); |
203 | |
204 | // Support splatting |
205 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
206 | MVT::v2f64}) |
207 | setOperationAction(ISD::SPLAT_VECTOR, T, Legal); |
208 | |
209 | // Custom lowering since wasm shifts must have a scalar shift amount |
210 | for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL}) |
211 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
212 | setOperationAction(Op, T, Custom); |
213 | |
214 | // Custom lower lane accesses to expand out variable indices |
215 | for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT}) |
216 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
217 | MVT::v2f64}) |
218 | setOperationAction(Op, T, Custom); |
219 | |
220 | // There is no i8x16.mul instruction |
221 | setOperationAction(ISD::MUL, MVT::v16i8, Expand); |
222 | |
223 | // There is no vector conditional select instruction |
224 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64, |
225 | MVT::v2f64}) |
226 | setOperationAction(ISD::SELECT_CC, T, Expand); |
227 | |
228 | // Expand integer operations supported for scalars but not SIMD |
229 | for (auto Op : |
230 | {ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR}) |
231 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
232 | setOperationAction(Op, T, Expand); |
233 | |
234 | // But we do have integer min and max operations |
235 | for (auto Op : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX}) |
236 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32}) |
237 | setOperationAction(Op, T, Legal); |
238 | |
239 | // And we have popcnt for i8x16. It can be used to expand ctlz/cttz. |
240 | setOperationAction(ISD::CTPOP, MVT::v16i8, Legal); |
241 | setOperationAction(ISD::CTLZ, MVT::v16i8, Expand); |
242 | setOperationAction(ISD::CTTZ, MVT::v16i8, Expand); |
243 | |
244 | // Custom lower bit counting operations for other types to scalarize them. |
245 | for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP}) |
246 | for (auto T : {MVT::v8i16, MVT::v4i32, MVT::v2i64}) |
247 | setOperationAction(Op, T, Custom); |
248 | |
249 | // Expand float operations supported for scalars but not SIMD |
250 | for (auto Op : {ISD::FCOPYSIGN, ISD::FLOG, ISD::FLOG2, ISD::FLOG10, |
251 | ISD::FEXP, ISD::FEXP2}) |
252 | for (auto T : {MVT::v4f32, MVT::v2f64}) |
253 | setOperationAction(Op, T, Expand); |
254 | |
255 | // Unsigned comparison operations are unavailable for i64x2 vectors. |
256 | for (auto CC : {ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE}) |
257 | setCondCodeAction(CC, MVT::v2i64, Custom); |
258 | |
259 | // 64x2 conversions are not in the spec |
260 | for (auto Op : |
261 | {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT}) |
262 | for (auto T : {MVT::v2i64, MVT::v2f64}) |
263 | setOperationAction(Op, T, Expand); |
264 | |
265 | // But saturating fp_to_int converstions are |
266 | for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) |
267 | setOperationAction(Op, MVT::v4i32, Custom); |
268 | |
269 | // Support vector extending |
270 | for (auto T : MVT::integer_fixedlen_vector_valuetypes()) { |
271 | setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, T, Custom); |
272 | setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, T, Custom); |
273 | } |
274 | } |
275 | |
276 | // As a special case, these operators use the type to mean the type to |
277 | // sign-extend from. |
278 | setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand); |
279 | if (!Subtarget->hasSignExt()) { |
280 | // Sign extends are legal only when extending a vector extract |
281 | auto Action = Subtarget->hasSIMD128() ? Custom : Expand; |
282 | for (auto T : {MVT::i8, MVT::i16, MVT::i32}) |
283 | setOperationAction(ISD::SIGN_EXTEND_INREG, T, Action); |
284 | } |
285 | for (auto T : MVT::integer_fixedlen_vector_valuetypes()) |
286 | setOperationAction(ISD::SIGN_EXTEND_INREG, T, Expand); |
287 | |
288 | // Dynamic stack allocation: use the default expansion. |
289 | setOperationAction(ISD::STACKSAVE, MVT::Other, Expand); |
290 | setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand); |
291 | setOperationAction(ISD::DYNAMIC_STACKALLOC, MVTPtr, Expand); |
292 | |
293 | setOperationAction(ISD::FrameIndex, MVT::i32, Custom); |
294 | setOperationAction(ISD::FrameIndex, MVT::i64, Custom); |
295 | setOperationAction(ISD::CopyToReg, MVT::Other, Custom); |
296 | |
297 | // Expand these forms; we pattern-match the forms that we can handle in isel. |
298 | for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) |
299 | for (auto Op : {ISD::BR_CC, ISD::SELECT_CC}) |
300 | setOperationAction(Op, T, Expand); |
301 | |
302 | // We have custom switch handling. |
303 | setOperationAction(ISD::BR_JT, MVT::Other, Custom); |
304 | |
305 | // WebAssembly doesn't have: |
306 | // - Floating-point extending loads. |
307 | // - Floating-point truncating stores. |
308 | // - i1 extending loads. |
309 | // - truncating SIMD stores and most extending loads |
310 | setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand); |
311 | setTruncStoreAction(MVT::f64, MVT::f32, Expand); |
312 | for (auto T : MVT::integer_valuetypes()) |
313 | for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD}) |
314 | setLoadExtAction(Ext, T, MVT::i1, Promote); |
315 | if (Subtarget->hasSIMD128()) { |
316 | for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32, |
317 | MVT::v2f64}) { |
318 | for (auto MemT : MVT::fixedlen_vector_valuetypes()) { |
319 | if (MVT(T) != MemT) { |
320 | setTruncStoreAction(T, MemT, Expand); |
321 | for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD}) |
322 | setLoadExtAction(Ext, T, MemT, Expand); |
323 | } |
324 | } |
325 | } |
326 | // But some vector extending loads are legal |
327 | for (auto Ext : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) { |
328 | setLoadExtAction(Ext, MVT::v8i16, MVT::v8i8, Legal); |
329 | setLoadExtAction(Ext, MVT::v4i32, MVT::v4i16, Legal); |
330 | setLoadExtAction(Ext, MVT::v2i64, MVT::v2i32, Legal); |
331 | } |
332 | setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Legal); |
333 | } |
334 | |
335 | // Don't do anything clever with build_pairs |
336 | setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand); |
337 | |
338 | // Trap lowers to wasm unreachable |
339 | setOperationAction(ISD::TRAP, MVT::Other, Legal); |
340 | setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal); |
341 | |
342 | // Exception handling intrinsics |
343 | setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom); |
344 | setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom); |
345 | setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom); |
346 | |
347 | setMaxAtomicSizeInBitsSupported(64); |
348 | |
349 | // Override the __gnu_f2h_ieee/__gnu_h2f_ieee names so that the f32 name is |
350 | // consistent with the f64 and f128 names. |
351 | setLibcallName(Call: RTLIB::FPEXT_F16_F32, Name: "__extendhfsf2" ); |
352 | setLibcallName(Call: RTLIB::FPROUND_F32_F16, Name: "__truncsfhf2" ); |
353 | |
354 | // Define the emscripten name for return address helper. |
355 | // TODO: when implementing other Wasm backends, make this generic or only do |
356 | // this on emscripten depending on what they end up doing. |
357 | setLibcallName(Call: RTLIB::RETURN_ADDRESS, Name: "emscripten_return_address" ); |
358 | |
359 | // Always convert switches to br_tables unless there is only one case, which |
360 | // is equivalent to a simple branch. This reduces code size for wasm, and we |
361 | // defer possible jump table optimizations to the VM. |
362 | setMinimumJumpTableEntries(2); |
363 | } |
364 | |
365 | MVT WebAssemblyTargetLowering::getPointerTy(const DataLayout &DL, |
366 | uint32_t AS) const { |
367 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF) |
368 | return MVT::externref; |
369 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF) |
370 | return MVT::funcref; |
371 | return TargetLowering::getPointerTy(DL, AS); |
372 | } |
373 | |
374 | MVT WebAssemblyTargetLowering::getPointerMemTy(const DataLayout &DL, |
375 | uint32_t AS) const { |
376 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF) |
377 | return MVT::externref; |
378 | if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF) |
379 | return MVT::funcref; |
380 | return TargetLowering::getPointerMemTy(DL, AS); |
381 | } |
382 | |
383 | TargetLowering::AtomicExpansionKind |
384 | WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const { |
385 | // We have wasm instructions for these |
386 | switch (AI->getOperation()) { |
387 | case AtomicRMWInst::Add: |
388 | case AtomicRMWInst::Sub: |
389 | case AtomicRMWInst::And: |
390 | case AtomicRMWInst::Or: |
391 | case AtomicRMWInst::Xor: |
392 | case AtomicRMWInst::Xchg: |
393 | return AtomicExpansionKind::None; |
394 | default: |
395 | break; |
396 | } |
397 | return AtomicExpansionKind::CmpXChg; |
398 | } |
399 | |
400 | bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const { |
401 | // Implementation copied from X86TargetLowering. |
402 | unsigned Opc = VecOp.getOpcode(); |
403 | |
404 | // Assume target opcodes can't be scalarized. |
405 | // TODO - do we have any exceptions? |
406 | if (Opc >= ISD::BUILTIN_OP_END) |
407 | return false; |
408 | |
409 | // If the vector op is not supported, try to convert to scalar. |
410 | EVT VecVT = VecOp.getValueType(); |
411 | if (!isOperationLegalOrCustomOrPromote(Op: Opc, VT: VecVT)) |
412 | return true; |
413 | |
414 | // If the vector op is supported, but the scalar op is not, the transform may |
415 | // not be worthwhile. |
416 | EVT ScalarVT = VecVT.getScalarType(); |
417 | return isOperationLegalOrCustomOrPromote(Op: Opc, VT: ScalarVT); |
418 | } |
419 | |
420 | FastISel *WebAssemblyTargetLowering::createFastISel( |
421 | FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo) const { |
422 | return WebAssembly::createFastISel(funcInfo&: FuncInfo, libInfo: LibInfo); |
423 | } |
424 | |
425 | MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/, |
426 | EVT VT) const { |
427 | unsigned BitWidth = NextPowerOf2(A: VT.getSizeInBits() - 1); |
428 | if (BitWidth > 1 && BitWidth < 8) |
429 | BitWidth = 8; |
430 | |
431 | if (BitWidth > 64) { |
432 | // The shift will be lowered to a libcall, and compiler-rt libcalls expect |
433 | // the count to be an i32. |
434 | BitWidth = 32; |
435 | assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) && |
436 | "32-bit shift counts ought to be enough for anyone" ); |
437 | } |
438 | |
439 | MVT Result = MVT::getIntegerVT(BitWidth); |
440 | assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE && |
441 | "Unable to represent scalar shift amount type" ); |
442 | return Result; |
443 | } |
444 | |
445 | // Lower an fp-to-int conversion operator from the LLVM opcode, which has an |
446 | // undefined result on invalid/overflow, to the WebAssembly opcode, which |
447 | // traps on invalid/overflow. |
448 | static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL, |
449 | MachineBasicBlock *BB, |
450 | const TargetInstrInfo &TII, |
451 | bool IsUnsigned, bool Int64, |
452 | bool Float64, unsigned LoweredOpcode) { |
453 | MachineRegisterInfo &MRI = BB->getParent()->getRegInfo(); |
454 | |
455 | Register OutReg = MI.getOperand(i: 0).getReg(); |
456 | Register InReg = MI.getOperand(i: 1).getReg(); |
457 | |
458 | unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32; |
459 | unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32; |
460 | unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32; |
461 | unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32; |
462 | unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32; |
463 | unsigned Eqz = WebAssembly::EQZ_I32; |
464 | unsigned And = WebAssembly::AND_I32; |
465 | int64_t Limit = Int64 ? INT64_MIN : INT32_MIN; |
466 | int64_t Substitute = IsUnsigned ? 0 : Limit; |
467 | double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit; |
468 | auto &Context = BB->getParent()->getFunction().getContext(); |
469 | Type *Ty = Float64 ? Type::getDoubleTy(C&: Context) : Type::getFloatTy(C&: Context); |
470 | |
471 | const BasicBlock *LLVMBB = BB->getBasicBlock(); |
472 | MachineFunction *F = BB->getParent(); |
473 | MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
474 | MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
475 | MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB); |
476 | |
477 | MachineFunction::iterator It = ++BB->getIterator(); |
478 | F->insert(MBBI: It, MBB: FalseMBB); |
479 | F->insert(MBBI: It, MBB: TrueMBB); |
480 | F->insert(MBBI: It, MBB: DoneMBB); |
481 | |
482 | // Transfer the remainder of BB and its successor edges to DoneMBB. |
483 | DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end()); |
484 | DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB); |
485 | |
486 | BB->addSuccessor(Succ: TrueMBB); |
487 | BB->addSuccessor(Succ: FalseMBB); |
488 | TrueMBB->addSuccessor(Succ: DoneMBB); |
489 | FalseMBB->addSuccessor(Succ: DoneMBB); |
490 | |
491 | unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg; |
492 | Tmp0 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
493 | Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
494 | CmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass); |
495 | EqzReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass); |
496 | FalseReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg)); |
497 | TrueReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg)); |
498 | |
499 | MI.eraseFromParent(); |
500 | // For signed numbers, we can do a single comparison to determine whether |
501 | // fabs(x) is within range. |
502 | if (IsUnsigned) { |
503 | Tmp0 = InReg; |
504 | } else { |
505 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Abs), DestReg: Tmp0).addReg(RegNo: InReg); |
506 | } |
507 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1) |
508 | .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: CmpVal))); |
509 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: LT), DestReg: CmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1); |
510 | |
511 | // For unsigned numbers, we have to do a separate comparison with zero. |
512 | if (IsUnsigned) { |
513 | Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg)); |
514 | Register SecondCmpReg = |
515 | MRI.createVirtualRegister(&WebAssembly::I32RegClass); |
516 | Register AndReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass); |
517 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1) |
518 | .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: 0.0))); |
519 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: GE), DestReg: SecondCmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1); |
520 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: And), DestReg: AndReg).addReg(RegNo: CmpReg).addReg(RegNo: SecondCmpReg); |
521 | CmpReg = AndReg; |
522 | } |
523 | |
524 | BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).addReg(RegNo: CmpReg); |
525 | |
526 | // Create the CFG diamond to select between doing the conversion or using |
527 | // the substitute value. |
528 | BuildMI(BB, DL, TII.get(WebAssembly::BR_IF)).addMBB(TrueMBB).addReg(EqzReg); |
529 | BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: LoweredOpcode), DestReg: FalseReg).addReg(RegNo: InReg); |
530 | BuildMI(FalseMBB, DL, TII.get(WebAssembly::BR)).addMBB(DoneMBB); |
531 | BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: IConst), DestReg: TrueReg).addImm(Val: Substitute); |
532 | BuildMI(BB&: *DoneMBB, I: DoneMBB->begin(), MIMD: DL, MCID: TII.get(Opcode: TargetOpcode::PHI), DestReg: OutReg) |
533 | .addReg(RegNo: FalseReg) |
534 | .addMBB(MBB: FalseMBB) |
535 | .addReg(RegNo: TrueReg) |
536 | .addMBB(MBB: TrueMBB); |
537 | |
538 | return DoneMBB; |
539 | } |
540 | |
541 | static MachineBasicBlock * |
542 | LowerCallResults(MachineInstr &CallResults, DebugLoc DL, MachineBasicBlock *BB, |
543 | const WebAssemblySubtarget *Subtarget, |
544 | const TargetInstrInfo &TII) { |
545 | MachineInstr &CallParams = *CallResults.getPrevNode(); |
546 | assert(CallParams.getOpcode() == WebAssembly::CALL_PARAMS); |
547 | assert(CallResults.getOpcode() == WebAssembly::CALL_RESULTS || |
548 | CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS); |
549 | |
550 | bool IsIndirect = |
551 | CallParams.getOperand(i: 0).isReg() || CallParams.getOperand(i: 0).isFI(); |
552 | bool IsRetCall = CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS; |
553 | |
554 | bool IsFuncrefCall = false; |
555 | if (IsIndirect && CallParams.getOperand(i: 0).isReg()) { |
556 | Register Reg = CallParams.getOperand(i: 0).getReg(); |
557 | const MachineFunction *MF = BB->getParent(); |
558 | const MachineRegisterInfo &MRI = MF->getRegInfo(); |
559 | const TargetRegisterClass *TRC = MRI.getRegClass(Reg); |
560 | IsFuncrefCall = (TRC == &WebAssembly::FUNCREFRegClass); |
561 | assert(!IsFuncrefCall || Subtarget->hasReferenceTypes()); |
562 | } |
563 | |
564 | unsigned CallOp; |
565 | if (IsIndirect && IsRetCall) { |
566 | CallOp = WebAssembly::RET_CALL_INDIRECT; |
567 | } else if (IsIndirect) { |
568 | CallOp = WebAssembly::CALL_INDIRECT; |
569 | } else if (IsRetCall) { |
570 | CallOp = WebAssembly::RET_CALL; |
571 | } else { |
572 | CallOp = WebAssembly::CALL; |
573 | } |
574 | |
575 | MachineFunction &MF = *BB->getParent(); |
576 | const MCInstrDesc &MCID = TII.get(Opcode: CallOp); |
577 | MachineInstrBuilder MIB(MF, MF.CreateMachineInstr(MCID, DL)); |
578 | |
579 | // See if we must truncate the function pointer. |
580 | // CALL_INDIRECT takes an i32, but in wasm64 we represent function pointers |
581 | // as 64-bit for uniformity with other pointer types. |
582 | // See also: WebAssemblyFastISel::selectCall |
583 | if (IsIndirect && MF.getSubtarget<WebAssemblySubtarget>().hasAddr64()) { |
584 | Register Reg32 = |
585 | MF.getRegInfo().createVirtualRegister(&WebAssembly::I32RegClass); |
586 | auto &FnPtr = CallParams.getOperand(i: 0); |
587 | BuildMI(*BB, CallResults.getIterator(), DL, |
588 | TII.get(WebAssembly::I32_WRAP_I64), Reg32) |
589 | .addReg(FnPtr.getReg()); |
590 | FnPtr.setReg(Reg32); |
591 | } |
592 | |
593 | // Move the function pointer to the end of the arguments for indirect calls |
594 | if (IsIndirect) { |
595 | auto FnPtr = CallParams.getOperand(i: 0); |
596 | CallParams.removeOperand(OpNo: 0); |
597 | |
598 | // For funcrefs, call_indirect is done through __funcref_call_table and the |
599 | // funcref is always installed in slot 0 of the table, therefore instead of |
600 | // having the function pointer added at the end of the params list, a zero |
601 | // (the index in |
602 | // __funcref_call_table is added). |
603 | if (IsFuncrefCall) { |
604 | Register RegZero = |
605 | MF.getRegInfo().createVirtualRegister(&WebAssembly::I32RegClass); |
606 | MachineInstrBuilder MIBC0 = |
607 | BuildMI(MF, DL, TII.get(WebAssembly::CONST_I32), RegZero).addImm(0); |
608 | |
609 | BB->insert(I: CallResults.getIterator(), M: MIBC0); |
610 | MachineInstrBuilder(MF, CallParams).addReg(RegNo: RegZero); |
611 | } else |
612 | CallParams.addOperand(Op: FnPtr); |
613 | } |
614 | |
615 | for (auto Def : CallResults.defs()) |
616 | MIB.add(MO: Def); |
617 | |
618 | if (IsIndirect) { |
619 | // Placeholder for the type index. |
620 | MIB.addImm(Val: 0); |
621 | // The table into which this call_indirect indexes. |
622 | MCSymbolWasm *Table = IsFuncrefCall |
623 | ? WebAssembly::getOrCreateFuncrefCallTableSymbol( |
624 | Ctx&: MF.getContext(), Subtarget) |
625 | : WebAssembly::getOrCreateFunctionTableSymbol( |
626 | Ctx&: MF.getContext(), Subtarget); |
627 | if (Subtarget->hasReferenceTypes()) { |
628 | MIB.addSym(Sym: Table); |
629 | } else { |
630 | // For the MVP there is at most one table whose number is 0, but we can't |
631 | // write a table symbol or issue relocations. Instead we just ensure the |
632 | // table is live and write a zero. |
633 | Table->setNoStrip(); |
634 | MIB.addImm(Val: 0); |
635 | } |
636 | } |
637 | |
638 | for (auto Use : CallParams.uses()) |
639 | MIB.add(MO: Use); |
640 | |
641 | BB->insert(I: CallResults.getIterator(), M: MIB); |
642 | CallParams.eraseFromParent(); |
643 | CallResults.eraseFromParent(); |
644 | |
645 | // If this is a funcref call, to avoid hidden GC roots, we need to clear the |
646 | // table slot with ref.null upon call_indirect return. |
647 | // |
648 | // This generates the following code, which comes right after a call_indirect |
649 | // of a funcref: |
650 | // |
651 | // i32.const 0 |
652 | // ref.null func |
653 | // table.set __funcref_call_table |
654 | if (IsIndirect && IsFuncrefCall) { |
655 | MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol( |
656 | Ctx&: MF.getContext(), Subtarget); |
657 | Register RegZero = |
658 | MF.getRegInfo().createVirtualRegister(&WebAssembly::I32RegClass); |
659 | MachineInstr *Const0 = |
660 | BuildMI(MF, DL, TII.get(WebAssembly::CONST_I32), RegZero).addImm(0); |
661 | BB->insertAfter(I: MIB.getInstr()->getIterator(), MI: Const0); |
662 | |
663 | Register RegFuncref = |
664 | MF.getRegInfo().createVirtualRegister(&WebAssembly::FUNCREFRegClass); |
665 | MachineInstr *RefNull = |
666 | BuildMI(MF, DL, TII.get(WebAssembly::REF_NULL_FUNCREF), RegFuncref); |
667 | BB->insertAfter(I: Const0->getIterator(), MI: RefNull); |
668 | |
669 | MachineInstr *TableSet = |
670 | BuildMI(MF, DL, TII.get(WebAssembly::TABLE_SET_FUNCREF)) |
671 | .addSym(Table) |
672 | .addReg(RegZero) |
673 | .addReg(RegFuncref); |
674 | BB->insertAfter(I: RefNull->getIterator(), MI: TableSet); |
675 | } |
676 | |
677 | return BB; |
678 | } |
679 | |
680 | MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter( |
681 | MachineInstr &MI, MachineBasicBlock *BB) const { |
682 | const TargetInstrInfo &TII = *Subtarget->getInstrInfo(); |
683 | DebugLoc DL = MI.getDebugLoc(); |
684 | |
685 | switch (MI.getOpcode()) { |
686 | default: |
687 | llvm_unreachable("Unexpected instr type to insert" ); |
688 | case WebAssembly::FP_TO_SINT_I32_F32: |
689 | return LowerFPToInt(MI, DL, BB, TII, false, false, false, |
690 | WebAssembly::I32_TRUNC_S_F32); |
691 | case WebAssembly::FP_TO_UINT_I32_F32: |
692 | return LowerFPToInt(MI, DL, BB, TII, true, false, false, |
693 | WebAssembly::I32_TRUNC_U_F32); |
694 | case WebAssembly::FP_TO_SINT_I64_F32: |
695 | return LowerFPToInt(MI, DL, BB, TII, false, true, false, |
696 | WebAssembly::I64_TRUNC_S_F32); |
697 | case WebAssembly::FP_TO_UINT_I64_F32: |
698 | return LowerFPToInt(MI, DL, BB, TII, true, true, false, |
699 | WebAssembly::I64_TRUNC_U_F32); |
700 | case WebAssembly::FP_TO_SINT_I32_F64: |
701 | return LowerFPToInt(MI, DL, BB, TII, false, false, true, |
702 | WebAssembly::I32_TRUNC_S_F64); |
703 | case WebAssembly::FP_TO_UINT_I32_F64: |
704 | return LowerFPToInt(MI, DL, BB, TII, true, false, true, |
705 | WebAssembly::I32_TRUNC_U_F64); |
706 | case WebAssembly::FP_TO_SINT_I64_F64: |
707 | return LowerFPToInt(MI, DL, BB, TII, false, true, true, |
708 | WebAssembly::I64_TRUNC_S_F64); |
709 | case WebAssembly::FP_TO_UINT_I64_F64: |
710 | return LowerFPToInt(MI, DL, BB, TII, true, true, true, |
711 | WebAssembly::I64_TRUNC_U_F64); |
712 | case WebAssembly::CALL_RESULTS: |
713 | case WebAssembly::RET_CALL_RESULTS: |
714 | return LowerCallResults(CallResults&: MI, DL, BB, Subtarget, TII); |
715 | } |
716 | } |
717 | |
718 | const char * |
719 | WebAssemblyTargetLowering::getTargetNodeName(unsigned Opcode) const { |
720 | switch (static_cast<WebAssemblyISD::NodeType>(Opcode)) { |
721 | case WebAssemblyISD::FIRST_NUMBER: |
722 | case WebAssemblyISD::FIRST_MEM_OPCODE: |
723 | break; |
724 | #define HANDLE_NODETYPE(NODE) \ |
725 | case WebAssemblyISD::NODE: \ |
726 | return "WebAssemblyISD::" #NODE; |
727 | #define HANDLE_MEM_NODETYPE(NODE) HANDLE_NODETYPE(NODE) |
728 | #include "WebAssemblyISD.def" |
729 | #undef HANDLE_MEM_NODETYPE |
730 | #undef HANDLE_NODETYPE |
731 | } |
732 | return nullptr; |
733 | } |
734 | |
735 | std::pair<unsigned, const TargetRegisterClass *> |
736 | WebAssemblyTargetLowering::getRegForInlineAsmConstraint( |
737 | const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const { |
738 | // First, see if this is a constraint that directly corresponds to a |
739 | // WebAssembly register class. |
740 | if (Constraint.size() == 1) { |
741 | switch (Constraint[0]) { |
742 | case 'r': |
743 | assert(VT != MVT::iPTR && "Pointer MVT not expected here" ); |
744 | if (Subtarget->hasSIMD128() && VT.isVector()) { |
745 | if (VT.getSizeInBits() == 128) |
746 | return std::make_pair(0U, &WebAssembly::V128RegClass); |
747 | } |
748 | if (VT.isInteger() && !VT.isVector()) { |
749 | if (VT.getSizeInBits() <= 32) |
750 | return std::make_pair(0U, &WebAssembly::I32RegClass); |
751 | if (VT.getSizeInBits() <= 64) |
752 | return std::make_pair(0U, &WebAssembly::I64RegClass); |
753 | } |
754 | if (VT.isFloatingPoint() && !VT.isVector()) { |
755 | switch (VT.getSizeInBits()) { |
756 | case 32: |
757 | return std::make_pair(0U, &WebAssembly::F32RegClass); |
758 | case 64: |
759 | return std::make_pair(0U, &WebAssembly::F64RegClass); |
760 | default: |
761 | break; |
762 | } |
763 | } |
764 | break; |
765 | default: |
766 | break; |
767 | } |
768 | } |
769 | |
770 | return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT); |
771 | } |
772 | |
773 | bool WebAssemblyTargetLowering::isCheapToSpeculateCttz(Type *Ty) const { |
774 | // Assume ctz is a relatively cheap operation. |
775 | return true; |
776 | } |
777 | |
778 | bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const { |
779 | // Assume clz is a relatively cheap operation. |
780 | return true; |
781 | } |
782 | |
783 | bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL, |
784 | const AddrMode &AM, |
785 | Type *Ty, unsigned AS, |
786 | Instruction *I) const { |
787 | // WebAssembly offsets are added as unsigned without wrapping. The |
788 | // isLegalAddressingMode gives us no way to determine if wrapping could be |
789 | // happening, so we approximate this by accepting only non-negative offsets. |
790 | if (AM.BaseOffs < 0) |
791 | return false; |
792 | |
793 | // WebAssembly has no scale register operands. |
794 | if (AM.Scale != 0) |
795 | return false; |
796 | |
797 | // Everything else is legal. |
798 | return true; |
799 | } |
800 | |
801 | bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses( |
802 | EVT /*VT*/, unsigned /*AddrSpace*/, Align /*Align*/, |
803 | MachineMemOperand::Flags /*Flags*/, unsigned *Fast) const { |
804 | // WebAssembly supports unaligned accesses, though it should be declared |
805 | // with the p2align attribute on loads and stores which do so, and there |
806 | // may be a performance impact. We tell LLVM they're "fast" because |
807 | // for the kinds of things that LLVM uses this for (merging adjacent stores |
808 | // of constants, etc.), WebAssembly implementations will either want the |
809 | // unaligned access or they'll split anyway. |
810 | if (Fast) |
811 | *Fast = 1; |
812 | return true; |
813 | } |
814 | |
815 | bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT, |
816 | AttributeList Attr) const { |
817 | // The current thinking is that wasm engines will perform this optimization, |
818 | // so we can save on code size. |
819 | return true; |
820 | } |
821 | |
822 | bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const { |
823 | EVT ExtT = ExtVal.getValueType(); |
824 | EVT MemT = cast<LoadSDNode>(Val: ExtVal->getOperand(Num: 0))->getValueType(ResNo: 0); |
825 | return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) || |
826 | (ExtT == MVT::v4i32 && MemT == MVT::v4i16) || |
827 | (ExtT == MVT::v2i64 && MemT == MVT::v2i32); |
828 | } |
829 | |
830 | bool WebAssemblyTargetLowering::isOffsetFoldingLegal( |
831 | const GlobalAddressSDNode *GA) const { |
832 | // Wasm doesn't support function addresses with offsets |
833 | const GlobalValue *GV = GA->getGlobal(); |
834 | return isa<Function>(Val: GV) ? false : TargetLowering::isOffsetFoldingLegal(GA); |
835 | } |
836 | |
837 | bool WebAssemblyTargetLowering::shouldSinkOperands( |
838 | Instruction *I, SmallVectorImpl<Use *> &Ops) const { |
839 | using namespace llvm::PatternMatch; |
840 | |
841 | if (!I->getType()->isVectorTy() || !I->isShift()) |
842 | return false; |
843 | |
844 | Value *V = I->getOperand(i: 1); |
845 | // We dont need to sink constant splat. |
846 | if (dyn_cast<Constant>(Val: V)) |
847 | return false; |
848 | |
849 | if (match(V, P: m_Shuffle(v1: m_InsertElt(Val: m_Value(), Elt: m_Value(), Idx: m_ZeroInt()), |
850 | v2: m_Value(), mask: m_ZeroMask()))) { |
851 | // Sink insert |
852 | Ops.push_back(Elt: &cast<Instruction>(Val: V)->getOperandUse(i: 0)); |
853 | // Sink shuffle |
854 | Ops.push_back(Elt: &I->getOperandUse(i: 1)); |
855 | return true; |
856 | } |
857 | |
858 | return false; |
859 | } |
860 | |
861 | EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL, |
862 | LLVMContext &C, |
863 | EVT VT) const { |
864 | if (VT.isVector()) |
865 | return VT.changeVectorElementTypeToInteger(); |
866 | |
867 | // So far, all branch instructions in Wasm take an I32 condition. |
868 | // The default TargetLowering::getSetCCResultType returns the pointer size, |
869 | // which would be useful to reduce instruction counts when testing |
870 | // against 64-bit pointers/values if at some point Wasm supports that. |
871 | return EVT::getIntegerVT(Context&: C, BitWidth: 32); |
872 | } |
873 | |
874 | bool WebAssemblyTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info, |
875 | const CallInst &I, |
876 | MachineFunction &MF, |
877 | unsigned Intrinsic) const { |
878 | switch (Intrinsic) { |
879 | case Intrinsic::wasm_memory_atomic_notify: |
880 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
881 | Info.memVT = MVT::i32; |
882 | Info.ptrVal = I.getArgOperand(i: 0); |
883 | Info.offset = 0; |
884 | Info.align = Align(4); |
885 | // atomic.notify instruction does not really load the memory specified with |
886 | // this argument, but MachineMemOperand should either be load or store, so |
887 | // we set this to a load. |
888 | // FIXME Volatile isn't really correct, but currently all LLVM atomic |
889 | // instructions are treated as volatiles in the backend, so we should be |
890 | // consistent. The same applies for wasm_atomic_wait intrinsics too. |
891 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
892 | return true; |
893 | case Intrinsic::wasm_memory_atomic_wait32: |
894 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
895 | Info.memVT = MVT::i32; |
896 | Info.ptrVal = I.getArgOperand(i: 0); |
897 | Info.offset = 0; |
898 | Info.align = Align(4); |
899 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
900 | return true; |
901 | case Intrinsic::wasm_memory_atomic_wait64: |
902 | Info.opc = ISD::INTRINSIC_W_CHAIN; |
903 | Info.memVT = MVT::i64; |
904 | Info.ptrVal = I.getArgOperand(i: 0); |
905 | Info.offset = 0; |
906 | Info.align = Align(8); |
907 | Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad; |
908 | return true; |
909 | default: |
910 | return false; |
911 | } |
912 | } |
913 | |
914 | void WebAssemblyTargetLowering::computeKnownBitsForTargetNode( |
915 | const SDValue Op, KnownBits &Known, const APInt &DemandedElts, |
916 | const SelectionDAG &DAG, unsigned Depth) const { |
917 | switch (Op.getOpcode()) { |
918 | default: |
919 | break; |
920 | case ISD::INTRINSIC_WO_CHAIN: { |
921 | unsigned IntNo = Op.getConstantOperandVal(i: 0); |
922 | switch (IntNo) { |
923 | default: |
924 | break; |
925 | case Intrinsic::wasm_bitmask: { |
926 | unsigned BitWidth = Known.getBitWidth(); |
927 | EVT VT = Op.getOperand(i: 1).getSimpleValueType(); |
928 | unsigned PossibleBits = VT.getVectorNumElements(); |
929 | APInt ZeroMask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - PossibleBits); |
930 | Known.Zero |= ZeroMask; |
931 | break; |
932 | } |
933 | } |
934 | } |
935 | } |
936 | } |
937 | |
938 | TargetLoweringBase::LegalizeTypeAction |
939 | WebAssemblyTargetLowering::getPreferredVectorAction(MVT VT) const { |
940 | if (VT.isFixedLengthVector()) { |
941 | MVT EltVT = VT.getVectorElementType(); |
942 | // We have legal vector types with these lane types, so widening the |
943 | // vector would let us use some of the lanes directly without having to |
944 | // extend or truncate values. |
945 | if (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 || |
946 | EltVT == MVT::i64 || EltVT == MVT::f32 || EltVT == MVT::f64) |
947 | return TypeWidenVector; |
948 | } |
949 | |
950 | return TargetLoweringBase::getPreferredVectorAction(VT); |
951 | } |
952 | |
953 | bool WebAssemblyTargetLowering::shouldSimplifyDemandedVectorElts( |
954 | SDValue Op, const TargetLoweringOpt &TLO) const { |
955 | // ISel process runs DAGCombiner after legalization; this step is called |
956 | // SelectionDAG optimization phase. This post-legalization combining process |
957 | // runs DAGCombiner on each node, and if there was a change to be made, |
958 | // re-runs legalization again on it and its user nodes to make sure |
959 | // everythiing is in a legalized state. |
960 | // |
961 | // The legalization calls lowering routines, and we do our custom lowering for |
962 | // build_vectors (LowerBUILD_VECTOR), which converts undef vector elements |
963 | // into zeros. But there is a set of routines in DAGCombiner that turns unused |
964 | // (= not demanded) nodes into undef, among which SimplifyDemandedVectorElts |
965 | // turns unused vector elements into undefs. But this routine does not work |
966 | // with our custom LowerBUILD_VECTOR, which turns undefs into zeros. This |
967 | // combination can result in a infinite loop, in which undefs are converted to |
968 | // zeros in legalization and back to undefs in combining. |
969 | // |
970 | // So after DAG is legalized, we prevent SimplifyDemandedVectorElts from |
971 | // running for build_vectors. |
972 | if (Op.getOpcode() == ISD::BUILD_VECTOR && TLO.LegalOps && TLO.LegalTys) |
973 | return false; |
974 | return true; |
975 | } |
976 | |
977 | //===----------------------------------------------------------------------===// |
978 | // WebAssembly Lowering private implementation. |
979 | //===----------------------------------------------------------------------===// |
980 | |
981 | //===----------------------------------------------------------------------===// |
982 | // Lowering Code |
983 | //===----------------------------------------------------------------------===// |
984 | |
985 | static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) { |
986 | MachineFunction &MF = DAG.getMachineFunction(); |
987 | DAG.getContext()->diagnose( |
988 | DI: DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc())); |
989 | } |
990 | |
991 | // Test whether the given calling convention is supported. |
992 | static bool callingConvSupported(CallingConv::ID CallConv) { |
993 | // We currently support the language-independent target-independent |
994 | // conventions. We don't yet have a way to annotate calls with properties like |
995 | // "cold", and we don't have any call-clobbered registers, so these are mostly |
996 | // all handled the same. |
997 | return CallConv == CallingConv::C || CallConv == CallingConv::Fast || |
998 | CallConv == CallingConv::Cold || |
999 | CallConv == CallingConv::PreserveMost || |
1000 | CallConv == CallingConv::PreserveAll || |
1001 | CallConv == CallingConv::CXX_FAST_TLS || |
1002 | CallConv == CallingConv::WASM_EmscriptenInvoke || |
1003 | CallConv == CallingConv::Swift; |
1004 | } |
1005 | |
1006 | SDValue |
1007 | WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI, |
1008 | SmallVectorImpl<SDValue> &InVals) const { |
1009 | SelectionDAG &DAG = CLI.DAG; |
1010 | SDLoc DL = CLI.DL; |
1011 | SDValue Chain = CLI.Chain; |
1012 | SDValue Callee = CLI.Callee; |
1013 | MachineFunction &MF = DAG.getMachineFunction(); |
1014 | auto Layout = MF.getDataLayout(); |
1015 | |
1016 | CallingConv::ID CallConv = CLI.CallConv; |
1017 | if (!callingConvSupported(CallConv)) |
1018 | fail(DL, DAG, |
1019 | Msg: "WebAssembly doesn't support language-specific or target-specific " |
1020 | "calling conventions yet" ); |
1021 | if (CLI.IsPatchPoint) |
1022 | fail(DL, DAG, Msg: "WebAssembly doesn't support patch point yet" ); |
1023 | |
1024 | if (CLI.IsTailCall) { |
1025 | auto NoTail = [&](const char *Msg) { |
1026 | if (CLI.CB && CLI.CB->isMustTailCall()) |
1027 | fail(DL, DAG, Msg); |
1028 | CLI.IsTailCall = false; |
1029 | }; |
1030 | |
1031 | if (!Subtarget->hasTailCall()) |
1032 | NoTail("WebAssembly 'tail-call' feature not enabled" ); |
1033 | |
1034 | // Varargs calls cannot be tail calls because the buffer is on the stack |
1035 | if (CLI.IsVarArg) |
1036 | NoTail("WebAssembly does not support varargs tail calls" ); |
1037 | |
1038 | // Do not tail call unless caller and callee return types match |
1039 | const Function &F = MF.getFunction(); |
1040 | const TargetMachine &TM = getTargetMachine(); |
1041 | Type *RetTy = F.getReturnType(); |
1042 | SmallVector<MVT, 4> CallerRetTys; |
1043 | SmallVector<MVT, 4> CalleeRetTys; |
1044 | computeLegalValueVTs(F, TM, Ty: RetTy, ValueVTs&: CallerRetTys); |
1045 | computeLegalValueVTs(F, TM, Ty: CLI.RetTy, ValueVTs&: CalleeRetTys); |
1046 | bool TypesMatch = CallerRetTys.size() == CalleeRetTys.size() && |
1047 | std::equal(first1: CallerRetTys.begin(), last1: CallerRetTys.end(), |
1048 | first2: CalleeRetTys.begin()); |
1049 | if (!TypesMatch) |
1050 | NoTail("WebAssembly tail call requires caller and callee return types to " |
1051 | "match" ); |
1052 | |
1053 | // If pointers to local stack values are passed, we cannot tail call |
1054 | if (CLI.CB) { |
1055 | for (auto &Arg : CLI.CB->args()) { |
1056 | Value *Val = Arg.get(); |
1057 | // Trace the value back through pointer operations |
1058 | while (true) { |
1059 | Value *Src = Val->stripPointerCastsAndAliases(); |
1060 | if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: Src)) |
1061 | Src = GEP->getPointerOperand(); |
1062 | if (Val == Src) |
1063 | break; |
1064 | Val = Src; |
1065 | } |
1066 | if (isa<AllocaInst>(Val)) { |
1067 | NoTail( |
1068 | "WebAssembly does not support tail calling with stack arguments" ); |
1069 | break; |
1070 | } |
1071 | } |
1072 | } |
1073 | } |
1074 | |
1075 | SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins; |
1076 | SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs; |
1077 | SmallVectorImpl<SDValue> &OutVals = CLI.OutVals; |
1078 | |
1079 | // The generic code may have added an sret argument. If we're lowering an |
1080 | // invoke function, the ABI requires that the function pointer be the first |
1081 | // argument, so we may have to swap the arguments. |
1082 | if (CallConv == CallingConv::WASM_EmscriptenInvoke && Outs.size() >= 2 && |
1083 | Outs[0].Flags.isSRet()) { |
1084 | std::swap(a&: Outs[0], b&: Outs[1]); |
1085 | std::swap(a&: OutVals[0], b&: OutVals[1]); |
1086 | } |
1087 | |
1088 | bool HasSwiftSelfArg = false; |
1089 | bool HasSwiftErrorArg = false; |
1090 | unsigned NumFixedArgs = 0; |
1091 | for (unsigned I = 0; I < Outs.size(); ++I) { |
1092 | const ISD::OutputArg &Out = Outs[I]; |
1093 | SDValue &OutVal = OutVals[I]; |
1094 | HasSwiftSelfArg |= Out.Flags.isSwiftSelf(); |
1095 | HasSwiftErrorArg |= Out.Flags.isSwiftError(); |
1096 | if (Out.Flags.isNest()) |
1097 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments" ); |
1098 | if (Out.Flags.isInAlloca()) |
1099 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments" ); |
1100 | if (Out.Flags.isInConsecutiveRegs()) |
1101 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments" ); |
1102 | if (Out.Flags.isInConsecutiveRegsLast()) |
1103 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments" ); |
1104 | if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) { |
1105 | auto &MFI = MF.getFrameInfo(); |
1106 | int FI = MFI.CreateStackObject(Size: Out.Flags.getByValSize(), |
1107 | Alignment: Out.Flags.getNonZeroByValAlign(), |
1108 | /*isSS=*/isSpillSlot: false); |
1109 | SDValue SizeNode = |
1110 | DAG.getConstant(Out.Flags.getByValSize(), DL, MVT::i32); |
1111 | SDValue FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout)); |
1112 | Chain = DAG.getMemcpy( |
1113 | Chain, dl: DL, Dst: FINode, Src: OutVal, Size: SizeNode, Alignment: Out.Flags.getNonZeroByValAlign(), |
1114 | /*isVolatile*/ isVol: false, /*AlwaysInline=*/false, |
1115 | /*isTailCall*/ false, DstPtrInfo: MachinePointerInfo(), SrcPtrInfo: MachinePointerInfo()); |
1116 | OutVal = FINode; |
1117 | } |
1118 | // Count the number of fixed args *after* legalization. |
1119 | NumFixedArgs += Out.IsFixed; |
1120 | } |
1121 | |
1122 | bool IsVarArg = CLI.IsVarArg; |
1123 | auto PtrVT = getPointerTy(DL: Layout); |
1124 | |
1125 | // For swiftcc, emit additional swiftself and swifterror arguments |
1126 | // if there aren't. These additional arguments are also added for callee |
1127 | // signature They are necessary to match callee and caller signature for |
1128 | // indirect call. |
1129 | if (CallConv == CallingConv::Swift) { |
1130 | if (!HasSwiftSelfArg) { |
1131 | NumFixedArgs++; |
1132 | ISD::OutputArg Arg; |
1133 | Arg.Flags.setSwiftSelf(); |
1134 | CLI.Outs.push_back(Elt: Arg); |
1135 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
1136 | CLI.OutVals.push_back(Elt: ArgVal); |
1137 | } |
1138 | if (!HasSwiftErrorArg) { |
1139 | NumFixedArgs++; |
1140 | ISD::OutputArg Arg; |
1141 | Arg.Flags.setSwiftError(); |
1142 | CLI.Outs.push_back(Elt: Arg); |
1143 | SDValue ArgVal = DAG.getUNDEF(VT: PtrVT); |
1144 | CLI.OutVals.push_back(Elt: ArgVal); |
1145 | } |
1146 | } |
1147 | |
1148 | // Analyze operands of the call, assigning locations to each operand. |
1149 | SmallVector<CCValAssign, 16> ArgLocs; |
1150 | CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext()); |
1151 | |
1152 | if (IsVarArg) { |
1153 | // Outgoing non-fixed arguments are placed in a buffer. First |
1154 | // compute their offsets and the total amount of buffer space needed. |
1155 | for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) { |
1156 | const ISD::OutputArg &Out = Outs[I]; |
1157 | SDValue &Arg = OutVals[I]; |
1158 | EVT VT = Arg.getValueType(); |
1159 | assert(VT != MVT::iPTR && "Legalized args should be concrete" ); |
1160 | Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext()); |
1161 | Align Alignment = |
1162 | std::max(a: Out.Flags.getNonZeroOrigAlign(), b: Layout.getABITypeAlign(Ty)); |
1163 | unsigned Offset = |
1164 | CCInfo.AllocateStack(Size: Layout.getTypeAllocSize(Ty), Alignment); |
1165 | CCInfo.addLoc(V: CCValAssign::getMem(ValNo: ArgLocs.size(), ValVT: VT.getSimpleVT(), |
1166 | Offset, LocVT: VT.getSimpleVT(), |
1167 | HTP: CCValAssign::Full)); |
1168 | } |
1169 | } |
1170 | |
1171 | unsigned NumBytes = CCInfo.getAlignedCallFrameSize(); |
1172 | |
1173 | SDValue FINode; |
1174 | if (IsVarArg && NumBytes) { |
1175 | // For non-fixed arguments, next emit stores to store the argument values |
1176 | // to the stack buffer at the offsets computed above. |
1177 | int FI = MF.getFrameInfo().CreateStackObject(Size: NumBytes, |
1178 | Alignment: Layout.getStackAlignment(), |
1179 | /*isSS=*/isSpillSlot: false); |
1180 | unsigned ValNo = 0; |
1181 | SmallVector<SDValue, 8> Chains; |
1182 | for (SDValue Arg : drop_begin(RangeOrContainer&: OutVals, N: NumFixedArgs)) { |
1183 | assert(ArgLocs[ValNo].getValNo() == ValNo && |
1184 | "ArgLocs should remain in order and only hold varargs args" ); |
1185 | unsigned Offset = ArgLocs[ValNo++].getLocMemOffset(); |
1186 | FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout)); |
1187 | SDValue Add = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: FINode, |
1188 | N2: DAG.getConstant(Val: Offset, DL, VT: PtrVT)); |
1189 | Chains.push_back( |
1190 | Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: Add, |
1191 | PtrInfo: MachinePointerInfo::getFixedStack(MF, FI, Offset))); |
1192 | } |
1193 | if (!Chains.empty()) |
1194 | Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains); |
1195 | } else if (IsVarArg) { |
1196 | FINode = DAG.getIntPtrConstant(Val: 0, DL); |
1197 | } |
1198 | |
1199 | if (Callee->getOpcode() == ISD::GlobalAddress) { |
1200 | // If the callee is a GlobalAddress node (quite common, every direct call |
1201 | // is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress |
1202 | // doesn't at MO_GOT which is not needed for direct calls. |
1203 | GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Callee); |
1204 | Callee = DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, |
1205 | VT: getPointerTy(DL: DAG.getDataLayout()), |
1206 | offset: GA->getOffset()); |
1207 | Callee = DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, |
1208 | VT: getPointerTy(DL: DAG.getDataLayout()), Operand: Callee); |
1209 | } |
1210 | |
1211 | // Compute the operands for the CALLn node. |
1212 | SmallVector<SDValue, 16> Ops; |
1213 | Ops.push_back(Elt: Chain); |
1214 | Ops.push_back(Elt: Callee); |
1215 | |
1216 | // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs |
1217 | // isn't reliable. |
1218 | Ops.append(in_start: OutVals.begin(), |
1219 | in_end: IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end()); |
1220 | // Add a pointer to the vararg buffer. |
1221 | if (IsVarArg) |
1222 | Ops.push_back(Elt: FINode); |
1223 | |
1224 | SmallVector<EVT, 8> InTys; |
1225 | for (const auto &In : Ins) { |
1226 | assert(!In.Flags.isByVal() && "byval is not valid for return values" ); |
1227 | assert(!In.Flags.isNest() && "nest is not valid for return values" ); |
1228 | if (In.Flags.isInAlloca()) |
1229 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca return values" ); |
1230 | if (In.Flags.isInConsecutiveRegs()) |
1231 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs return values" ); |
1232 | if (In.Flags.isInConsecutiveRegsLast()) |
1233 | fail(DL, DAG, |
1234 | Msg: "WebAssembly hasn't implemented cons regs last return values" ); |
1235 | // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in |
1236 | // registers. |
1237 | InTys.push_back(Elt: In.VT); |
1238 | } |
1239 | |
1240 | // Lastly, if this is a call to a funcref we need to add an instruction |
1241 | // table.set to the chain and transform the call. |
1242 | if (CLI.CB && WebAssembly::isWebAssemblyFuncrefType( |
1243 | Ty: CLI.CB->getCalledOperand()->getType())) { |
1244 | // In the absence of function references proposal where a funcref call is |
1245 | // lowered to call_ref, using reference types we generate a table.set to set |
1246 | // the funcref to a special table used solely for this purpose, followed by |
1247 | // a call_indirect. Here we just generate the table set, and return the |
1248 | // SDValue of the table.set so that LowerCall can finalize the lowering by |
1249 | // generating the call_indirect. |
1250 | SDValue Chain = Ops[0]; |
1251 | |
1252 | MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol( |
1253 | Ctx&: MF.getContext(), Subtarget); |
1254 | SDValue Sym = DAG.getMCSymbol(Sym: Table, VT: PtrVT); |
1255 | SDValue TableSlot = DAG.getConstant(0, DL, MVT::i32); |
1256 | SDValue TableSetOps[] = {Chain, Sym, TableSlot, Callee}; |
1257 | SDValue TableSet = DAG.getMemIntrinsicNode( |
1258 | WebAssemblyISD::TABLE_SET, DL, DAG.getVTList(MVT::Other), TableSetOps, |
1259 | MVT::funcref, |
1260 | // Machine Mem Operand args |
1261 | MachinePointerInfo( |
1262 | WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF), |
1263 | CLI.CB->getCalledOperand()->getPointerAlignment(DAG.getDataLayout()), |
1264 | MachineMemOperand::MOStore); |
1265 | |
1266 | Ops[0] = TableSet; // The new chain is the TableSet itself |
1267 | } |
1268 | |
1269 | if (CLI.IsTailCall) { |
1270 | // ret_calls do not return values to the current frame |
1271 | SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue); |
1272 | return DAG.getNode(Opcode: WebAssemblyISD::RET_CALL, DL, VTList: NodeTys, Ops); |
1273 | } |
1274 | |
1275 | InTys.push_back(MVT::Other); |
1276 | SDVTList InTyList = DAG.getVTList(VTs: InTys); |
1277 | SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::CALL, DL, VTList: InTyList, Ops); |
1278 | |
1279 | for (size_t I = 0; I < Ins.size(); ++I) |
1280 | InVals.push_back(Elt: Res.getValue(R: I)); |
1281 | |
1282 | // Return the chain |
1283 | return Res.getValue(R: Ins.size()); |
1284 | } |
1285 | |
1286 | bool WebAssemblyTargetLowering::CanLowerReturn( |
1287 | CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/, |
1288 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1289 | LLVMContext & /*Context*/) const { |
1290 | // WebAssembly can only handle returning tuples with multivalue enabled |
1291 | return WebAssembly::canLowerReturn(ResultSize: Outs.size(), Subtarget); |
1292 | } |
1293 | |
1294 | SDValue WebAssemblyTargetLowering::LowerReturn( |
1295 | SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/, |
1296 | const SmallVectorImpl<ISD::OutputArg> &Outs, |
1297 | const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL, |
1298 | SelectionDAG &DAG) const { |
1299 | assert(WebAssembly::canLowerReturn(Outs.size(), Subtarget) && |
1300 | "MVP WebAssembly can only return up to one value" ); |
1301 | if (!callingConvSupported(CallConv)) |
1302 | fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions" ); |
1303 | |
1304 | SmallVector<SDValue, 4> RetOps(1, Chain); |
1305 | RetOps.append(in_start: OutVals.begin(), in_end: OutVals.end()); |
1306 | Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps); |
1307 | |
1308 | // Record the number and types of the return values. |
1309 | for (const ISD::OutputArg &Out : Outs) { |
1310 | assert(!Out.Flags.isByVal() && "byval is not valid for return values" ); |
1311 | assert(!Out.Flags.isNest() && "nest is not valid for return values" ); |
1312 | assert(Out.IsFixed && "non-fixed return value is not valid" ); |
1313 | if (Out.Flags.isInAlloca()) |
1314 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca results" ); |
1315 | if (Out.Flags.isInConsecutiveRegs()) |
1316 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs results" ); |
1317 | if (Out.Flags.isInConsecutiveRegsLast()) |
1318 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last results" ); |
1319 | } |
1320 | |
1321 | return Chain; |
1322 | } |
1323 | |
1324 | SDValue WebAssemblyTargetLowering::LowerFormalArguments( |
1325 | SDValue Chain, CallingConv::ID CallConv, bool IsVarArg, |
1326 | const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL, |
1327 | SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const { |
1328 | if (!callingConvSupported(CallConv)) |
1329 | fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions" ); |
1330 | |
1331 | MachineFunction &MF = DAG.getMachineFunction(); |
1332 | auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>(); |
1333 | |
1334 | // Set up the incoming ARGUMENTS value, which serves to represent the liveness |
1335 | // of the incoming values before they're represented by virtual registers. |
1336 | MF.getRegInfo().addLiveIn(WebAssembly::ARGUMENTS); |
1337 | |
1338 | bool HasSwiftErrorArg = false; |
1339 | bool HasSwiftSelfArg = false; |
1340 | for (const ISD::InputArg &In : Ins) { |
1341 | HasSwiftSelfArg |= In.Flags.isSwiftSelf(); |
1342 | HasSwiftErrorArg |= In.Flags.isSwiftError(); |
1343 | if (In.Flags.isInAlloca()) |
1344 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments" ); |
1345 | if (In.Flags.isNest()) |
1346 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments" ); |
1347 | if (In.Flags.isInConsecutiveRegs()) |
1348 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments" ); |
1349 | if (In.Flags.isInConsecutiveRegsLast()) |
1350 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments" ); |
1351 | // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in |
1352 | // registers. |
1353 | InVals.push_back(In.Used ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT, |
1354 | DAG.getTargetConstant(InVals.size(), |
1355 | DL, MVT::i32)) |
1356 | : DAG.getUNDEF(In.VT)); |
1357 | |
1358 | // Record the number and types of arguments. |
1359 | MFI->addParam(VT: In.VT); |
1360 | } |
1361 | |
1362 | // For swiftcc, emit additional swiftself and swifterror arguments |
1363 | // if there aren't. These additional arguments are also added for callee |
1364 | // signature They are necessary to match callee and caller signature for |
1365 | // indirect call. |
1366 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
1367 | if (CallConv == CallingConv::Swift) { |
1368 | if (!HasSwiftSelfArg) { |
1369 | MFI->addParam(VT: PtrVT); |
1370 | } |
1371 | if (!HasSwiftErrorArg) { |
1372 | MFI->addParam(VT: PtrVT); |
1373 | } |
1374 | } |
1375 | // Varargs are copied into a buffer allocated by the caller, and a pointer to |
1376 | // the buffer is passed as an argument. |
1377 | if (IsVarArg) { |
1378 | MVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
1379 | Register VarargVreg = |
1380 | MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: PtrVT)); |
1381 | MFI->setVarargBufferVreg(VarargVreg); |
1382 | Chain = DAG.getCopyToReg( |
1383 | Chain, DL, VarargVreg, |
1384 | DAG.getNode(WebAssemblyISD::ARGUMENT, DL, PtrVT, |
1385 | DAG.getTargetConstant(Ins.size(), DL, MVT::i32))); |
1386 | MFI->addParam(VT: PtrVT); |
1387 | } |
1388 | |
1389 | // Record the number and types of arguments and results. |
1390 | SmallVector<MVT, 4> Params; |
1391 | SmallVector<MVT, 4> Results; |
1392 | computeSignatureVTs(Ty: MF.getFunction().getFunctionType(), TargetFunc: &MF.getFunction(), |
1393 | ContextFunc: MF.getFunction(), TM: DAG.getTarget(), Params, Results); |
1394 | for (MVT VT : Results) |
1395 | MFI->addResult(VT); |
1396 | // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify |
1397 | // the param logic here with ComputeSignatureVTs |
1398 | assert(MFI->getParams().size() == Params.size() && |
1399 | std::equal(MFI->getParams().begin(), MFI->getParams().end(), |
1400 | Params.begin())); |
1401 | |
1402 | return Chain; |
1403 | } |
1404 | |
1405 | void WebAssemblyTargetLowering::ReplaceNodeResults( |
1406 | SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const { |
1407 | switch (N->getOpcode()) { |
1408 | case ISD::SIGN_EXTEND_INREG: |
1409 | // Do not add any results, signifying that N should not be custom lowered |
1410 | // after all. This happens because simd128 turns on custom lowering for |
1411 | // SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an |
1412 | // illegal type. |
1413 | break; |
1414 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
1415 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
1416 | // Do not add any results, signifying that N should not be custom lowered. |
1417 | // EXTEND_VECTOR_INREG is implemented for some vectors, but not all. |
1418 | break; |
1419 | default: |
1420 | llvm_unreachable( |
1421 | "ReplaceNodeResults not implemented for this op for WebAssembly!" ); |
1422 | } |
1423 | } |
1424 | |
1425 | //===----------------------------------------------------------------------===// |
1426 | // Custom lowering hooks. |
1427 | //===----------------------------------------------------------------------===// |
1428 | |
1429 | SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op, |
1430 | SelectionDAG &DAG) const { |
1431 | SDLoc DL(Op); |
1432 | switch (Op.getOpcode()) { |
1433 | default: |
1434 | llvm_unreachable("unimplemented operation lowering" ); |
1435 | return SDValue(); |
1436 | case ISD::FrameIndex: |
1437 | return LowerFrameIndex(Op, DAG); |
1438 | case ISD::GlobalAddress: |
1439 | return LowerGlobalAddress(Op, DAG); |
1440 | case ISD::GlobalTLSAddress: |
1441 | return LowerGlobalTLSAddress(Op, DAG); |
1442 | case ISD::ExternalSymbol: |
1443 | return LowerExternalSymbol(Op, DAG); |
1444 | case ISD::JumpTable: |
1445 | return LowerJumpTable(Op, DAG); |
1446 | case ISD::BR_JT: |
1447 | return LowerBR_JT(Op, DAG); |
1448 | case ISD::VASTART: |
1449 | return LowerVASTART(Op, DAG); |
1450 | case ISD::BlockAddress: |
1451 | case ISD::BRIND: |
1452 | fail(DL, DAG, Msg: "WebAssembly hasn't implemented computed gotos" ); |
1453 | return SDValue(); |
1454 | case ISD::RETURNADDR: |
1455 | return LowerRETURNADDR(Op, DAG); |
1456 | case ISD::FRAMEADDR: |
1457 | return LowerFRAMEADDR(Op, DAG); |
1458 | case ISD::CopyToReg: |
1459 | return LowerCopyToReg(Op, DAG); |
1460 | case ISD::EXTRACT_VECTOR_ELT: |
1461 | case ISD::INSERT_VECTOR_ELT: |
1462 | return LowerAccessVectorElement(Op, DAG); |
1463 | case ISD::INTRINSIC_VOID: |
1464 | case ISD::INTRINSIC_WO_CHAIN: |
1465 | case ISD::INTRINSIC_W_CHAIN: |
1466 | return LowerIntrinsic(Op, DAG); |
1467 | case ISD::SIGN_EXTEND_INREG: |
1468 | return LowerSIGN_EXTEND_INREG(Op, DAG); |
1469 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
1470 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
1471 | return LowerEXTEND_VECTOR_INREG(Op, DAG); |
1472 | case ISD::BUILD_VECTOR: |
1473 | return LowerBUILD_VECTOR(Op, DAG); |
1474 | case ISD::VECTOR_SHUFFLE: |
1475 | return LowerVECTOR_SHUFFLE(Op, DAG); |
1476 | case ISD::SETCC: |
1477 | return LowerSETCC(Op, DAG); |
1478 | case ISD::SHL: |
1479 | case ISD::SRA: |
1480 | case ISD::SRL: |
1481 | return LowerShift(Op, DAG); |
1482 | case ISD::FP_TO_SINT_SAT: |
1483 | case ISD::FP_TO_UINT_SAT: |
1484 | return LowerFP_TO_INT_SAT(Op, DAG); |
1485 | case ISD::LOAD: |
1486 | return LowerLoad(Op, DAG); |
1487 | case ISD::STORE: |
1488 | return LowerStore(Op, DAG); |
1489 | case ISD::CTPOP: |
1490 | case ISD::CTLZ: |
1491 | case ISD::CTTZ: |
1492 | return DAG.UnrollVectorOp(N: Op.getNode()); |
1493 | } |
1494 | } |
1495 | |
1496 | static bool IsWebAssemblyGlobal(SDValue Op) { |
1497 | if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op)) |
1498 | return WebAssembly::isWasmVarAddressSpace(AS: GA->getAddressSpace()); |
1499 | |
1500 | return false; |
1501 | } |
1502 | |
1503 | static std::optional<unsigned> IsWebAssemblyLocal(SDValue Op, |
1504 | SelectionDAG &DAG) { |
1505 | const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Op); |
1506 | if (!FI) |
1507 | return std::nullopt; |
1508 | |
1509 | auto &MF = DAG.getMachineFunction(); |
1510 | return WebAssemblyFrameLowering::getLocalForStackObject(MF, FrameIndex: FI->getIndex()); |
1511 | } |
1512 | |
1513 | SDValue WebAssemblyTargetLowering::LowerStore(SDValue Op, |
1514 | SelectionDAG &DAG) const { |
1515 | SDLoc DL(Op); |
1516 | StoreSDNode *SN = cast<StoreSDNode>(Val: Op.getNode()); |
1517 | const SDValue &Value = SN->getValue(); |
1518 | const SDValue &Base = SN->getBasePtr(); |
1519 | const SDValue &Offset = SN->getOffset(); |
1520 | |
1521 | if (IsWebAssemblyGlobal(Op: Base)) { |
1522 | if (!Offset->isUndef()) |
1523 | report_fatal_error(reason: "unexpected offset when storing to webassembly global" , |
1524 | gen_crash_diag: false); |
1525 | |
1526 | SDVTList Tys = DAG.getVTList(MVT::Other); |
1527 | SDValue Ops[] = {SN->getChain(), Value, Base}; |
1528 | return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_SET, dl: DL, VTList: Tys, Ops, |
1529 | MemVT: SN->getMemoryVT(), MMO: SN->getMemOperand()); |
1530 | } |
1531 | |
1532 | if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) { |
1533 | if (!Offset->isUndef()) |
1534 | report_fatal_error(reason: "unexpected offset when storing to webassembly local" , |
1535 | gen_crash_diag: false); |
1536 | |
1537 | SDValue Idx = DAG.getTargetConstant(*Local, Base, MVT::i32); |
1538 | SDVTList Tys = DAG.getVTList(MVT::Other); // The chain. |
1539 | SDValue Ops[] = {SN->getChain(), Idx, Value}; |
1540 | return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_SET, DL, VTList: Tys, Ops); |
1541 | } |
1542 | |
1543 | if (WebAssembly::isWasmVarAddressSpace(AS: SN->getAddressSpace())) |
1544 | report_fatal_error( |
1545 | reason: "Encountered an unlowerable store to the wasm_var address space" , |
1546 | gen_crash_diag: false); |
1547 | |
1548 | return Op; |
1549 | } |
1550 | |
1551 | SDValue WebAssemblyTargetLowering::LowerLoad(SDValue Op, |
1552 | SelectionDAG &DAG) const { |
1553 | SDLoc DL(Op); |
1554 | LoadSDNode *LN = cast<LoadSDNode>(Val: Op.getNode()); |
1555 | const SDValue &Base = LN->getBasePtr(); |
1556 | const SDValue &Offset = LN->getOffset(); |
1557 | |
1558 | if (IsWebAssemblyGlobal(Op: Base)) { |
1559 | if (!Offset->isUndef()) |
1560 | report_fatal_error( |
1561 | reason: "unexpected offset when loading from webassembly global" , gen_crash_diag: false); |
1562 | |
1563 | SDVTList Tys = DAG.getVTList(LN->getValueType(0), MVT::Other); |
1564 | SDValue Ops[] = {LN->getChain(), Base}; |
1565 | return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_GET, dl: DL, VTList: Tys, Ops, |
1566 | MemVT: LN->getMemoryVT(), MMO: LN->getMemOperand()); |
1567 | } |
1568 | |
1569 | if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) { |
1570 | if (!Offset->isUndef()) |
1571 | report_fatal_error( |
1572 | reason: "unexpected offset when loading from webassembly local" , gen_crash_diag: false); |
1573 | |
1574 | SDValue Idx = DAG.getTargetConstant(*Local, Base, MVT::i32); |
1575 | EVT LocalVT = LN->getValueType(ResNo: 0); |
1576 | SDValue LocalGet = DAG.getNode(Opcode: WebAssemblyISD::LOCAL_GET, DL, VT: LocalVT, |
1577 | Ops: {LN->getChain(), Idx}); |
1578 | SDValue Result = DAG.getMergeValues(Ops: {LocalGet, LN->getChain()}, dl: DL); |
1579 | assert(Result->getNumValues() == 2 && "Loads must carry a chain!" ); |
1580 | return Result; |
1581 | } |
1582 | |
1583 | if (WebAssembly::isWasmVarAddressSpace(AS: LN->getAddressSpace())) |
1584 | report_fatal_error( |
1585 | reason: "Encountered an unlowerable load from the wasm_var address space" , |
1586 | gen_crash_diag: false); |
1587 | |
1588 | return Op; |
1589 | } |
1590 | |
1591 | SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op, |
1592 | SelectionDAG &DAG) const { |
1593 | SDValue Src = Op.getOperand(i: 2); |
1594 | if (isa<FrameIndexSDNode>(Val: Src.getNode())) { |
1595 | // CopyToReg nodes don't support FrameIndex operands. Other targets select |
1596 | // the FI to some LEA-like instruction, but since we don't have that, we |
1597 | // need to insert some kind of instruction that can take an FI operand and |
1598 | // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy |
1599 | // local.copy between Op and its FI operand. |
1600 | SDValue Chain = Op.getOperand(i: 0); |
1601 | SDLoc DL(Op); |
1602 | Register Reg = cast<RegisterSDNode>(Val: Op.getOperand(i: 1))->getReg(); |
1603 | EVT VT = Src.getValueType(); |
1604 | SDValue Copy(DAG.getMachineNode(VT == MVT::i32 ? WebAssembly::COPY_I32 |
1605 | : WebAssembly::COPY_I64, |
1606 | DL, VT, Src), |
1607 | 0); |
1608 | return Op.getNode()->getNumValues() == 1 |
1609 | ? DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy) |
1610 | : DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy, |
1611 | Glue: Op.getNumOperands() == 4 ? Op.getOperand(i: 3) |
1612 | : SDValue()); |
1613 | } |
1614 | return SDValue(); |
1615 | } |
1616 | |
1617 | SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op, |
1618 | SelectionDAG &DAG) const { |
1619 | int FI = cast<FrameIndexSDNode>(Val&: Op)->getIndex(); |
1620 | return DAG.getTargetFrameIndex(FI, VT: Op.getValueType()); |
1621 | } |
1622 | |
1623 | SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op, |
1624 | SelectionDAG &DAG) const { |
1625 | SDLoc DL(Op); |
1626 | |
1627 | if (!Subtarget->getTargetTriple().isOSEmscripten()) { |
1628 | fail(DL, DAG, |
1629 | Msg: "Non-Emscripten WebAssembly hasn't implemented " |
1630 | "__builtin_return_address" ); |
1631 | return SDValue(); |
1632 | } |
1633 | |
1634 | if (verifyReturnAddressArgumentIsConstant(Op, DAG)) |
1635 | return SDValue(); |
1636 | |
1637 | unsigned Depth = Op.getConstantOperandVal(i: 0); |
1638 | MakeLibCallOptions CallOptions; |
1639 | return makeLibCall(DAG, RTLIB::RETURN_ADDRESS, Op.getValueType(), |
1640 | {DAG.getConstant(Depth, DL, MVT::i32)}, CallOptions, DL) |
1641 | .first; |
1642 | } |
1643 | |
1644 | SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op, |
1645 | SelectionDAG &DAG) const { |
1646 | // Non-zero depths are not supported by WebAssembly currently. Use the |
1647 | // legalizer's default expansion, which is to return 0 (what this function is |
1648 | // documented to do). |
1649 | if (Op.getConstantOperandVal(i: 0) > 0) |
1650 | return SDValue(); |
1651 | |
1652 | DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true); |
1653 | EVT VT = Op.getValueType(); |
1654 | Register FP = |
1655 | Subtarget->getRegisterInfo()->getFrameRegister(MF: DAG.getMachineFunction()); |
1656 | return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), Reg: FP, VT); |
1657 | } |
1658 | |
1659 | SDValue |
1660 | WebAssemblyTargetLowering::LowerGlobalTLSAddress(SDValue Op, |
1661 | SelectionDAG &DAG) const { |
1662 | SDLoc DL(Op); |
1663 | const auto *GA = cast<GlobalAddressSDNode>(Val&: Op); |
1664 | |
1665 | MachineFunction &MF = DAG.getMachineFunction(); |
1666 | if (!MF.getSubtarget<WebAssemblySubtarget>().hasBulkMemory()) |
1667 | report_fatal_error(reason: "cannot use thread-local storage without bulk memory" , |
1668 | gen_crash_diag: false); |
1669 | |
1670 | const GlobalValue *GV = GA->getGlobal(); |
1671 | |
1672 | // Currently only Emscripten supports dynamic linking with threads. Therefore, |
1673 | // on other targets, if we have thread-local storage, only the local-exec |
1674 | // model is possible. |
1675 | auto model = Subtarget->getTargetTriple().isOSEmscripten() |
1676 | ? GV->getThreadLocalMode() |
1677 | : GlobalValue::LocalExecTLSModel; |
1678 | |
1679 | // Unsupported TLS modes |
1680 | assert(model != GlobalValue::NotThreadLocal); |
1681 | assert(model != GlobalValue::InitialExecTLSModel); |
1682 | |
1683 | if (model == GlobalValue::LocalExecTLSModel || |
1684 | model == GlobalValue::LocalDynamicTLSModel || |
1685 | (model == GlobalValue::GeneralDynamicTLSModel && |
1686 | getTargetMachine().shouldAssumeDSOLocal(GV))) { |
1687 | // For DSO-local TLS variables we use offset from __tls_base |
1688 | |
1689 | MVT PtrVT = getPointerTy(DL: DAG.getDataLayout()); |
1690 | auto GlobalGet = PtrVT == MVT::i64 ? WebAssembly::GLOBAL_GET_I64 |
1691 | : WebAssembly::GLOBAL_GET_I32; |
1692 | const char *BaseName = MF.createExternalSymbolName(Name: "__tls_base" ); |
1693 | |
1694 | SDValue BaseAddr( |
1695 | DAG.getMachineNode(GlobalGet, DL, PtrVT, |
1696 | DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)), |
1697 | 0); |
1698 | |
1699 | SDValue TLSOffset = DAG.getTargetGlobalAddress( |
1700 | GV, DL, VT: PtrVT, offset: GA->getOffset(), TargetFlags: WebAssemblyII::MO_TLS_BASE_REL); |
1701 | SDValue SymOffset = |
1702 | DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: TLSOffset); |
1703 | |
1704 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymOffset); |
1705 | } |
1706 | |
1707 | assert(model == GlobalValue::GeneralDynamicTLSModel); |
1708 | |
1709 | EVT VT = Op.getValueType(); |
1710 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
1711 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, |
1712 | offset: GA->getOffset(), |
1713 | TargetFlags: WebAssemblyII::MO_GOT_TLS)); |
1714 | } |
1715 | |
1716 | SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op, |
1717 | SelectionDAG &DAG) const { |
1718 | SDLoc DL(Op); |
1719 | const auto *GA = cast<GlobalAddressSDNode>(Val&: Op); |
1720 | EVT VT = Op.getValueType(); |
1721 | assert(GA->getTargetFlags() == 0 && |
1722 | "Unexpected target flags on generic GlobalAddressSDNode" ); |
1723 | if (!WebAssembly::isValidAddressSpace(AS: GA->getAddressSpace())) |
1724 | fail(DL, DAG, Msg: "Invalid address space for WebAssembly target" ); |
1725 | |
1726 | unsigned OperandFlags = 0; |
1727 | const GlobalValue *GV = GA->getGlobal(); |
1728 | // Since WebAssembly tables cannot yet be shared accross modules, we don't |
1729 | // need special treatment for tables in PIC mode. |
1730 | if (isPositionIndependent() && |
1731 | !WebAssembly::isWebAssemblyTableType(Ty: GV->getValueType())) { |
1732 | if (getTargetMachine().shouldAssumeDSOLocal(GV)) { |
1733 | MachineFunction &MF = DAG.getMachineFunction(); |
1734 | MVT PtrVT = getPointerTy(DL: MF.getDataLayout()); |
1735 | const char *BaseName; |
1736 | if (GV->getValueType()->isFunctionTy()) { |
1737 | BaseName = MF.createExternalSymbolName(Name: "__table_base" ); |
1738 | OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL; |
1739 | } else { |
1740 | BaseName = MF.createExternalSymbolName(Name: "__memory_base" ); |
1741 | OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL; |
1742 | } |
1743 | SDValue BaseAddr = |
1744 | DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, |
1745 | Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)); |
1746 | |
1747 | SDValue SymAddr = DAG.getNode( |
1748 | Opcode: WebAssemblyISD::WrapperREL, DL, VT, |
1749 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, offset: GA->getOffset(), |
1750 | TargetFlags: OperandFlags)); |
1751 | |
1752 | return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: BaseAddr, N2: SymAddr); |
1753 | } |
1754 | OperandFlags = WebAssemblyII::MO_GOT; |
1755 | } |
1756 | |
1757 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
1758 | Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, |
1759 | offset: GA->getOffset(), TargetFlags: OperandFlags)); |
1760 | } |
1761 | |
1762 | SDValue |
1763 | WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op, |
1764 | SelectionDAG &DAG) const { |
1765 | SDLoc DL(Op); |
1766 | const auto *ES = cast<ExternalSymbolSDNode>(Val&: Op); |
1767 | EVT VT = Op.getValueType(); |
1768 | assert(ES->getTargetFlags() == 0 && |
1769 | "Unexpected target flags on generic ExternalSymbolSDNode" ); |
1770 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT, |
1771 | Operand: DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), VT)); |
1772 | } |
1773 | |
1774 | SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op, |
1775 | SelectionDAG &DAG) const { |
1776 | // There's no need for a Wrapper node because we always incorporate a jump |
1777 | // table operand into a BR_TABLE instruction, rather than ever |
1778 | // materializing it in a register. |
1779 | const JumpTableSDNode *JT = cast<JumpTableSDNode>(Val&: Op); |
1780 | return DAG.getTargetJumpTable(JTI: JT->getIndex(), VT: Op.getValueType(), |
1781 | TargetFlags: JT->getTargetFlags()); |
1782 | } |
1783 | |
1784 | SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op, |
1785 | SelectionDAG &DAG) const { |
1786 | SDLoc DL(Op); |
1787 | SDValue Chain = Op.getOperand(i: 0); |
1788 | const auto *JT = cast<JumpTableSDNode>(Val: Op.getOperand(i: 1)); |
1789 | SDValue Index = Op.getOperand(i: 2); |
1790 | assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags" ); |
1791 | |
1792 | SmallVector<SDValue, 8> Ops; |
1793 | Ops.push_back(Elt: Chain); |
1794 | Ops.push_back(Elt: Index); |
1795 | |
1796 | MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo(); |
1797 | const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs; |
1798 | |
1799 | // Add an operand for each case. |
1800 | for (auto *MBB : MBBs) |
1801 | Ops.push_back(Elt: DAG.getBasicBlock(MBB)); |
1802 | |
1803 | // Add the first MBB as a dummy default target for now. This will be replaced |
1804 | // with the proper default target (and the preceding range check eliminated) |
1805 | // if possible by WebAssemblyFixBrTableDefaults. |
1806 | Ops.push_back(Elt: DAG.getBasicBlock(MBB: *MBBs.begin())); |
1807 | return DAG.getNode(WebAssemblyISD::BR_TABLE, DL, MVT::Other, Ops); |
1808 | } |
1809 | |
1810 | SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op, |
1811 | SelectionDAG &DAG) const { |
1812 | SDLoc DL(Op); |
1813 | EVT PtrVT = getPointerTy(DL: DAG.getMachineFunction().getDataLayout()); |
1814 | |
1815 | auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>(); |
1816 | const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue(); |
1817 | |
1818 | SDValue ArgN = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: DL, |
1819 | Reg: MFI->getVarargBufferVreg(), VT: PtrVT); |
1820 | return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: ArgN, Ptr: Op.getOperand(i: 1), |
1821 | PtrInfo: MachinePointerInfo(SV)); |
1822 | } |
1823 | |
1824 | SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op, |
1825 | SelectionDAG &DAG) const { |
1826 | MachineFunction &MF = DAG.getMachineFunction(); |
1827 | unsigned IntNo; |
1828 | switch (Op.getOpcode()) { |
1829 | case ISD::INTRINSIC_VOID: |
1830 | case ISD::INTRINSIC_W_CHAIN: |
1831 | IntNo = Op.getConstantOperandVal(i: 1); |
1832 | break; |
1833 | case ISD::INTRINSIC_WO_CHAIN: |
1834 | IntNo = Op.getConstantOperandVal(i: 0); |
1835 | break; |
1836 | default: |
1837 | llvm_unreachable("Invalid intrinsic" ); |
1838 | } |
1839 | SDLoc DL(Op); |
1840 | |
1841 | switch (IntNo) { |
1842 | default: |
1843 | return SDValue(); // Don't custom lower most intrinsics. |
1844 | |
1845 | case Intrinsic::wasm_lsda: { |
1846 | auto PtrVT = getPointerTy(DL: MF.getDataLayout()); |
1847 | const char *SymName = MF.createExternalSymbolName( |
1848 | Name: "GCC_except_table" + std::to_string(val: MF.getFunctionNumber())); |
1849 | if (isPositionIndependent()) { |
1850 | SDValue Node = DAG.getTargetExternalSymbol( |
1851 | Sym: SymName, VT: PtrVT, TargetFlags: WebAssemblyII::MO_MEMORY_BASE_REL); |
1852 | const char *BaseName = MF.createExternalSymbolName(Name: "__memory_base" ); |
1853 | SDValue BaseAddr = |
1854 | DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, |
1855 | Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT)); |
1856 | SDValue SymAddr = |
1857 | DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: Node); |
1858 | return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymAddr); |
1859 | } |
1860 | SDValue Node = DAG.getTargetExternalSymbol(Sym: SymName, VT: PtrVT); |
1861 | return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, Operand: Node); |
1862 | } |
1863 | |
1864 | case Intrinsic::wasm_shuffle: { |
1865 | // Drop in-chain and replace undefs, but otherwise pass through unchanged |
1866 | SDValue Ops[18]; |
1867 | size_t OpIdx = 0; |
1868 | Ops[OpIdx++] = Op.getOperand(i: 1); |
1869 | Ops[OpIdx++] = Op.getOperand(i: 2); |
1870 | while (OpIdx < 18) { |
1871 | const SDValue &MaskIdx = Op.getOperand(i: OpIdx + 1); |
1872 | if (MaskIdx.isUndef() || MaskIdx.getNode()->getAsZExtVal() >= 32) { |
1873 | bool isTarget = MaskIdx.getNode()->getOpcode() == ISD::TargetConstant; |
1874 | Ops[OpIdx++] = DAG.getConstant(0, DL, MVT::i32, isTarget); |
1875 | } else { |
1876 | Ops[OpIdx++] = MaskIdx; |
1877 | } |
1878 | } |
1879 | return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops); |
1880 | } |
1881 | } |
1882 | } |
1883 | |
1884 | SDValue |
1885 | WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, |
1886 | SelectionDAG &DAG) const { |
1887 | SDLoc DL(Op); |
1888 | // If sign extension operations are disabled, allow sext_inreg only if operand |
1889 | // is a vector extract of an i8 or i16 lane. SIMD does not depend on sign |
1890 | // extension operations, but allowing sext_inreg in this context lets us have |
1891 | // simple patterns to select extract_lane_s instructions. Expanding sext_inreg |
1892 | // everywhere would be simpler in this file, but would necessitate large and |
1893 | // brittle patterns to undo the expansion and select extract_lane_s |
1894 | // instructions. |
1895 | assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128()); |
1896 | if (Op.getOperand(i: 0).getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
1897 | return SDValue(); |
1898 | |
1899 | const SDValue & = Op.getOperand(i: 0); |
1900 | MVT VecT = Extract.getOperand(i: 0).getSimpleValueType(); |
1901 | if (VecT.getVectorElementType().getSizeInBits() > 32) |
1902 | return SDValue(); |
1903 | MVT = |
1904 | cast<VTSDNode>(Val: Op.getOperand(i: 1).getNode())->getVT().getSimpleVT(); |
1905 | MVT = |
1906 | MVT::getVectorVT(VT: ExtractedLaneT, NumElements: 128 / ExtractedLaneT.getSizeInBits()); |
1907 | if (ExtractedVecT == VecT) |
1908 | return Op; |
1909 | |
1910 | // Bitcast vector to appropriate type to ensure ISel pattern coverage |
1911 | const SDNode *Index = Extract.getOperand(i: 1).getNode(); |
1912 | if (!isa<ConstantSDNode>(Val: Index)) |
1913 | return SDValue(); |
1914 | unsigned IndexVal = Index->getAsZExtVal(); |
1915 | unsigned Scale = |
1916 | ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements(); |
1917 | assert(Scale > 1); |
1918 | SDValue NewIndex = |
1919 | DAG.getConstant(Val: IndexVal * Scale, DL, VT: Index->getValueType(ResNo: 0)); |
1920 | SDValue = DAG.getNode( |
1921 | Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: Extract.getValueType(), |
1922 | N1: DAG.getBitcast(VT: ExtractedVecT, V: Extract.getOperand(i: 0)), N2: NewIndex); |
1923 | return DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: Op.getValueType(), N1: NewExtract, |
1924 | N2: Op.getOperand(i: 1)); |
1925 | } |
1926 | |
1927 | SDValue |
1928 | WebAssemblyTargetLowering::LowerEXTEND_VECTOR_INREG(SDValue Op, |
1929 | SelectionDAG &DAG) const { |
1930 | SDLoc DL(Op); |
1931 | EVT VT = Op.getValueType(); |
1932 | SDValue Src = Op.getOperand(i: 0); |
1933 | EVT SrcVT = Src.getValueType(); |
1934 | |
1935 | if (SrcVT.getVectorElementType() == MVT::i1 || |
1936 | SrcVT.getVectorElementType() == MVT::i64) |
1937 | return SDValue(); |
1938 | |
1939 | assert(VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits() == 0 && |
1940 | "Unexpected extension factor." ); |
1941 | unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits(); |
1942 | |
1943 | if (Scale != 2 && Scale != 4 && Scale != 8) |
1944 | return SDValue(); |
1945 | |
1946 | unsigned Ext; |
1947 | switch (Op.getOpcode()) { |
1948 | case ISD::ZERO_EXTEND_VECTOR_INREG: |
1949 | Ext = WebAssemblyISD::EXTEND_LOW_U; |
1950 | break; |
1951 | case ISD::SIGN_EXTEND_VECTOR_INREG: |
1952 | Ext = WebAssemblyISD::EXTEND_LOW_S; |
1953 | break; |
1954 | } |
1955 | |
1956 | SDValue Ret = Src; |
1957 | while (Scale != 1) { |
1958 | Ret = DAG.getNode(Opcode: Ext, DL, |
1959 | VT: Ret.getValueType() |
1960 | .widenIntegerVectorElementType(Context&: *DAG.getContext()) |
1961 | .getHalfNumVectorElementsVT(Context&: *DAG.getContext()), |
1962 | Operand: Ret); |
1963 | Scale /= 2; |
1964 | } |
1965 | assert(Ret.getValueType() == VT); |
1966 | return Ret; |
1967 | } |
1968 | |
1969 | static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) { |
1970 | SDLoc DL(Op); |
1971 | if (Op.getValueType() != MVT::v2f64) |
1972 | return SDValue(); |
1973 | |
1974 | auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec, |
1975 | unsigned &Index) -> bool { |
1976 | switch (Op.getOpcode()) { |
1977 | case ISD::SINT_TO_FP: |
1978 | Opcode = WebAssemblyISD::CONVERT_LOW_S; |
1979 | break; |
1980 | case ISD::UINT_TO_FP: |
1981 | Opcode = WebAssemblyISD::CONVERT_LOW_U; |
1982 | break; |
1983 | case ISD::FP_EXTEND: |
1984 | Opcode = WebAssemblyISD::PROMOTE_LOW; |
1985 | break; |
1986 | default: |
1987 | return false; |
1988 | } |
1989 | |
1990 | auto = Op.getOperand(i: 0); |
1991 | if (ExtractVector.getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
1992 | return false; |
1993 | |
1994 | if (!isa<ConstantSDNode>(Val: ExtractVector.getOperand(i: 1).getNode())) |
1995 | return false; |
1996 | |
1997 | SrcVec = ExtractVector.getOperand(i: 0); |
1998 | Index = ExtractVector.getConstantOperandVal(i: 1); |
1999 | return true; |
2000 | }; |
2001 | |
2002 | unsigned LHSOpcode, RHSOpcode, LHSIndex, RHSIndex; |
2003 | SDValue LHSSrcVec, RHSSrcVec; |
2004 | if (!GetConvertedLane(Op.getOperand(i: 0), LHSOpcode, LHSSrcVec, LHSIndex) || |
2005 | !GetConvertedLane(Op.getOperand(i: 1), RHSOpcode, RHSSrcVec, RHSIndex)) |
2006 | return SDValue(); |
2007 | |
2008 | if (LHSOpcode != RHSOpcode) |
2009 | return SDValue(); |
2010 | |
2011 | MVT ExpectedSrcVT; |
2012 | switch (LHSOpcode) { |
2013 | case WebAssemblyISD::CONVERT_LOW_S: |
2014 | case WebAssemblyISD::CONVERT_LOW_U: |
2015 | ExpectedSrcVT = MVT::v4i32; |
2016 | break; |
2017 | case WebAssemblyISD::PROMOTE_LOW: |
2018 | ExpectedSrcVT = MVT::v4f32; |
2019 | break; |
2020 | } |
2021 | if (LHSSrcVec.getValueType() != ExpectedSrcVT) |
2022 | return SDValue(); |
2023 | |
2024 | auto Src = LHSSrcVec; |
2025 | if (LHSIndex != 0 || RHSIndex != 1 || LHSSrcVec != RHSSrcVec) { |
2026 | // Shuffle the source vector so that the converted lanes are the low lanes. |
2027 | Src = DAG.getVectorShuffle( |
2028 | VT: ExpectedSrcVT, dl: DL, N1: LHSSrcVec, N2: RHSSrcVec, |
2029 | Mask: {static_cast<int>(LHSIndex), static_cast<int>(RHSIndex) + 4, -1, -1}); |
2030 | } |
2031 | return DAG.getNode(LHSOpcode, DL, MVT::v2f64, Src); |
2032 | } |
2033 | |
2034 | SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op, |
2035 | SelectionDAG &DAG) const { |
2036 | if (auto ConvertLow = LowerConvertLow(Op, DAG)) |
2037 | return ConvertLow; |
2038 | |
2039 | SDLoc DL(Op); |
2040 | const EVT VecT = Op.getValueType(); |
2041 | const EVT LaneT = Op.getOperand(i: 0).getValueType(); |
2042 | const size_t Lanes = Op.getNumOperands(); |
2043 | bool CanSwizzle = VecT == MVT::v16i8; |
2044 | |
2045 | // BUILD_VECTORs are lowered to the instruction that initializes the highest |
2046 | // possible number of lanes at once followed by a sequence of replace_lane |
2047 | // instructions to individually initialize any remaining lanes. |
2048 | |
2049 | // TODO: Tune this. For example, lanewise swizzling is very expensive, so |
2050 | // swizzled lanes should be given greater weight. |
2051 | |
2052 | // TODO: Investigate looping rather than always extracting/replacing specific |
2053 | // lanes to fill gaps. |
2054 | |
2055 | auto IsConstant = [](const SDValue &V) { |
2056 | return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP; |
2057 | }; |
2058 | |
2059 | // Returns the source vector and index vector pair if they exist. Checks for: |
2060 | // (extract_vector_elt |
2061 | // $src, |
2062 | // (sign_extend_inreg (extract_vector_elt $indices, $i)) |
2063 | // ) |
2064 | auto GetSwizzleSrcs = [](size_t I, const SDValue &Lane) { |
2065 | auto Bail = std::make_pair(x: SDValue(), y: SDValue()); |
2066 | if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
2067 | return Bail; |
2068 | const SDValue &SwizzleSrc = Lane->getOperand(Num: 0); |
2069 | const SDValue &IndexExt = Lane->getOperand(Num: 1); |
2070 | if (IndexExt->getOpcode() != ISD::SIGN_EXTEND_INREG) |
2071 | return Bail; |
2072 | const SDValue &Index = IndexExt->getOperand(Num: 0); |
2073 | if (Index->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
2074 | return Bail; |
2075 | const SDValue &SwizzleIndices = Index->getOperand(Num: 0); |
2076 | if (SwizzleSrc.getValueType() != MVT::v16i8 || |
2077 | SwizzleIndices.getValueType() != MVT::v16i8 || |
2078 | Index->getOperand(1)->getOpcode() != ISD::Constant || |
2079 | Index->getConstantOperandVal(1) != I) |
2080 | return Bail; |
2081 | return std::make_pair(x: SwizzleSrc, y: SwizzleIndices); |
2082 | }; |
2083 | |
2084 | // If the lane is extracted from another vector at a constant index, return |
2085 | // that vector. The source vector must not have more lanes than the dest |
2086 | // because the shufflevector indices are in terms of the destination lanes and |
2087 | // would not be able to address the smaller individual source lanes. |
2088 | auto GetShuffleSrc = [&](const SDValue &Lane) { |
2089 | if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT) |
2090 | return SDValue(); |
2091 | if (!isa<ConstantSDNode>(Val: Lane->getOperand(Num: 1).getNode())) |
2092 | return SDValue(); |
2093 | if (Lane->getOperand(Num: 0).getValueType().getVectorNumElements() > |
2094 | VecT.getVectorNumElements()) |
2095 | return SDValue(); |
2096 | return Lane->getOperand(Num: 0); |
2097 | }; |
2098 | |
2099 | using ValueEntry = std::pair<SDValue, size_t>; |
2100 | SmallVector<ValueEntry, 16> SplatValueCounts; |
2101 | |
2102 | using SwizzleEntry = std::pair<std::pair<SDValue, SDValue>, size_t>; |
2103 | SmallVector<SwizzleEntry, 16> SwizzleCounts; |
2104 | |
2105 | using ShuffleEntry = std::pair<SDValue, size_t>; |
2106 | SmallVector<ShuffleEntry, 16> ShuffleCounts; |
2107 | |
2108 | auto AddCount = [](auto &Counts, const auto &Val) { |
2109 | auto CountIt = |
2110 | llvm::find_if(Counts, [&Val](auto E) { return E.first == Val; }); |
2111 | if (CountIt == Counts.end()) { |
2112 | Counts.emplace_back(Val, 1); |
2113 | } else { |
2114 | CountIt->second++; |
2115 | } |
2116 | }; |
2117 | |
2118 | auto GetMostCommon = [](auto &Counts) { |
2119 | auto CommonIt = |
2120 | std::max_element(Counts.begin(), Counts.end(), llvm::less_second()); |
2121 | assert(CommonIt != Counts.end() && "Unexpected all-undef build_vector" ); |
2122 | return *CommonIt; |
2123 | }; |
2124 | |
2125 | size_t NumConstantLanes = 0; |
2126 | |
2127 | // Count eligible lanes for each type of vector creation op |
2128 | for (size_t I = 0; I < Lanes; ++I) { |
2129 | const SDValue &Lane = Op->getOperand(Num: I); |
2130 | if (Lane.isUndef()) |
2131 | continue; |
2132 | |
2133 | AddCount(SplatValueCounts, Lane); |
2134 | |
2135 | if (IsConstant(Lane)) |
2136 | NumConstantLanes++; |
2137 | if (auto ShuffleSrc = GetShuffleSrc(Lane)) |
2138 | AddCount(ShuffleCounts, ShuffleSrc); |
2139 | if (CanSwizzle) { |
2140 | auto SwizzleSrcs = GetSwizzleSrcs(I, Lane); |
2141 | if (SwizzleSrcs.first) |
2142 | AddCount(SwizzleCounts, SwizzleSrcs); |
2143 | } |
2144 | } |
2145 | |
2146 | SDValue SplatValue; |
2147 | size_t NumSplatLanes; |
2148 | std::tie(args&: SplatValue, args&: NumSplatLanes) = GetMostCommon(SplatValueCounts); |
2149 | |
2150 | SDValue SwizzleSrc; |
2151 | SDValue SwizzleIndices; |
2152 | size_t NumSwizzleLanes = 0; |
2153 | if (SwizzleCounts.size()) |
2154 | std::forward_as_tuple(args: std::tie(args&: SwizzleSrc, args&: SwizzleIndices), |
2155 | args&: NumSwizzleLanes) = GetMostCommon(SwizzleCounts); |
2156 | |
2157 | // Shuffles can draw from up to two vectors, so find the two most common |
2158 | // sources. |
2159 | SDValue ShuffleSrc1, ShuffleSrc2; |
2160 | size_t NumShuffleLanes = 0; |
2161 | if (ShuffleCounts.size()) { |
2162 | std::tie(args&: ShuffleSrc1, args&: NumShuffleLanes) = GetMostCommon(ShuffleCounts); |
2163 | llvm::erase_if(C&: ShuffleCounts, |
2164 | P: [&](const auto &Pair) { return Pair.first == ShuffleSrc1; }); |
2165 | } |
2166 | if (ShuffleCounts.size()) { |
2167 | size_t AdditionalShuffleLanes; |
2168 | std::tie(args&: ShuffleSrc2, args&: AdditionalShuffleLanes) = |
2169 | GetMostCommon(ShuffleCounts); |
2170 | NumShuffleLanes += AdditionalShuffleLanes; |
2171 | } |
2172 | |
2173 | // Predicate returning true if the lane is properly initialized by the |
2174 | // original instruction |
2175 | std::function<bool(size_t, const SDValue &)> IsLaneConstructed; |
2176 | SDValue Result; |
2177 | // Prefer swizzles over shuffles over vector consts over splats |
2178 | if (NumSwizzleLanes >= NumShuffleLanes && |
2179 | NumSwizzleLanes >= NumConstantLanes && NumSwizzleLanes >= NumSplatLanes) { |
2180 | Result = DAG.getNode(Opcode: WebAssemblyISD::SWIZZLE, DL, VT: VecT, N1: SwizzleSrc, |
2181 | N2: SwizzleIndices); |
2182 | auto Swizzled = std::make_pair(x&: SwizzleSrc, y&: SwizzleIndices); |
2183 | IsLaneConstructed = [&, Swizzled](size_t I, const SDValue &Lane) { |
2184 | return Swizzled == GetSwizzleSrcs(I, Lane); |
2185 | }; |
2186 | } else if (NumShuffleLanes >= NumConstantLanes && |
2187 | NumShuffleLanes >= NumSplatLanes) { |
2188 | size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits() / 8; |
2189 | size_t DestLaneCount = VecT.getVectorNumElements(); |
2190 | size_t Scale1 = 1; |
2191 | size_t Scale2 = 1; |
2192 | SDValue Src1 = ShuffleSrc1; |
2193 | SDValue Src2 = ShuffleSrc2 ? ShuffleSrc2 : DAG.getUNDEF(VT: VecT); |
2194 | if (Src1.getValueType() != VecT) { |
2195 | size_t LaneSize = |
2196 | Src1.getValueType().getVectorElementType().getFixedSizeInBits() / 8; |
2197 | assert(LaneSize > DestLaneSize); |
2198 | Scale1 = LaneSize / DestLaneSize; |
2199 | Src1 = DAG.getBitcast(VT: VecT, V: Src1); |
2200 | } |
2201 | if (Src2.getValueType() != VecT) { |
2202 | size_t LaneSize = |
2203 | Src2.getValueType().getVectorElementType().getFixedSizeInBits() / 8; |
2204 | assert(LaneSize > DestLaneSize); |
2205 | Scale2 = LaneSize / DestLaneSize; |
2206 | Src2 = DAG.getBitcast(VT: VecT, V: Src2); |
2207 | } |
2208 | |
2209 | int Mask[16]; |
2210 | assert(DestLaneCount <= 16); |
2211 | for (size_t I = 0; I < DestLaneCount; ++I) { |
2212 | const SDValue &Lane = Op->getOperand(Num: I); |
2213 | SDValue Src = GetShuffleSrc(Lane); |
2214 | if (Src == ShuffleSrc1) { |
2215 | Mask[I] = Lane->getConstantOperandVal(Num: 1) * Scale1; |
2216 | } else if (Src && Src == ShuffleSrc2) { |
2217 | Mask[I] = DestLaneCount + Lane->getConstantOperandVal(Num: 1) * Scale2; |
2218 | } else { |
2219 | Mask[I] = -1; |
2220 | } |
2221 | } |
2222 | ArrayRef<int> MaskRef(Mask, DestLaneCount); |
2223 | Result = DAG.getVectorShuffle(VT: VecT, dl: DL, N1: Src1, N2: Src2, Mask: MaskRef); |
2224 | IsLaneConstructed = [&](size_t, const SDValue &Lane) { |
2225 | auto Src = GetShuffleSrc(Lane); |
2226 | return Src == ShuffleSrc1 || (Src && Src == ShuffleSrc2); |
2227 | }; |
2228 | } else if (NumConstantLanes >= NumSplatLanes) { |
2229 | SmallVector<SDValue, 16> ConstLanes; |
2230 | for (const SDValue &Lane : Op->op_values()) { |
2231 | if (IsConstant(Lane)) { |
2232 | // Values may need to be fixed so that they will sign extend to be |
2233 | // within the expected range during ISel. Check whether the value is in |
2234 | // bounds based on the lane bit width and if it is out of bounds, lop |
2235 | // off the extra bits and subtract 2^n to reflect giving the high bit |
2236 | // value -2^(n-1) rather than +2^(n-1). Skip the i64 case because it |
2237 | // cannot possibly be out of range. |
2238 | auto *Const = dyn_cast<ConstantSDNode>(Val: Lane.getNode()); |
2239 | int64_t Val = Const ? Const->getSExtValue() : 0; |
2240 | uint64_t LaneBits = 128 / Lanes; |
2241 | assert((LaneBits == 64 || Val >= -(1ll << (LaneBits - 1))) && |
2242 | "Unexpected out of bounds negative value" ); |
2243 | if (Const && LaneBits != 64 && Val > (1ll << (LaneBits - 1)) - 1) { |
2244 | uint64_t Mask = (1ll << LaneBits) - 1; |
2245 | auto NewVal = (((uint64_t)Val & Mask) - (1ll << LaneBits)) & Mask; |
2246 | ConstLanes.push_back(Elt: DAG.getConstant(Val: NewVal, DL: SDLoc(Lane), VT: LaneT)); |
2247 | } else { |
2248 | ConstLanes.push_back(Elt: Lane); |
2249 | } |
2250 | } else if (LaneT.isFloatingPoint()) { |
2251 | ConstLanes.push_back(Elt: DAG.getConstantFP(Val: 0, DL, VT: LaneT)); |
2252 | } else { |
2253 | ConstLanes.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: LaneT)); |
2254 | } |
2255 | } |
2256 | Result = DAG.getBuildVector(VT: VecT, DL, Ops: ConstLanes); |
2257 | IsLaneConstructed = [&IsConstant](size_t _, const SDValue &Lane) { |
2258 | return IsConstant(Lane); |
2259 | }; |
2260 | } else { |
2261 | // Use a splat (which might be selected as a load splat) |
2262 | Result = DAG.getSplatBuildVector(VT: VecT, DL, Op: SplatValue); |
2263 | IsLaneConstructed = [&SplatValue](size_t _, const SDValue &Lane) { |
2264 | return Lane == SplatValue; |
2265 | }; |
2266 | } |
2267 | |
2268 | assert(Result); |
2269 | assert(IsLaneConstructed); |
2270 | |
2271 | // Add replace_lane instructions for any unhandled values |
2272 | for (size_t I = 0; I < Lanes; ++I) { |
2273 | const SDValue &Lane = Op->getOperand(Num: I); |
2274 | if (!Lane.isUndef() && !IsLaneConstructed(I, Lane)) |
2275 | Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane, |
2276 | DAG.getConstant(I, DL, MVT::i32)); |
2277 | } |
2278 | |
2279 | return Result; |
2280 | } |
2281 | |
2282 | SDValue |
2283 | WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, |
2284 | SelectionDAG &DAG) const { |
2285 | SDLoc DL(Op); |
2286 | ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: Op.getNode())->getMask(); |
2287 | MVT VecType = Op.getOperand(i: 0).getSimpleValueType(); |
2288 | assert(VecType.is128BitVector() && "Unexpected shuffle vector type" ); |
2289 | size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8; |
2290 | |
2291 | // Space for two vector args and sixteen mask indices |
2292 | SDValue Ops[18]; |
2293 | size_t OpIdx = 0; |
2294 | Ops[OpIdx++] = Op.getOperand(i: 0); |
2295 | Ops[OpIdx++] = Op.getOperand(i: 1); |
2296 | |
2297 | // Expand mask indices to byte indices and materialize them as operands |
2298 | for (int M : Mask) { |
2299 | for (size_t J = 0; J < LaneBytes; ++J) { |
2300 | // Lower undefs (represented by -1 in mask) to {0..J}, which use a |
2301 | // whole lane of vector input, to allow further reduction at VM. E.g. |
2302 | // match an 8x16 byte shuffle to an equivalent cheaper 32x4 shuffle. |
2303 | uint64_t ByteIndex = M == -1 ? J : (uint64_t)M * LaneBytes + J; |
2304 | Ops[OpIdx++] = DAG.getConstant(ByteIndex, DL, MVT::i32); |
2305 | } |
2306 | } |
2307 | |
2308 | return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops); |
2309 | } |
2310 | |
2311 | SDValue WebAssemblyTargetLowering::LowerSETCC(SDValue Op, |
2312 | SelectionDAG &DAG) const { |
2313 | SDLoc DL(Op); |
2314 | // The legalizer does not know how to expand the unsupported comparison modes |
2315 | // of i64x2 vectors, so we manually unroll them here. |
2316 | assert(Op->getOperand(0)->getSimpleValueType(0) == MVT::v2i64); |
2317 | SmallVector<SDValue, 2> LHS, RHS; |
2318 | DAG.ExtractVectorElements(Op: Op->getOperand(Num: 0), Args&: LHS); |
2319 | DAG.ExtractVectorElements(Op: Op->getOperand(Num: 1), Args&: RHS); |
2320 | const SDValue &CC = Op->getOperand(Num: 2); |
2321 | auto MakeLane = [&](unsigned I) { |
2322 | return DAG.getNode(ISD::SELECT_CC, DL, MVT::i64, LHS[I], RHS[I], |
2323 | DAG.getConstant(uint64_t(-1), DL, MVT::i64), |
2324 | DAG.getConstant(uint64_t(0), DL, MVT::i64), CC); |
2325 | }; |
2326 | return DAG.getBuildVector(Op->getValueType(ResNo: 0), DL, |
2327 | {MakeLane(0), MakeLane(1)}); |
2328 | } |
2329 | |
2330 | SDValue |
2331 | WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op, |
2332 | SelectionDAG &DAG) const { |
2333 | // Allow constant lane indices, expand variable lane indices |
2334 | SDNode *IdxNode = Op.getOperand(i: Op.getNumOperands() - 1).getNode(); |
2335 | if (isa<ConstantSDNode>(Val: IdxNode)) { |
2336 | // Ensure the index type is i32 to match the tablegen patterns |
2337 | uint64_t Idx = IdxNode->getAsZExtVal(); |
2338 | SmallVector<SDValue, 3> Ops(Op.getNode()->ops()); |
2339 | Ops[Op.getNumOperands() - 1] = |
2340 | DAG.getConstant(Idx, SDLoc(IdxNode), MVT::i32); |
2341 | return DAG.getNode(Opcode: Op.getOpcode(), DL: SDLoc(Op), VT: Op.getValueType(), Ops); |
2342 | } |
2343 | // Perform default expansion |
2344 | return SDValue(); |
2345 | } |
2346 | |
2347 | static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) { |
2348 | EVT LaneT = Op.getSimpleValueType().getVectorElementType(); |
2349 | // 32-bit and 64-bit unrolled shifts will have proper semantics |
2350 | if (LaneT.bitsGE(MVT::i32)) |
2351 | return DAG.UnrollVectorOp(N: Op.getNode()); |
2352 | // Otherwise mask the shift value to get proper semantics from 32-bit shift |
2353 | SDLoc DL(Op); |
2354 | size_t NumLanes = Op.getSimpleValueType().getVectorNumElements(); |
2355 | SDValue Mask = DAG.getConstant(LaneT.getSizeInBits() - 1, DL, MVT::i32); |
2356 | unsigned ShiftOpcode = Op.getOpcode(); |
2357 | SmallVector<SDValue, 16> ShiftedElements; |
2358 | DAG.ExtractVectorElements(Op.getOperand(0), ShiftedElements, 0, 0, MVT::i32); |
2359 | SmallVector<SDValue, 16> ShiftElements; |
2360 | DAG.ExtractVectorElements(Op.getOperand(1), ShiftElements, 0, 0, MVT::i32); |
2361 | SmallVector<SDValue, 16> UnrolledOps; |
2362 | for (size_t i = 0; i < NumLanes; ++i) { |
2363 | SDValue MaskedShiftValue = |
2364 | DAG.getNode(ISD::AND, DL, MVT::i32, ShiftElements[i], Mask); |
2365 | SDValue ShiftedValue = ShiftedElements[i]; |
2366 | if (ShiftOpcode == ISD::SRA) |
2367 | ShiftedValue = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, |
2368 | ShiftedValue, DAG.getValueType(LaneT)); |
2369 | UnrolledOps.push_back( |
2370 | DAG.getNode(ShiftOpcode, DL, MVT::i32, ShiftedValue, MaskedShiftValue)); |
2371 | } |
2372 | return DAG.getBuildVector(VT: Op.getValueType(), DL, Ops: UnrolledOps); |
2373 | } |
2374 | |
2375 | SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op, |
2376 | SelectionDAG &DAG) const { |
2377 | SDLoc DL(Op); |
2378 | |
2379 | // Only manually lower vector shifts |
2380 | assert(Op.getSimpleValueType().isVector()); |
2381 | |
2382 | uint64_t LaneBits = Op.getValueType().getScalarSizeInBits(); |
2383 | auto ShiftVal = Op.getOperand(i: 1); |
2384 | |
2385 | // Try to skip bitmask operation since it is implied inside shift instruction |
2386 | auto SkipImpliedMask = [](SDValue MaskOp, uint64_t MaskBits) { |
2387 | if (MaskOp.getOpcode() != ISD::AND) |
2388 | return MaskOp; |
2389 | SDValue LHS = MaskOp.getOperand(i: 0); |
2390 | SDValue RHS = MaskOp.getOperand(i: 1); |
2391 | if (MaskOp.getValueType().isVector()) { |
2392 | APInt MaskVal; |
2393 | if (!ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal)) |
2394 | std::swap(a&: LHS, b&: RHS); |
2395 | |
2396 | if (ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal) && |
2397 | MaskVal == MaskBits) |
2398 | MaskOp = LHS; |
2399 | } else { |
2400 | if (!isa<ConstantSDNode>(Val: RHS.getNode())) |
2401 | std::swap(a&: LHS, b&: RHS); |
2402 | |
2403 | auto ConstantRHS = dyn_cast<ConstantSDNode>(Val: RHS.getNode()); |
2404 | if (ConstantRHS && ConstantRHS->getAPIntValue() == MaskBits) |
2405 | MaskOp = LHS; |
2406 | } |
2407 | |
2408 | return MaskOp; |
2409 | }; |
2410 | |
2411 | // Skip vector and operation |
2412 | ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1); |
2413 | ShiftVal = DAG.getSplatValue(V: ShiftVal); |
2414 | if (!ShiftVal) |
2415 | return unrollVectorShift(Op, DAG); |
2416 | |
2417 | // Skip scalar and operation |
2418 | ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1); |
2419 | // Use anyext because none of the high bits can affect the shift |
2420 | ShiftVal = DAG.getAnyExtOrTrunc(ShiftVal, DL, MVT::i32); |
2421 | |
2422 | unsigned Opcode; |
2423 | switch (Op.getOpcode()) { |
2424 | case ISD::SHL: |
2425 | Opcode = WebAssemblyISD::VEC_SHL; |
2426 | break; |
2427 | case ISD::SRA: |
2428 | Opcode = WebAssemblyISD::VEC_SHR_S; |
2429 | break; |
2430 | case ISD::SRL: |
2431 | Opcode = WebAssemblyISD::VEC_SHR_U; |
2432 | break; |
2433 | default: |
2434 | llvm_unreachable("unexpected opcode" ); |
2435 | } |
2436 | |
2437 | return DAG.getNode(Opcode, DL, VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: ShiftVal); |
2438 | } |
2439 | |
2440 | SDValue WebAssemblyTargetLowering::LowerFP_TO_INT_SAT(SDValue Op, |
2441 | SelectionDAG &DAG) const { |
2442 | SDLoc DL(Op); |
2443 | EVT ResT = Op.getValueType(); |
2444 | EVT SatVT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT(); |
2445 | |
2446 | if ((ResT == MVT::i32 || ResT == MVT::i64) && |
2447 | (SatVT == MVT::i32 || SatVT == MVT::i64)) |
2448 | return Op; |
2449 | |
2450 | if (ResT == MVT::v4i32 && SatVT == MVT::i32) |
2451 | return Op; |
2452 | |
2453 | return SDValue(); |
2454 | } |
2455 | |
2456 | //===----------------------------------------------------------------------===// |
2457 | // Custom DAG combine hooks |
2458 | //===----------------------------------------------------------------------===// |
2459 | static SDValue |
2460 | performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
2461 | auto &DAG = DCI.DAG; |
2462 | auto Shuffle = cast<ShuffleVectorSDNode>(Val: N); |
2463 | |
2464 | // Hoist vector bitcasts that don't change the number of lanes out of unary |
2465 | // shuffles, where they are less likely to get in the way of other combines. |
2466 | // (shuffle (vNxT1 (bitcast (vNxT0 x))), undef, mask) -> |
2467 | // (vNxT1 (bitcast (vNxT0 (shuffle x, undef, mask)))) |
2468 | SDValue Bitcast = N->getOperand(Num: 0); |
2469 | if (Bitcast.getOpcode() != ISD::BITCAST) |
2470 | return SDValue(); |
2471 | if (!N->getOperand(Num: 1).isUndef()) |
2472 | return SDValue(); |
2473 | SDValue CastOp = Bitcast.getOperand(i: 0); |
2474 | EVT SrcType = CastOp.getValueType(); |
2475 | EVT DstType = Bitcast.getValueType(); |
2476 | if (!SrcType.is128BitVector() || |
2477 | SrcType.getVectorNumElements() != DstType.getVectorNumElements()) |
2478 | return SDValue(); |
2479 | SDValue NewShuffle = DAG.getVectorShuffle( |
2480 | VT: SrcType, dl: SDLoc(N), N1: CastOp, N2: DAG.getUNDEF(VT: SrcType), Mask: Shuffle->getMask()); |
2481 | return DAG.getBitcast(VT: DstType, V: NewShuffle); |
2482 | } |
2483 | |
2484 | /// Convert ({u,s}itofp vec) --> ({u,s}itofp ({s,z}ext vec)) so it doesn't get |
2485 | /// split up into scalar instructions during legalization, and the vector |
2486 | /// extending instructions are selected in performVectorExtendCombine below. |
2487 | static SDValue |
2488 | performVectorExtendToFPCombine(SDNode *N, |
2489 | TargetLowering::DAGCombinerInfo &DCI) { |
2490 | auto &DAG = DCI.DAG; |
2491 | assert(N->getOpcode() == ISD::UINT_TO_FP || |
2492 | N->getOpcode() == ISD::SINT_TO_FP); |
2493 | |
2494 | EVT InVT = N->getOperand(Num: 0)->getValueType(ResNo: 0); |
2495 | EVT ResVT = N->getValueType(ResNo: 0); |
2496 | MVT ExtVT; |
2497 | if (ResVT == MVT::v4f32 && (InVT == MVT::v4i16 || InVT == MVT::v4i8)) |
2498 | ExtVT = MVT::v4i32; |
2499 | else if (ResVT == MVT::v2f64 && (InVT == MVT::v2i16 || InVT == MVT::v2i8)) |
2500 | ExtVT = MVT::v2i32; |
2501 | else |
2502 | return SDValue(); |
2503 | |
2504 | unsigned Op = |
2505 | N->getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND; |
2506 | SDValue Conv = DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ExtVT, Operand: N->getOperand(Num: 0)); |
2507 | return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: ResVT, Operand: Conv); |
2508 | } |
2509 | |
2510 | static SDValue |
2511 | performVectorExtendCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
2512 | auto &DAG = DCI.DAG; |
2513 | assert(N->getOpcode() == ISD::SIGN_EXTEND || |
2514 | N->getOpcode() == ISD::ZERO_EXTEND); |
2515 | |
2516 | // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if |
2517 | // possible before the extract_subvector can be expanded. |
2518 | auto = N->getOperand(Num: 0); |
2519 | if (Extract.getOpcode() != ISD::EXTRACT_SUBVECTOR) |
2520 | return SDValue(); |
2521 | auto Source = Extract.getOperand(i: 0); |
2522 | auto *IndexNode = dyn_cast<ConstantSDNode>(Val: Extract.getOperand(i: 1)); |
2523 | if (IndexNode == nullptr) |
2524 | return SDValue(); |
2525 | auto Index = IndexNode->getZExtValue(); |
2526 | |
2527 | // Only v8i8, v4i16, and v2i32 extracts can be widened, and only if the |
2528 | // extracted subvector is the low or high half of its source. |
2529 | EVT ResVT = N->getValueType(ResNo: 0); |
2530 | if (ResVT == MVT::v8i16) { |
2531 | if (Extract.getValueType() != MVT::v8i8 || |
2532 | Source.getValueType() != MVT::v16i8 || (Index != 0 && Index != 8)) |
2533 | return SDValue(); |
2534 | } else if (ResVT == MVT::v4i32) { |
2535 | if (Extract.getValueType() != MVT::v4i16 || |
2536 | Source.getValueType() != MVT::v8i16 || (Index != 0 && Index != 4)) |
2537 | return SDValue(); |
2538 | } else if (ResVT == MVT::v2i64) { |
2539 | if (Extract.getValueType() != MVT::v2i32 || |
2540 | Source.getValueType() != MVT::v4i32 || (Index != 0 && Index != 2)) |
2541 | return SDValue(); |
2542 | } else { |
2543 | return SDValue(); |
2544 | } |
2545 | |
2546 | bool IsSext = N->getOpcode() == ISD::SIGN_EXTEND; |
2547 | bool IsLow = Index == 0; |
2548 | |
2549 | unsigned Op = IsSext ? (IsLow ? WebAssemblyISD::EXTEND_LOW_S |
2550 | : WebAssemblyISD::EXTEND_HIGH_S) |
2551 | : (IsLow ? WebAssemblyISD::EXTEND_LOW_U |
2552 | : WebAssemblyISD::EXTEND_HIGH_U); |
2553 | |
2554 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
2555 | } |
2556 | |
2557 | static SDValue |
2558 | performVectorTruncZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) { |
2559 | auto &DAG = DCI.DAG; |
2560 | |
2561 | auto GetWasmConversionOp = [](unsigned Op) { |
2562 | switch (Op) { |
2563 | case ISD::FP_TO_SINT_SAT: |
2564 | return WebAssemblyISD::TRUNC_SAT_ZERO_S; |
2565 | case ISD::FP_TO_UINT_SAT: |
2566 | return WebAssemblyISD::TRUNC_SAT_ZERO_U; |
2567 | case ISD::FP_ROUND: |
2568 | return WebAssemblyISD::DEMOTE_ZERO; |
2569 | } |
2570 | llvm_unreachable("unexpected op" ); |
2571 | }; |
2572 | |
2573 | auto IsZeroSplat = [](SDValue SplatVal) { |
2574 | auto *Splat = dyn_cast<BuildVectorSDNode>(Val: SplatVal.getNode()); |
2575 | APInt SplatValue, SplatUndef; |
2576 | unsigned SplatBitSize; |
2577 | bool HasAnyUndefs; |
2578 | // Endianness doesn't matter in this context because we are looking for |
2579 | // an all-zero value. |
2580 | return Splat && |
2581 | Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize, |
2582 | HasAnyUndefs) && |
2583 | SplatValue == 0; |
2584 | }; |
2585 | |
2586 | if (N->getOpcode() == ISD::CONCAT_VECTORS) { |
2587 | // Combine this: |
2588 | // |
2589 | // (concat_vectors (v2i32 (fp_to_{s,u}int_sat $x, 32)), (v2i32 (splat 0))) |
2590 | // |
2591 | // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x). |
2592 | // |
2593 | // Or this: |
2594 | // |
2595 | // (concat_vectors (v2f32 (fp_round (v2f64 $x))), (v2f32 (splat 0))) |
2596 | // |
2597 | // into (f32x4.demote_zero_f64x2 $x). |
2598 | EVT ResVT; |
2599 | EVT ExpectedConversionType; |
2600 | auto Conversion = N->getOperand(Num: 0); |
2601 | auto ConversionOp = Conversion.getOpcode(); |
2602 | switch (ConversionOp) { |
2603 | case ISD::FP_TO_SINT_SAT: |
2604 | case ISD::FP_TO_UINT_SAT: |
2605 | ResVT = MVT::v4i32; |
2606 | ExpectedConversionType = MVT::v2i32; |
2607 | break; |
2608 | case ISD::FP_ROUND: |
2609 | ResVT = MVT::v4f32; |
2610 | ExpectedConversionType = MVT::v2f32; |
2611 | break; |
2612 | default: |
2613 | return SDValue(); |
2614 | } |
2615 | |
2616 | if (N->getValueType(ResNo: 0) != ResVT) |
2617 | return SDValue(); |
2618 | |
2619 | if (Conversion.getValueType() != ExpectedConversionType) |
2620 | return SDValue(); |
2621 | |
2622 | auto Source = Conversion.getOperand(i: 0); |
2623 | if (Source.getValueType() != MVT::v2f64) |
2624 | return SDValue(); |
2625 | |
2626 | if (!IsZeroSplat(N->getOperand(Num: 1)) || |
2627 | N->getOperand(Num: 1).getValueType() != ExpectedConversionType) |
2628 | return SDValue(); |
2629 | |
2630 | unsigned Op = GetWasmConversionOp(ConversionOp); |
2631 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
2632 | } |
2633 | |
2634 | // Combine this: |
2635 | // |
2636 | // (fp_to_{s,u}int_sat (concat_vectors $x, (v2f64 (splat 0))), 32) |
2637 | // |
2638 | // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x). |
2639 | // |
2640 | // Or this: |
2641 | // |
2642 | // (v4f32 (fp_round (concat_vectors $x, (v2f64 (splat 0))))) |
2643 | // |
2644 | // into (f32x4.demote_zero_f64x2 $x). |
2645 | EVT ResVT; |
2646 | auto ConversionOp = N->getOpcode(); |
2647 | switch (ConversionOp) { |
2648 | case ISD::FP_TO_SINT_SAT: |
2649 | case ISD::FP_TO_UINT_SAT: |
2650 | ResVT = MVT::v4i32; |
2651 | break; |
2652 | case ISD::FP_ROUND: |
2653 | ResVT = MVT::v4f32; |
2654 | break; |
2655 | default: |
2656 | llvm_unreachable("unexpected op" ); |
2657 | } |
2658 | |
2659 | if (N->getValueType(ResNo: 0) != ResVT) |
2660 | return SDValue(); |
2661 | |
2662 | auto Concat = N->getOperand(Num: 0); |
2663 | if (Concat.getValueType() != MVT::v4f64) |
2664 | return SDValue(); |
2665 | |
2666 | auto Source = Concat.getOperand(i: 0); |
2667 | if (Source.getValueType() != MVT::v2f64) |
2668 | return SDValue(); |
2669 | |
2670 | if (!IsZeroSplat(Concat.getOperand(1)) || |
2671 | Concat.getOperand(1).getValueType() != MVT::v2f64) |
2672 | return SDValue(); |
2673 | |
2674 | unsigned Op = GetWasmConversionOp(ConversionOp); |
2675 | return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source); |
2676 | } |
2677 | |
2678 | // Helper to extract VectorWidth bits from Vec, starting from IdxVal. |
2679 | static SDValue (SDValue Vec, unsigned IdxVal, SelectionDAG &DAG, |
2680 | const SDLoc &DL, unsigned VectorWidth) { |
2681 | EVT VT = Vec.getValueType(); |
2682 | EVT ElVT = VT.getVectorElementType(); |
2683 | unsigned Factor = VT.getSizeInBits() / VectorWidth; |
2684 | EVT ResultVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ElVT, |
2685 | NumElements: VT.getVectorNumElements() / Factor); |
2686 | |
2687 | // Extract the relevant VectorWidth bits. Generate an EXTRACT_SUBVECTOR |
2688 | unsigned ElemsPerChunk = VectorWidth / ElVT.getSizeInBits(); |
2689 | assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2" ); |
2690 | |
2691 | // This is the index of the first element of the VectorWidth-bit chunk |
2692 | // we want. Since ElemsPerChunk is a power of 2 just need to clear bits. |
2693 | IdxVal &= ~(ElemsPerChunk - 1); |
2694 | |
2695 | // If the input is a buildvector just emit a smaller one. |
2696 | if (Vec.getOpcode() == ISD::BUILD_VECTOR) |
2697 | return DAG.getBuildVector(VT: ResultVT, DL, |
2698 | Ops: Vec->ops().slice(N: IdxVal, M: ElemsPerChunk)); |
2699 | |
2700 | SDValue VecIdx = DAG.getIntPtrConstant(Val: IdxVal, DL); |
2701 | return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: ResultVT, N1: Vec, N2: VecIdx); |
2702 | } |
2703 | |
2704 | // Helper to recursively truncate vector elements in half with NARROW_U. DstVT |
2705 | // is the expected destination value type after recursion. In is the initial |
2706 | // input. Note that the input should have enough leading zero bits to prevent |
2707 | // NARROW_U from saturating results. |
2708 | static SDValue truncateVectorWithNARROW(EVT DstVT, SDValue In, const SDLoc &DL, |
2709 | SelectionDAG &DAG) { |
2710 | EVT SrcVT = In.getValueType(); |
2711 | |
2712 | // No truncation required, we might get here due to recursive calls. |
2713 | if (SrcVT == DstVT) |
2714 | return In; |
2715 | |
2716 | unsigned SrcSizeInBits = SrcVT.getSizeInBits(); |
2717 | unsigned NumElems = SrcVT.getVectorNumElements(); |
2718 | if (!isPowerOf2_32(Value: NumElems)) |
2719 | return SDValue(); |
2720 | assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation" ); |
2721 | assert(SrcSizeInBits > DstVT.getSizeInBits() && "Illegal truncation" ); |
2722 | |
2723 | LLVMContext &Ctx = *DAG.getContext(); |
2724 | EVT PackedSVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: SrcVT.getScalarSizeInBits() / 2); |
2725 | |
2726 | // Narrow to the largest type possible: |
2727 | // vXi64/vXi32 -> i16x8.narrow_i32x4_u and vXi16 -> i8x16.narrow_i16x8_u. |
2728 | EVT InVT = MVT::i16, OutVT = MVT::i8; |
2729 | if (SrcVT.getScalarSizeInBits() > 16) { |
2730 | InVT = MVT::i32; |
2731 | OutVT = MVT::i16; |
2732 | } |
2733 | unsigned SubSizeInBits = SrcSizeInBits / 2; |
2734 | InVT = EVT::getVectorVT(Context&: Ctx, VT: InVT, NumElements: SubSizeInBits / InVT.getSizeInBits()); |
2735 | OutVT = EVT::getVectorVT(Context&: Ctx, VT: OutVT, NumElements: SubSizeInBits / OutVT.getSizeInBits()); |
2736 | |
2737 | // Split lower/upper subvectors. |
2738 | SDValue Lo = extractSubVector(Vec: In, IdxVal: 0, DAG, DL, VectorWidth: SubSizeInBits); |
2739 | SDValue Hi = extractSubVector(Vec: In, IdxVal: NumElems / 2, DAG, DL, VectorWidth: SubSizeInBits); |
2740 | |
2741 | // 256bit -> 128bit truncate - Narrow lower/upper 128-bit subvectors. |
2742 | if (SrcVT.is256BitVector() && DstVT.is128BitVector()) { |
2743 | Lo = DAG.getBitcast(VT: InVT, V: Lo); |
2744 | Hi = DAG.getBitcast(VT: InVT, V: Hi); |
2745 | SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::NARROW_U, DL, VT: OutVT, N1: Lo, N2: Hi); |
2746 | return DAG.getBitcast(VT: DstVT, V: Res); |
2747 | } |
2748 | |
2749 | // Recursively narrow lower/upper subvectors, concat result and narrow again. |
2750 | EVT PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems / 2); |
2751 | Lo = truncateVectorWithNARROW(DstVT: PackedVT, In: Lo, DL, DAG); |
2752 | Hi = truncateVectorWithNARROW(DstVT: PackedVT, In: Hi, DL, DAG); |
2753 | |
2754 | PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems); |
2755 | SDValue Res = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: PackedVT, N1: Lo, N2: Hi); |
2756 | return truncateVectorWithNARROW(DstVT, In: Res, DL, DAG); |
2757 | } |
2758 | |
2759 | static SDValue performTruncateCombine(SDNode *N, |
2760 | TargetLowering::DAGCombinerInfo &DCI) { |
2761 | auto &DAG = DCI.DAG; |
2762 | |
2763 | SDValue In = N->getOperand(Num: 0); |
2764 | EVT InVT = In.getValueType(); |
2765 | if (!InVT.isSimple()) |
2766 | return SDValue(); |
2767 | |
2768 | EVT OutVT = N->getValueType(ResNo: 0); |
2769 | if (!OutVT.isVector()) |
2770 | return SDValue(); |
2771 | |
2772 | EVT OutSVT = OutVT.getVectorElementType(); |
2773 | EVT InSVT = InVT.getVectorElementType(); |
2774 | // Currently only cover truncate to v16i8 or v8i16. |
2775 | if (!((InSVT == MVT::i16 || InSVT == MVT::i32 || InSVT == MVT::i64) && |
2776 | (OutSVT == MVT::i8 || OutSVT == MVT::i16) && OutVT.is128BitVector())) |
2777 | return SDValue(); |
2778 | |
2779 | SDLoc DL(N); |
2780 | APInt Mask = APInt::getLowBitsSet(numBits: InVT.getScalarSizeInBits(), |
2781 | loBitsSet: OutVT.getScalarSizeInBits()); |
2782 | In = DAG.getNode(Opcode: ISD::AND, DL, VT: InVT, N1: In, N2: DAG.getConstant(Val: Mask, DL, VT: InVT)); |
2783 | return truncateVectorWithNARROW(DstVT: OutVT, In, DL, DAG); |
2784 | } |
2785 | |
2786 | static SDValue performBitcastCombine(SDNode *N, |
2787 | TargetLowering::DAGCombinerInfo &DCI) { |
2788 | auto &DAG = DCI.DAG; |
2789 | SDLoc DL(N); |
2790 | SDValue Src = N->getOperand(Num: 0); |
2791 | EVT VT = N->getValueType(ResNo: 0); |
2792 | EVT SrcVT = Src.getValueType(); |
2793 | |
2794 | // bitcast <N x i1> to iN |
2795 | // ==> bitmask |
2796 | if (DCI.isBeforeLegalize() && VT.isScalarInteger() && |
2797 | SrcVT.isFixedLengthVector() && SrcVT.getScalarType() == MVT::i1) { |
2798 | unsigned NumElts = SrcVT.getVectorNumElements(); |
2799 | if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16) |
2800 | return SDValue(); |
2801 | EVT Width = MVT::getIntegerVT(BitWidth: 128 / NumElts); |
2802 | return DAG.getZExtOrTrunc( |
2803 | DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32, |
2804 | {DAG.getConstant(Intrinsic::wasm_bitmask, DL, MVT::i32), |
2805 | DAG.getSExtOrTrunc(N->getOperand(0), DL, |
2806 | SrcVT.changeVectorElementType(Width))}), |
2807 | DL, VT); |
2808 | } |
2809 | |
2810 | return SDValue(); |
2811 | } |
2812 | |
2813 | static SDValue performSETCCCombine(SDNode *N, |
2814 | TargetLowering::DAGCombinerInfo &DCI) { |
2815 | auto &DAG = DCI.DAG; |
2816 | |
2817 | SDValue LHS = N->getOperand(Num: 0); |
2818 | SDValue RHS = N->getOperand(Num: 1); |
2819 | ISD::CondCode Cond = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get(); |
2820 | SDLoc DL(N); |
2821 | EVT VT = N->getValueType(ResNo: 0); |
2822 | |
2823 | // setcc (iN (bitcast (vNi1 X))), 0, ne |
2824 | // ==> any_true (vNi1 X) |
2825 | // setcc (iN (bitcast (vNi1 X))), 0, eq |
2826 | // ==> xor (any_true (vNi1 X)), -1 |
2827 | // setcc (iN (bitcast (vNi1 X))), -1, eq |
2828 | // ==> all_true (vNi1 X) |
2829 | // setcc (iN (bitcast (vNi1 X))), -1, ne |
2830 | // ==> xor (all_true (vNi1 X)), -1 |
2831 | if (DCI.isBeforeLegalize() && VT.isScalarInteger() && |
2832 | (Cond == ISD::SETEQ || Cond == ISD::SETNE) && |
2833 | (isNullConstant(V: RHS) || isAllOnesConstant(V: RHS)) && |
2834 | LHS->getOpcode() == ISD::BITCAST) { |
2835 | EVT FromVT = LHS->getOperand(Num: 0).getValueType(); |
2836 | if (FromVT.isFixedLengthVector() && |
2837 | FromVT.getVectorElementType() == MVT::i1) { |
2838 | int Intrin = isNullConstant(RHS) ? Intrinsic::wasm_anytrue |
2839 | : Intrinsic::wasm_alltrue; |
2840 | unsigned NumElts = FromVT.getVectorNumElements(); |
2841 | if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16) |
2842 | return SDValue(); |
2843 | EVT Width = MVT::getIntegerVT(BitWidth: 128 / NumElts); |
2844 | SDValue Ret = DAG.getZExtOrTrunc( |
2845 | DAG.getNode( |
2846 | ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32, |
2847 | {DAG.getConstant(Intrin, DL, MVT::i32), |
2848 | DAG.getSExtOrTrunc(LHS->getOperand(0), DL, |
2849 | FromVT.changeVectorElementType(Width))}), |
2850 | DL, MVT::i1); |
2851 | if ((isNullConstant(V: RHS) && (Cond == ISD::SETEQ)) || |
2852 | (isAllOnesConstant(V: RHS) && (Cond == ISD::SETNE))) { |
2853 | Ret = DAG.getNOT(DL, Ret, MVT::i1); |
2854 | } |
2855 | return DAG.getZExtOrTrunc(Op: Ret, DL, VT); |
2856 | } |
2857 | } |
2858 | |
2859 | return SDValue(); |
2860 | } |
2861 | |
2862 | SDValue |
2863 | WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N, |
2864 | DAGCombinerInfo &DCI) const { |
2865 | switch (N->getOpcode()) { |
2866 | default: |
2867 | return SDValue(); |
2868 | case ISD::BITCAST: |
2869 | return performBitcastCombine(N, DCI); |
2870 | case ISD::SETCC: |
2871 | return performSETCCCombine(N, DCI); |
2872 | case ISD::VECTOR_SHUFFLE: |
2873 | return performVECTOR_SHUFFLECombine(N, DCI); |
2874 | case ISD::SIGN_EXTEND: |
2875 | case ISD::ZERO_EXTEND: |
2876 | return performVectorExtendCombine(N, DCI); |
2877 | case ISD::UINT_TO_FP: |
2878 | case ISD::SINT_TO_FP: |
2879 | return performVectorExtendToFPCombine(N, DCI); |
2880 | case ISD::FP_TO_SINT_SAT: |
2881 | case ISD::FP_TO_UINT_SAT: |
2882 | case ISD::FP_ROUND: |
2883 | case ISD::CONCAT_VECTORS: |
2884 | return performVectorTruncZeroCombine(N, DCI); |
2885 | case ISD::TRUNCATE: |
2886 | return performTruncateCombine(N, DCI); |
2887 | } |
2888 | } |
2889 | |