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/MachineInstrBuilder.h"
24#include "llvm/CodeGen/MachineJumpTableInfo.h"
25#include "llvm/CodeGen/MachineModuleInfo.h"
26#include "llvm/CodeGen/MachineRegisterInfo.h"
27#include "llvm/CodeGen/SDPatternMatch.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/Support/ErrorHandling.h"
36#include "llvm/Support/KnownBits.h"
37#include "llvm/Support/MathExtras.h"
38#include "llvm/Target/TargetOptions.h"
39using namespace llvm;
40
41#define DEBUG_TYPE "wasm-lower"
42
43WebAssemblyTargetLowering::WebAssemblyTargetLowering(
44 const TargetMachine &TM, const WebAssemblySubtarget &STI)
45 : TargetLowering(TM, STI), Subtarget(&STI) {
46 auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32;
47
48 // Set the load count for memcmp expand optimization
49 MaxLoadsPerMemcmp = 8;
50 MaxLoadsPerMemcmpOptSize = 4;
51
52 // Booleans always contain 0 or 1.
53 setBooleanContents(ZeroOrOneBooleanContent);
54 // Except in SIMD vectors
55 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
56 // We don't know the microarchitecture here, so just reduce register pressure.
57 setSchedulingPreference(Sched::RegPressure);
58 // Tell ISel that we have a stack pointer.
59 setStackPointerRegisterToSaveRestore(
60 Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32);
61 // Set up the register classes.
62 addRegisterClass(VT: MVT::i32, RC: &WebAssembly::I32RegClass);
63 addRegisterClass(VT: MVT::i64, RC: &WebAssembly::I64RegClass);
64 addRegisterClass(VT: MVT::f32, RC: &WebAssembly::F32RegClass);
65 addRegisterClass(VT: MVT::f64, RC: &WebAssembly::F64RegClass);
66 if (Subtarget->hasSIMD128()) {
67 addRegisterClass(VT: MVT::v16i8, RC: &WebAssembly::V128RegClass);
68 addRegisterClass(VT: MVT::v8i16, RC: &WebAssembly::V128RegClass);
69 addRegisterClass(VT: MVT::v4i32, RC: &WebAssembly::V128RegClass);
70 addRegisterClass(VT: MVT::v4f32, RC: &WebAssembly::V128RegClass);
71 addRegisterClass(VT: MVT::v2i64, RC: &WebAssembly::V128RegClass);
72 addRegisterClass(VT: MVT::v2f64, RC: &WebAssembly::V128RegClass);
73 }
74 if (Subtarget->hasFP16()) {
75 addRegisterClass(VT: MVT::v8f16, RC: &WebAssembly::V128RegClass);
76 }
77 if (Subtarget->hasReferenceTypes()) {
78 addRegisterClass(VT: MVT::externref, RC: &WebAssembly::EXTERNREFRegClass);
79 addRegisterClass(VT: MVT::funcref, RC: &WebAssembly::FUNCREFRegClass);
80 if (Subtarget->hasExceptionHandling()) {
81 addRegisterClass(VT: MVT::exnref, RC: &WebAssembly::EXNREFRegClass);
82 }
83 }
84 // Compute derived properties from the register classes.
85 computeRegisterProperties(TRI: Subtarget->getRegisterInfo());
86
87 // Transform loads and stores to pointers in address space 1 to loads and
88 // stores to WebAssembly global variables, outside linear memory.
89 for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) {
90 setOperationAction(Op: ISD::LOAD, VT: T, Action: Custom);
91 setOperationAction(Op: ISD::STORE, VT: T, Action: Custom);
92 }
93 if (Subtarget->hasSIMD128()) {
94 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
95 MVT::v2f64}) {
96 setOperationAction(Op: ISD::LOAD, VT: T, Action: Custom);
97 setOperationAction(Op: ISD::STORE, VT: T, Action: Custom);
98 }
99 }
100 if (Subtarget->hasFP16()) {
101 setOperationAction(Op: ISD::LOAD, VT: MVT::v8f16, Action: Custom);
102 setOperationAction(Op: ISD::STORE, VT: MVT::v8f16, Action: Custom);
103 }
104 if (Subtarget->hasReferenceTypes()) {
105 // We need custom load and store lowering for both externref, funcref and
106 // Other. The MVT::Other here represents tables of reference types.
107 for (auto T : {MVT::externref, MVT::funcref, MVT::Other}) {
108 setOperationAction(Op: ISD::LOAD, VT: T, Action: Custom);
109 setOperationAction(Op: ISD::STORE, VT: T, Action: Custom);
110 }
111 }
112
113 setOperationAction(Op: ISD::GlobalAddress, VT: MVTPtr, Action: Custom);
114 setOperationAction(Op: ISD::GlobalTLSAddress, VT: MVTPtr, Action: Custom);
115 setOperationAction(Op: ISD::ExternalSymbol, VT: MVTPtr, Action: Custom);
116 setOperationAction(Op: ISD::JumpTable, VT: MVTPtr, Action: Custom);
117 setOperationAction(Op: ISD::BlockAddress, VT: MVTPtr, Action: Custom);
118 setOperationAction(Op: ISD::BRIND, VT: MVT::Other, Action: Custom);
119 setOperationAction(Op: ISD::CLEAR_CACHE, VT: MVT::Other, Action: Custom);
120
121 // Take the default expansion for va_arg, va_copy, and va_end. There is no
122 // default action for va_start, so we do that custom.
123 setOperationAction(Op: ISD::VASTART, VT: MVT::Other, Action: Custom);
124 setOperationAction(Op: ISD::VAARG, VT: MVT::Other, Action: Expand);
125 setOperationAction(Op: ISD::VACOPY, VT: MVT::Other, Action: Expand);
126 setOperationAction(Op: ISD::VAEND, VT: MVT::Other, Action: Expand);
127
128 for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64, MVT::v8f16}) {
129 if (!Subtarget->hasFP16() && T == MVT::v8f16) {
130 continue;
131 }
132 // Don't expand the floating-point types to constant pools.
133 setOperationAction(Op: ISD::ConstantFP, VT: T, Action: Legal);
134 // Expand floating-point comparisons.
135 for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
136 ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
137 setCondCodeAction(CCs: CC, VT: T, Action: Expand);
138 // Expand floating-point library function operators.
139 for (auto Op : {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FMA})
140 setOperationAction(Op, VT: T, Action: Expand);
141 // Expand vector FREM, but use a libcall rather than an expansion for scalar
142 if (MVT(T).isVector())
143 setOperationAction(Op: ISD::FREM, VT: T, Action: Expand);
144 else
145 setOperationAction(Op: ISD::FREM, VT: T, Action: LibCall);
146 // Note supported floating-point library function operators that otherwise
147 // default to expand.
148 for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT,
149 ISD::FRINT, ISD::FROUNDEVEN})
150 setOperationAction(Op, VT: T, Action: Legal);
151 // Support minimum and maximum, which otherwise default to expand.
152 setOperationAction(Op: ISD::FMINIMUM, VT: T, Action: Legal);
153 setOperationAction(Op: ISD::FMAXIMUM, VT: T, Action: Legal);
154 if (Subtarget->hasSIMD128() && MVT(T).isVector()) {
155 setOperationAction(Op: ISD::PSEUDO_FMIN, VT: T, Action: Legal);
156 setOperationAction(Op: ISD::PSEUDO_FMAX, VT: T, Action: Legal);
157 }
158 // When experimental v8f16 support is enabled these instructions don't need
159 // to be expanded.
160 if (T != MVT::v8f16) {
161 setOperationAction(Op: ISD::FP16_TO_FP, VT: T, Action: Expand);
162 setOperationAction(Op: ISD::FP_TO_FP16, VT: T, Action: Expand);
163 }
164 if (Subtarget->hasFP16() && T == MVT::f32) {
165 setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: T, MemVT: MVT::f16, Action: Legal);
166 setTruncStoreAction(ValVT: T, MemVT: MVT::f16, Action: Legal);
167 } else {
168 setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: T, MemVT: MVT::f16, Action: Expand);
169 setTruncStoreAction(ValVT: T, MemVT: MVT::f16, Action: Expand);
170 }
171 }
172
173 // Expand unavailable integer operations.
174 for (auto Op :
175 {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU,
176 ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS,
177 ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) {
178 for (auto T : {MVT::i32, MVT::i64})
179 setOperationAction(Op, VT: T, Action: Expand);
180 if (Subtarget->hasSIMD128())
181 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
182 setOperationAction(Op, VT: T, Action: Expand);
183 }
184
185 if (Subtarget->hasWideArithmetic()) {
186 setOperationAction(Op: ISD::ADD, VT: MVT::i128, Action: Custom);
187 setOperationAction(Op: ISD::SUB, VT: MVT::i128, Action: Custom);
188 setOperationAction(Op: ISD::SMUL_LOHI, VT: MVT::i64, Action: Custom);
189 setOperationAction(Op: ISD::UMUL_LOHI, VT: MVT::i64, Action: Custom);
190 setOperationAction(Op: ISD::UADDO, VT: MVT::i64, Action: Custom);
191 }
192
193 if (Subtarget->hasNontrappingFPToInt())
194 for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT})
195 for (auto T : {MVT::i32, MVT::i64})
196 setOperationAction(Op, VT: T, Action: Custom);
197
198 if (Subtarget->hasRelaxedSIMD()) {
199 setOperationAction(
200 Ops: {ISD::FMINNUM, ISD::FMINIMUMNUM, ISD::FMAXNUM, ISD::FMAXIMUMNUM},
201 VTs: {MVT::v4f32, MVT::v2f64}, Action: Custom);
202 }
203
204 // Combine expands these operations, because wasi-libc and emscripten do not
205 // yet have the dedicated libcalls.
206 setTargetDAGCombine(
207 {ISD::FMINIMUM, ISD::FMAXIMUM, ISD::FMINIMUMNUM, ISD::FMAXIMUMNUM});
208
209 // SIMD-specific configuration
210 if (Subtarget->hasSIMD128()) {
211
212 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
213
214 // Combine wide-vector muls, with extend inputs, to extmul_half.
215 setTargetDAGCombine(ISD::MUL);
216 setTargetDAGCombine(ISD::SHL);
217
218 // Combine vector mask reductions into alltrue/anytrue
219 setTargetDAGCombine(ISD::SETCC);
220
221 // Convert vector to integer bitcasts to bitmask
222 setTargetDAGCombine(ISD::BITCAST);
223
224 // Hoist bitcasts out of shuffles
225 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
226
227 // Combine extends of extract_subvectors into widening ops
228 setTargetDAGCombine({ISD::SIGN_EXTEND, ISD::ZERO_EXTEND});
229
230 // Combine int_to_fp or fp_extend of extract_vectors and vice versa into
231 // conversions ops
232 setTargetDAGCombine({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_EXTEND,
233 ISD::EXTRACT_SUBVECTOR});
234
235 // Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa
236 // into conversion ops
237 setTargetDAGCombine({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT,
238 ISD::FP_TO_SINT, ISD::FP_TO_UINT, ISD::FP_ROUND,
239 ISD::CONCAT_VECTORS});
240
241 setTargetDAGCombine(ISD::TRUNCATE);
242
243 // Support saturating add/sub for i8x16 and i16x8
244 for (auto Op : {ISD::SADDSAT, ISD::UADDSAT, ISD::SSUBSAT, ISD::USUBSAT})
245 for (auto T : {MVT::v16i8, MVT::v8i16})
246 setOperationAction(Op, VT: T, Action: Legal);
247
248 // Support integer abs
249 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
250 setOperationAction(Op: ISD::ABS, VT: T, Action: Legal);
251
252 // Custom lower BUILD_VECTORs to minimize number of replace_lanes
253 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
254 MVT::v2f64})
255 setOperationAction(Op: ISD::BUILD_VECTOR, VT: T, Action: Custom);
256
257 if (Subtarget->hasFP16()) {
258 setOperationAction(Op: ISD::BUILD_VECTOR, VT: MVT::f16, Action: Custom);
259 setOperationAction(Op: ISD::INSERT_VECTOR_ELT, VT: MVT::f16, Action: Custom);
260 setOperationAction(Op: ISD::FP_ROUND, VT: MVT::v4f16, Action: Custom);
261 }
262
263 // We have custom shuffle lowering to expose the shuffle mask
264 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
265 MVT::v2f64})
266 setOperationAction(Op: ISD::VECTOR_SHUFFLE, VT: T, Action: Custom);
267
268 if (Subtarget->hasFP16())
269 setOperationAction(Op: ISD::VECTOR_SHUFFLE, VT: MVT::v8f16, Action: Custom);
270
271 // Support splatting
272 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
273 MVT::v2f64})
274 setOperationAction(Op: ISD::SPLAT_VECTOR, VT: T, Action: Legal);
275
276 setOperationAction(Ops: ISD::AVGCEILU, VTs: {MVT::v8i16, MVT::v16i8}, Action: Legal);
277
278 // Custom lowering since wasm shifts must have a scalar shift amount
279 for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL})
280 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
281 setOperationAction(Op, VT: T, Action: Custom);
282
283 // Custom lower lane accesses to expand out variable indices
284 for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT})
285 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
286 MVT::v2f64})
287 setOperationAction(Op, VT: T, Action: Custom);
288
289 // There is no i8x16.mul instruction
290 setOperationAction(Op: ISD::MUL, VT: MVT::v16i8, Action: Expand);
291
292 // Expand integer operations supported for scalars but not SIMD
293 for (auto Op :
294 {ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR})
295 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
296 setOperationAction(Op, VT: T, Action: Expand);
297
298 // But we do have integer min and max operations
299 for (auto Op : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
300 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
301 setOperationAction(Op, VT: T, Action: Legal);
302
303 // And we have popcnt for i8x16. It can be used to expand ctlz/cttz.
304 setOperationAction(Op: ISD::CTPOP, VT: MVT::v16i8, Action: Legal);
305 setOperationAction(Op: ISD::CTLZ, VT: MVT::v16i8, Action: Expand);
306 setOperationAction(Op: ISD::CTTZ, VT: MVT::v16i8, Action: Expand);
307
308 // Custom lower bit counting operations for other types to scalarize them.
309 for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP})
310 for (auto T : {MVT::v8i16, MVT::v4i32, MVT::v2i64})
311 setOperationAction(Op, VT: T, Action: Custom);
312
313 // Expand float operations supported for scalars but not SIMD
314 for (auto Op : {ISD::FCOPYSIGN, ISD::FLOG, ISD::FLOG2, ISD::FLOG10,
315 ISD::FEXP, ISD::FEXP2, ISD::FEXP10})
316 for (auto T : {MVT::v4f32, MVT::v2f64})
317 setOperationAction(Op, VT: T, Action: Expand);
318
319 // Unsigned comparison operations are unavailable for i64x2 vectors.
320 for (auto CC : {ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE})
321 setCondCodeAction(CCs: CC, VT: MVT::v2i64, Action: Custom);
322
323 // 64x2 conversions are not in the spec
324 for (auto Op :
325 {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT})
326 for (auto T : {MVT::v2i64, MVT::v2f64})
327 setOperationAction(Op, VT: T, Action: Expand);
328
329 // But saturating fp_to_int conversions are
330 for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT}) {
331 setOperationAction(Op, VT: MVT::v4i32, Action: Custom);
332 if (Subtarget->hasFP16()) {
333 setOperationAction(Op, VT: MVT::v8i16, Action: Custom);
334 }
335 }
336
337 // Support vector extending
338 for (auto T : MVT::integer_fixedlen_vector_valuetypes()) {
339 setOperationAction(Op: ISD::ANY_EXTEND_VECTOR_INREG, VT: T, Action: Custom);
340 setOperationAction(Op: ISD::SIGN_EXTEND_VECTOR_INREG, VT: T, Action: Custom);
341 setOperationAction(Op: ISD::ZERO_EXTEND_VECTOR_INREG, VT: T, Action: Custom);
342 }
343
344 if (Subtarget->hasFP16()) {
345 setOperationAction(Op: ISD::FMA, VT: MVT::v8f16, Action: Legal);
346 }
347
348 if (Subtarget->hasRelaxedSIMD()) {
349 setOperationAction(Op: ISD::FMULADD, VT: MVT::v4f32, Action: Legal);
350 setOperationAction(Op: ISD::FMULADD, VT: MVT::v2f64, Action: Legal);
351 }
352
353 // Partial MLA reductions.
354 for (auto Op : {ISD::PARTIAL_REDUCE_SMLA, ISD::PARTIAL_REDUCE_UMLA}) {
355 setPartialReduceMLAAction(Opc: Op, AccVT: MVT::v4i32, InputVT: MVT::v16i8, Action: Legal);
356 setPartialReduceMLAAction(Opc: Op, AccVT: MVT::v4i32, InputVT: MVT::v8i16, Action: Legal);
357 }
358 }
359
360 // As a special case, these operators use the type to mean the type to
361 // sign-extend from.
362 setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: MVT::i1, Action: Expand);
363 if (!Subtarget->hasSignExt()) {
364 // Sign extends are legal only when extending a vector extract
365 auto Action = Subtarget->hasSIMD128() ? Custom : Expand;
366 for (auto T : {MVT::i8, MVT::i16, MVT::i32})
367 setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: T, Action);
368 }
369 for (auto T : MVT::integer_fixedlen_vector_valuetypes())
370 setOperationAction(Op: ISD::SIGN_EXTEND_INREG, VT: T, Action: Expand);
371
372 // Dynamic stack allocation: use the default expansion.
373 setOperationAction(Op: ISD::STACKSAVE, VT: MVT::Other, Action: Expand);
374 setOperationAction(Op: ISD::STACKRESTORE, VT: MVT::Other, Action: Expand);
375 setOperationAction(Op: ISD::DYNAMIC_STACKALLOC, VT: MVTPtr, Action: Expand);
376
377 setOperationAction(Op: ISD::FrameIndex, VT: MVT::i32, Action: Custom);
378 setOperationAction(Op: ISD::FrameIndex, VT: MVT::i64, Action: Custom);
379 setOperationAction(Op: ISD::CopyToReg, VT: MVT::Other, Action: Custom);
380
381 // Expand these forms; we pattern-match the forms that we can handle in isel.
382 for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64})
383 for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
384 setOperationAction(Op, VT: T, Action: Expand);
385
386 if (Subtarget->hasReferenceTypes())
387 for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
388 for (auto T : {MVT::externref, MVT::funcref})
389 setOperationAction(Op, VT: T, Action: Expand);
390
391 // There is no vector conditional select instruction
392 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
393 MVT::v2f64, MVT::v8f16})
394 setOperationAction(Op: ISD::SELECT_CC, VT: T, Action: Expand);
395
396 // We have custom switch handling.
397 setOperationAction(Op: ISD::BR_JT, VT: MVT::Other, Action: Custom);
398
399 // WebAssembly doesn't have:
400 // - Floating-point extending loads.
401 // - Floating-point truncating stores.
402 // - i1 extending loads.
403 // - truncating SIMD stores and most extending loads
404 setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand);
405 setTruncStoreAction(ValVT: MVT::f64, MemVT: MVT::f32, Action: Expand);
406 for (auto T : MVT::integer_valuetypes())
407 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
408 setLoadExtAction(ExtType: Ext, ValVT: T, MemVT: MVT::i1, Action: Promote);
409 if (Subtarget->hasSIMD128()) {
410 for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32,
411 MVT::v2f64}) {
412 for (auto MemT : MVT::fixedlen_vector_valuetypes()) {
413 if (MVT(T) != MemT) {
414 setTruncStoreAction(ValVT: T, MemVT: MemT, Action: Expand);
415 for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
416 setLoadExtAction(ExtType: Ext, ValVT: T, MemVT: MemT, Action: Expand);
417 }
418 }
419 }
420 // But some vector extending loads are legal
421 for (auto Ext : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) {
422 setLoadExtAction(ExtType: Ext, ValVT: MVT::v8i16, MemVT: MVT::v8i8, Action: Legal);
423 setLoadExtAction(ExtType: Ext, ValVT: MVT::v4i32, MemVT: MVT::v4i16, Action: Legal);
424 setLoadExtAction(ExtType: Ext, ValVT: MVT::v2i64, MemVT: MVT::v2i32, Action: Legal);
425 }
426 setLoadExtAction(ExtType: ISD::EXTLOAD, ValVT: MVT::v2f64, MemVT: MVT::v2f32, Action: Legal);
427 }
428
429 // Don't do anything clever with build_pairs
430 setOperationAction(Op: ISD::BUILD_PAIR, VT: MVT::i64, Action: Expand);
431
432 // Trap lowers to wasm unreachable
433 setOperationAction(Op: ISD::TRAP, VT: MVT::Other, Action: Legal);
434 setOperationAction(Op: ISD::DEBUGTRAP, VT: MVT::Other, Action: Legal);
435
436 // Exception handling intrinsics
437 setOperationAction(Op: ISD::INTRINSIC_WO_CHAIN, VT: MVT::Other, Action: Custom);
438 setOperationAction(Op: ISD::INTRINSIC_W_CHAIN, VT: MVT::Other, Action: Custom);
439 setOperationAction(Op: ISD::INTRINSIC_VOID, VT: MVT::Other, Action: Custom);
440
441 setMaxAtomicSizeInBitsSupported(64);
442
443 // Always convert switches to br_tables unless there is only one case, which
444 // is equivalent to a simple branch. This reduces code size for wasm, and we
445 // defer possible jump table optimizations to the VM.
446 setMinimumJumpTableEntries(2);
447}
448
449TargetLowering::AtomicExpansionKind
450WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR(
451 const AtomicRMWInst *AI) const {
452 // We have wasm instructions for these
453 switch (AI->getOperation()) {
454 case AtomicRMWInst::Add:
455 case AtomicRMWInst::Sub:
456 case AtomicRMWInst::And:
457 case AtomicRMWInst::Or:
458 case AtomicRMWInst::Xor:
459 case AtomicRMWInst::Xchg:
460 return AtomicExpansionKind::None;
461 default:
462 break;
463 }
464 return AtomicExpansionKind::CmpXChg;
465}
466
467bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
468 // Implementation copied from X86TargetLowering.
469 unsigned Opc = VecOp.getOpcode();
470
471 // Assume target opcodes can't be scalarized.
472 // TODO - do we have any exceptions?
473 if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opcode: Opc))
474 return false;
475
476 // If the vector op is not supported, try to convert to scalar.
477 EVT VecVT = VecOp.getValueType();
478 if (!isOperationLegalOrCustomOrPromote(Op: Opc, VT: VecVT))
479 return true;
480
481 // If the vector op is supported, but the scalar op is not, the transform may
482 // not be worthwhile.
483 EVT ScalarVT = VecVT.getScalarType();
484 return isOperationLegalOrCustomOrPromote(Op: Opc, VT: ScalarVT);
485}
486
487FastISel *WebAssemblyTargetLowering::createFastISel(
488 FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo,
489 const LibcallLoweringInfo *LibcallLowering) const {
490 return WebAssembly::createFastISel(funcInfo&: FuncInfo, libInfo: LibInfo, libcallLowering: LibcallLowering);
491}
492
493MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
494 EVT VT) const {
495 unsigned BitWidth = NextPowerOf2(A: VT.getSizeInBits() - 1);
496 if (BitWidth > 1 && BitWidth < 8)
497 BitWidth = 8;
498
499 if (BitWidth > 64) {
500 // The shift will be lowered to a libcall, and compiler-rt libcalls expect
501 // the count to be an i32.
502 BitWidth = 32;
503 assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
504 "32-bit shift counts ought to be enough for anyone");
505 }
506
507 MVT Result = MVT::getIntegerVT(BitWidth);
508 assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
509 "Unable to represent scalar shift amount type");
510 return Result;
511}
512
513// Lower an fp-to-int conversion operator from the LLVM opcode, which has an
514// undefined result on invalid/overflow, to the WebAssembly opcode, which
515// traps on invalid/overflow.
516static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL,
517 MachineBasicBlock *BB,
518 const TargetInstrInfo &TII,
519 bool IsUnsigned, bool Int64,
520 bool Float64, unsigned LoweredOpcode) {
521 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
522
523 Register OutReg = MI.getOperand(i: 0).getReg();
524 Register InReg = MI.getOperand(i: 1).getReg();
525
526 unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32;
527 unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32;
528 unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32;
529 unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32;
530 unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32;
531 unsigned Eqz = WebAssembly::EQZ_I32;
532 unsigned And = WebAssembly::AND_I32;
533 int64_t Limit = Int64 ? INT64_MIN : INT32_MIN;
534 int64_t Substitute = IsUnsigned ? 0 : Limit;
535 double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit;
536 auto &Context = BB->getParent()->getFunction().getContext();
537 Type *Ty = Float64 ? Type::getDoubleTy(C&: Context) : Type::getFloatTy(C&: Context);
538
539 const BasicBlock *LLVMBB = BB->getBasicBlock();
540 MachineFunction *F = BB->getParent();
541 MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
542 MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
543 MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
544
545 MachineFunction::iterator It = ++BB->getIterator();
546 F->insert(MBBI: It, MBB: FalseMBB);
547 F->insert(MBBI: It, MBB: TrueMBB);
548 F->insert(MBBI: It, MBB: DoneMBB);
549
550 // Transfer the remainder of BB and its successor edges to DoneMBB.
551 DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end());
552 DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB);
553
554 BB->addSuccessor(Succ: TrueMBB);
555 BB->addSuccessor(Succ: FalseMBB);
556 TrueMBB->addSuccessor(Succ: DoneMBB);
557 FalseMBB->addSuccessor(Succ: DoneMBB);
558
559 unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg;
560 Tmp0 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg));
561 Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg));
562 CmpReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
563 EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
564 FalseReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg));
565 TrueReg = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: OutReg));
566
567 MI.eraseFromParent();
568 // For signed numbers, we can do a single comparison to determine whether
569 // fabs(x) is within range.
570 if (IsUnsigned) {
571 Tmp0 = InReg;
572 } else {
573 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Abs), DestReg: Tmp0).addReg(RegNo: InReg);
574 }
575 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1)
576 .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: CmpVal)));
577 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: LT), DestReg: CmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1);
578
579 // For unsigned numbers, we have to do a separate comparison with zero.
580 if (IsUnsigned) {
581 Tmp1 = MRI.createVirtualRegister(RegClass: MRI.getRegClass(Reg: InReg));
582 Register SecondCmpReg =
583 MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
584 Register AndReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
585 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: FConst), DestReg: Tmp1)
586 .addFPImm(Val: cast<ConstantFP>(Val: ConstantFP::get(Ty, V: 0.0)));
587 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: GE), DestReg: SecondCmpReg).addReg(RegNo: Tmp0).addReg(RegNo: Tmp1);
588 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: And), DestReg: AndReg).addReg(RegNo: CmpReg).addReg(RegNo: SecondCmpReg);
589 CmpReg = AndReg;
590 }
591
592 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).addReg(RegNo: CmpReg);
593
594 // Create the CFG diamond to select between doing the conversion or using
595 // the substitute value.
596 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: TrueMBB).addReg(RegNo: EqzReg);
597 BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: LoweredOpcode), DestReg: FalseReg).addReg(RegNo: InReg);
598 BuildMI(BB: FalseMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB);
599 BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: IConst), DestReg: TrueReg).addImm(Val: Substitute);
600 BuildMI(BB&: *DoneMBB, I: DoneMBB->begin(), MIMD: DL, MCID: TII.get(Opcode: TargetOpcode::PHI), DestReg: OutReg)
601 .addReg(RegNo: FalseReg)
602 .addMBB(MBB: FalseMBB)
603 .addReg(RegNo: TrueReg)
604 .addMBB(MBB: TrueMBB);
605
606 return DoneMBB;
607}
608
609// Lower a `MEMCPY` instruction into a CFG triangle around a `MEMORY_COPY`
610// instruction to handle the zero-length case.
611static MachineBasicBlock *LowerMemcpy(MachineInstr &MI, DebugLoc DL,
612 MachineBasicBlock *BB,
613 const TargetInstrInfo &TII, bool Int64) {
614 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
615
616 MachineOperand DstMem = MI.getOperand(i: 0);
617 MachineOperand SrcMem = MI.getOperand(i: 1);
618 MachineOperand Dst = MI.getOperand(i: 2);
619 MachineOperand Src = MI.getOperand(i: 3);
620 MachineOperand Len = MI.getOperand(i: 4);
621
622 // If the length is a constant, we don't actually need the check.
623 if (MachineInstr *Def = MRI.getVRegDef(Reg: Len.getReg())) {
624 if (Def->getOpcode() == WebAssembly::CONST_I32 ||
625 Def->getOpcode() == WebAssembly::CONST_I64) {
626 if (Def->getOperand(i: 1).getImm() == 0) {
627 // A zero-length memcpy is a no-op.
628 MI.eraseFromParent();
629 return BB;
630 }
631 // A non-zero-length memcpy doesn't need a zero check.
632 unsigned MemoryCopy =
633 Int64 ? WebAssembly::MEMORY_COPY_A64 : WebAssembly::MEMORY_COPY_A32;
634 BuildMI(BB&: *BB, I&: MI, MIMD: DL, MCID: TII.get(Opcode: MemoryCopy))
635 .add(MO: DstMem)
636 .add(MO: SrcMem)
637 .add(MO: Dst)
638 .add(MO: Src)
639 .add(MO: Len);
640 MI.eraseFromParent();
641 return BB;
642 }
643 }
644
645 // We're going to add an extra use to `Len` to test if it's zero; that
646 // use shouldn't be a kill, even if the original use is.
647 MachineOperand NoKillLen = Len;
648 NoKillLen.setIsKill(false);
649
650 // Decide on which `MachineInstr` opcode we're going to use.
651 unsigned Eqz = Int64 ? WebAssembly::EQZ_I64 : WebAssembly::EQZ_I32;
652 unsigned MemoryCopy =
653 Int64 ? WebAssembly::MEMORY_COPY_A64 : WebAssembly::MEMORY_COPY_A32;
654
655 // Create two new basic blocks; one for the new `memory.fill` that we can
656 // branch over, and one for the rest of the instructions after the original
657 // `memory.fill`.
658 const BasicBlock *LLVMBB = BB->getBasicBlock();
659 MachineFunction *F = BB->getParent();
660 MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
661 MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
662
663 MachineFunction::iterator It = ++BB->getIterator();
664 F->insert(MBBI: It, MBB: TrueMBB);
665 F->insert(MBBI: It, MBB: DoneMBB);
666
667 // Transfer the remainder of BB and its successor edges to DoneMBB.
668 DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end());
669 DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB);
670
671 // Connect the CFG edges.
672 BB->addSuccessor(Succ: TrueMBB);
673 BB->addSuccessor(Succ: DoneMBB);
674 TrueMBB->addSuccessor(Succ: DoneMBB);
675
676 // Create a virtual register for the `Eqz` result.
677 unsigned EqzReg;
678 EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
679
680 // Erase the original `memory.copy`.
681 MI.eraseFromParent();
682
683 // Test if `Len` is zero.
684 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).add(MO: NoKillLen);
685
686 // Insert a new `memory.copy`.
687 BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: MemoryCopy))
688 .add(MO: DstMem)
689 .add(MO: SrcMem)
690 .add(MO: Dst)
691 .add(MO: Src)
692 .add(MO: Len);
693
694 // Create the CFG triangle.
695 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: DoneMBB).addReg(RegNo: EqzReg);
696 BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB);
697
698 return DoneMBB;
699}
700
701// Lower a `MEMSET` instruction into a CFG triangle around a `MEMORY_FILL`
702// instruction to handle the zero-length case.
703static MachineBasicBlock *LowerMemset(MachineInstr &MI, DebugLoc DL,
704 MachineBasicBlock *BB,
705 const TargetInstrInfo &TII, bool Int64) {
706 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
707
708 MachineOperand Mem = MI.getOperand(i: 0);
709 MachineOperand Dst = MI.getOperand(i: 1);
710 MachineOperand Val = MI.getOperand(i: 2);
711 MachineOperand Len = MI.getOperand(i: 3);
712
713 // If the length is a constant, we don't actually need the check.
714 if (MachineInstr *Def = MRI.getVRegDef(Reg: Len.getReg())) {
715 if (Def->getOpcode() == WebAssembly::CONST_I32 ||
716 Def->getOpcode() == WebAssembly::CONST_I64) {
717 if (Def->getOperand(i: 1).getImm() == 0) {
718 // A zero-length memset is a no-op.
719 MI.eraseFromParent();
720 return BB;
721 }
722 // A non-zero-length memset doesn't need a zero check.
723 unsigned MemoryFill =
724 Int64 ? WebAssembly::MEMORY_FILL_A64 : WebAssembly::MEMORY_FILL_A32;
725 BuildMI(BB&: *BB, I&: MI, MIMD: DL, MCID: TII.get(Opcode: MemoryFill))
726 .add(MO: Mem)
727 .add(MO: Dst)
728 .add(MO: Val)
729 .add(MO: Len);
730 MI.eraseFromParent();
731 return BB;
732 }
733 }
734
735 // We're going to add an extra use to `Len` to test if it's zero; that
736 // use shouldn't be a kill, even if the original use is.
737 MachineOperand NoKillLen = Len;
738 NoKillLen.setIsKill(false);
739
740 // Decide on which `MachineInstr` opcode we're going to use.
741 unsigned Eqz = Int64 ? WebAssembly::EQZ_I64 : WebAssembly::EQZ_I32;
742 unsigned MemoryFill =
743 Int64 ? WebAssembly::MEMORY_FILL_A64 : WebAssembly::MEMORY_FILL_A32;
744
745 // Create two new basic blocks; one for the new `memory.fill` that we can
746 // branch over, and one for the rest of the instructions after the original
747 // `memory.fill`.
748 const BasicBlock *LLVMBB = BB->getBasicBlock();
749 MachineFunction *F = BB->getParent();
750 MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
751 MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(BB: LLVMBB);
752
753 MachineFunction::iterator It = ++BB->getIterator();
754 F->insert(MBBI: It, MBB: TrueMBB);
755 F->insert(MBBI: It, MBB: DoneMBB);
756
757 // Transfer the remainder of BB and its successor edges to DoneMBB.
758 DoneMBB->splice(Where: DoneMBB->begin(), Other: BB, From: std::next(x: MI.getIterator()), To: BB->end());
759 DoneMBB->transferSuccessorsAndUpdatePHIs(FromMBB: BB);
760
761 // Connect the CFG edges.
762 BB->addSuccessor(Succ: TrueMBB);
763 BB->addSuccessor(Succ: DoneMBB);
764 TrueMBB->addSuccessor(Succ: DoneMBB);
765
766 // Create a virtual register for the `Eqz` result.
767 unsigned EqzReg;
768 EqzReg = MRI.createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
769
770 // Erase the original `memory.fill`.
771 MI.eraseFromParent();
772
773 // Test if `Len` is zero.
774 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: Eqz), DestReg: EqzReg).add(MO: NoKillLen);
775
776 // Insert a new `memory.copy`.
777 BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: MemoryFill)).add(MO: Mem).add(MO: Dst).add(MO: Val).add(MO: Len);
778
779 // Create the CFG triangle.
780 BuildMI(BB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR_IF)).addMBB(MBB: DoneMBB).addReg(RegNo: EqzReg);
781 BuildMI(BB: TrueMBB, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::BR)).addMBB(MBB: DoneMBB);
782
783 return DoneMBB;
784}
785
786static MachineBasicBlock *
787LowerCallResults(MachineInstr &CallResults, DebugLoc DL, MachineBasicBlock *BB,
788 const WebAssemblySubtarget *Subtarget,
789 const TargetInstrInfo &TII) {
790 MachineInstr &CallParams = *CallResults.getPrevNode();
791 assert(CallParams.getOpcode() == WebAssembly::CALL_PARAMS);
792 assert(CallResults.getOpcode() == WebAssembly::CALL_RESULTS ||
793 CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS);
794
795 bool IsIndirect =
796 CallParams.getOperand(i: 0).isReg() || CallParams.getOperand(i: 0).isFI();
797 bool IsRetCall = CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS;
798
799 bool IsFuncrefCall = false;
800 if (IsIndirect && CallParams.getOperand(i: 0).isReg()) {
801 Register Reg = CallParams.getOperand(i: 0).getReg();
802 const MachineFunction *MF = BB->getParent();
803 const MachineRegisterInfo &MRI = MF->getRegInfo();
804 const TargetRegisterClass *TRC = MRI.getRegClass(Reg);
805 IsFuncrefCall = (TRC == &WebAssembly::FUNCREFRegClass);
806 assert(!IsFuncrefCall || Subtarget->hasReferenceTypes());
807 }
808
809 unsigned CallOp;
810 if (IsIndirect && IsRetCall) {
811 CallOp = WebAssembly::RET_CALL_INDIRECT;
812 } else if (IsIndirect) {
813 CallOp = WebAssembly::CALL_INDIRECT;
814 } else if (IsRetCall) {
815 CallOp = WebAssembly::RET_CALL;
816 } else {
817 CallOp = WebAssembly::CALL;
818 }
819
820 MachineFunction &MF = *BB->getParent();
821 const MCInstrDesc &MCID = TII.get(Opcode: CallOp);
822 MachineInstrBuilder MIB(MF, MF.CreateMachineInstr(MCID, DL));
823
824 // Move the function pointer to the end of the arguments for indirect calls
825 if (IsIndirect) {
826 auto FnPtr = CallParams.getOperand(i: 0);
827 CallParams.removeOperand(OpNo: 0);
828
829 // For funcrefs, call_indirect is done through __funcref_call_table and the
830 // funcref is always installed in slot 0 of the table, therefore instead of
831 // having the function pointer added at the end of the params list, a zero
832 // (the index in
833 // __funcref_call_table is added).
834 if (IsFuncrefCall) {
835 Register RegZero =
836 MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
837 MachineInstrBuilder MIBC0 =
838 BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::CONST_I32), DestReg: RegZero).addImm(Val: 0);
839
840 BB->insert(I: CallResults.getIterator(), M: MIBC0);
841 MachineInstrBuilder(MF, CallParams).addReg(RegNo: RegZero);
842 } else
843 CallParams.addOperand(Op: FnPtr);
844 }
845
846 for (auto Def : CallResults.defs())
847 MIB.add(MO: Def);
848
849 if (IsIndirect) {
850 // Placeholder for the type index.
851 // This gets replaced with the correct value in WebAssemblyMCInstLower.cpp
852 MIB.addImm(Val: 0);
853 // The table into which this call_indirect indexes.
854 MCSymbolWasm *Table = IsFuncrefCall
855 ? WebAssembly::getOrCreateFuncrefCallTableSymbol(
856 Ctx&: MF.getContext(), Subtarget)
857 : WebAssembly::getOrCreateFunctionTableSymbol(
858 Ctx&: MF.getContext(), Subtarget);
859 if (Subtarget->hasCallIndirectOverlong()) {
860 MIB.addSym(Sym: Table);
861 } else {
862 // For the MVP there is at most one table whose number is 0, but we can't
863 // write a table symbol or issue relocations. Instead we just ensure the
864 // table is live and write a zero.
865 Table->setNoStrip();
866 MIB.addImm(Val: 0);
867 }
868 }
869
870 for (auto Use : CallParams.uses())
871 MIB.add(MO: Use);
872
873 BB->insert(I: CallResults.getIterator(), M: MIB);
874 CallParams.eraseFromParent();
875 CallResults.eraseFromParent();
876
877 // If this is a funcref call, to avoid hidden GC roots, we need to clear the
878 // table slot with ref.null upon call_indirect return.
879 //
880 // This generates the following code, which comes right after a call_indirect
881 // of a funcref:
882 //
883 // i32.const 0
884 // ref.null func
885 // table.set __funcref_call_table
886 if (IsIndirect && IsFuncrefCall) {
887 MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol(
888 Ctx&: MF.getContext(), Subtarget);
889 Register RegZero =
890 MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::I32RegClass);
891 MachineInstr *Const0 =
892 BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::CONST_I32), DestReg: RegZero).addImm(Val: 0);
893 BB->insertAfter(I: MIB.getInstr()->getIterator(), MI: Const0);
894
895 Register RegFuncref =
896 MF.getRegInfo().createVirtualRegister(RegClass: &WebAssembly::FUNCREFRegClass);
897 MachineInstr *RefNull =
898 BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::REF_NULL_FUNCREF), DestReg: RegFuncref);
899 BB->insertAfter(I: Const0->getIterator(), MI: RefNull);
900
901 MachineInstr *TableSet =
902 BuildMI(MF, MIMD: DL, MCID: TII.get(Opcode: WebAssembly::TABLE_SET_FUNCREF))
903 .addSym(Sym: Table)
904 .addReg(RegNo: RegZero)
905 .addReg(RegNo: RegFuncref);
906 BB->insertAfter(I: RefNull->getIterator(), MI: TableSet);
907 }
908
909 return BB;
910}
911
912MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter(
913 MachineInstr &MI, MachineBasicBlock *BB) const {
914 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
915 DebugLoc DL = MI.getDebugLoc();
916
917 switch (MI.getOpcode()) {
918 default:
919 llvm_unreachable("Unexpected instr type to insert");
920 case WebAssembly::FP_TO_SINT_I32_F32:
921 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: false, Float64: false,
922 LoweredOpcode: WebAssembly::I32_TRUNC_S_F32);
923 case WebAssembly::FP_TO_UINT_I32_F32:
924 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: false, Float64: false,
925 LoweredOpcode: WebAssembly::I32_TRUNC_U_F32);
926 case WebAssembly::FP_TO_SINT_I64_F32:
927 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: true, Float64: false,
928 LoweredOpcode: WebAssembly::I64_TRUNC_S_F32);
929 case WebAssembly::FP_TO_UINT_I64_F32:
930 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: true, Float64: false,
931 LoweredOpcode: WebAssembly::I64_TRUNC_U_F32);
932 case WebAssembly::FP_TO_SINT_I32_F64:
933 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: false, Float64: true,
934 LoweredOpcode: WebAssembly::I32_TRUNC_S_F64);
935 case WebAssembly::FP_TO_UINT_I32_F64:
936 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: false, Float64: true,
937 LoweredOpcode: WebAssembly::I32_TRUNC_U_F64);
938 case WebAssembly::FP_TO_SINT_I64_F64:
939 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: false, Int64: true, Float64: true,
940 LoweredOpcode: WebAssembly::I64_TRUNC_S_F64);
941 case WebAssembly::FP_TO_UINT_I64_F64:
942 return LowerFPToInt(MI, DL, BB, TII, IsUnsigned: true, Int64: true, Float64: true,
943 LoweredOpcode: WebAssembly::I64_TRUNC_U_F64);
944 case WebAssembly::MEMCPY_A32:
945 return LowerMemcpy(MI, DL, BB, TII, Int64: false);
946 case WebAssembly::MEMCPY_A64:
947 return LowerMemcpy(MI, DL, BB, TII, Int64: true);
948 case WebAssembly::MEMSET_A32:
949 return LowerMemset(MI, DL, BB, TII, Int64: false);
950 case WebAssembly::MEMSET_A64:
951 return LowerMemset(MI, DL, BB, TII, Int64: true);
952 case WebAssembly::CALL_RESULTS:
953 case WebAssembly::RET_CALL_RESULTS:
954 return LowerCallResults(CallResults&: MI, DL, BB, Subtarget, TII);
955 }
956}
957
958std::pair<unsigned, const TargetRegisterClass *>
959WebAssemblyTargetLowering::getRegForInlineAsmConstraint(
960 const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
961 // First, see if this is a constraint that directly corresponds to a
962 // WebAssembly register class.
963 if (Constraint.size() == 1) {
964 switch (Constraint[0]) {
965 case 'r':
966 assert(VT != MVT::iPTR && "Pointer MVT not expected here");
967 if (Subtarget->hasSIMD128() && VT.isVector()) {
968 if (VT.getSizeInBits() == 128)
969 return std::make_pair(x: 0U, y: &WebAssembly::V128RegClass);
970 }
971 if (VT.isInteger() && !VT.isVector()) {
972 if (VT.getSizeInBits() <= 32)
973 return std::make_pair(x: 0U, y: &WebAssembly::I32RegClass);
974 if (VT.getSizeInBits() <= 64)
975 return std::make_pair(x: 0U, y: &WebAssembly::I64RegClass);
976 }
977 if (VT.isFloatingPoint() && !VT.isVector()) {
978 switch (VT.getSizeInBits()) {
979 case 32:
980 return std::make_pair(x: 0U, y: &WebAssembly::F32RegClass);
981 case 64:
982 return std::make_pair(x: 0U, y: &WebAssembly::F64RegClass);
983 default:
984 break;
985 }
986 }
987 break;
988 default:
989 break;
990 }
991 }
992
993 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
994}
995
996bool WebAssemblyTargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
997 // Assume ctz is a relatively cheap operation.
998 return true;
999}
1000
1001bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
1002 // Assume clz is a relatively cheap operation.
1003 return true;
1004}
1005
1006bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL,
1007 const AddrMode &AM,
1008 Type *Ty, unsigned AS,
1009 Instruction *I) const {
1010 // WebAssembly offsets are added as unsigned without wrapping. The
1011 // isLegalAddressingMode gives us no way to determine if wrapping could be
1012 // happening, so we approximate this by accepting only non-negative offsets.
1013 if (AM.BaseOffs < 0)
1014 return false;
1015
1016 // WebAssembly has no scale register operands.
1017 if (AM.Scale != 0)
1018 return false;
1019
1020 // Everything else is legal.
1021 return true;
1022}
1023
1024bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses(
1025 EVT /*VT*/, unsigned /*AddrSpace*/, Align /*Align*/,
1026 MachineMemOperand::Flags /*Flags*/, unsigned *Fast) const {
1027 // WebAssembly supports unaligned accesses, though it should be declared
1028 // with the p2align attribute on loads and stores which do so, and there
1029 // may be a performance impact. We tell LLVM they're "fast" because
1030 // for the kinds of things that LLVM uses this for (merging adjacent stores
1031 // of constants, etc.), WebAssembly implementations will either want the
1032 // unaligned access or they'll split anyway.
1033 if (Fast)
1034 *Fast = 1;
1035 return true;
1036}
1037
1038bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT,
1039 AttributeList Attr) const {
1040 // The current thinking is that wasm engines will perform this optimization,
1041 // so we can save on code size.
1042 return true;
1043}
1044
1045bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
1046 EVT ExtT = ExtVal.getValueType();
1047 SDValue N0 = peekThroughFreeze(V: ExtVal->getOperand(Num: 0));
1048 auto *Load = dyn_cast<LoadSDNode>(Val&: N0);
1049 if (!Load)
1050 return false;
1051 EVT MemT = Load->getValueType(ResNo: 0);
1052 return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) ||
1053 (ExtT == MVT::v4i32 && MemT == MVT::v4i16) ||
1054 (ExtT == MVT::v2i64 && MemT == MVT::v2i32);
1055}
1056
1057bool WebAssemblyTargetLowering::isOffsetFoldingLegal(
1058 const GlobalAddressSDNode *GA) const {
1059 // Wasm doesn't support function addresses with offsets
1060 const GlobalValue *GV = GA->getGlobal();
1061 return isa<Function>(Val: GV) ? false : TargetLowering::isOffsetFoldingLegal(GA);
1062}
1063
1064EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
1065 LLVMContext &C,
1066 EVT VT) const {
1067 if (VT.isVector()) {
1068 if (VT.getVectorElementType() == MVT::f16 && !Subtarget->hasFP16())
1069 return VT.changeElementType(Context&: C, EltVT: MVT::i1);
1070
1071 return VT.changeVectorElementTypeToInteger();
1072 }
1073
1074 // So far, all branch instructions in Wasm take an I32 condition.
1075 // The default TargetLowering::getSetCCResultType returns the pointer size,
1076 // which would be useful to reduce instruction counts when testing
1077 // against 64-bit pointers/values if at some point Wasm supports that.
1078 return EVT::getIntegerVT(Context&: C, BitWidth: 32);
1079}
1080
1081void WebAssemblyTargetLowering::getTgtMemIntrinsic(
1082 SmallVectorImpl<IntrinsicInfo> &Infos, const CallBase &I,
1083 MachineFunction &MF, unsigned Intrinsic) const {
1084 IntrinsicInfo Info;
1085 switch (Intrinsic) {
1086 case Intrinsic::wasm_memory_atomic_notify:
1087 Info.opc = ISD::INTRINSIC_W_CHAIN;
1088 Info.memVT = MVT::i32;
1089 Info.ptrVal = I.getArgOperand(i: 0);
1090 Info.offset = 0;
1091 Info.align = Align(4);
1092 // atomic.notify instruction does not really load the memory specified with
1093 // this argument, but MachineMemOperand should either be load or store, so
1094 // we set this to a load.
1095 // FIXME Volatile isn't really correct, but currently all LLVM atomic
1096 // instructions are treated as volatiles in the backend, so we should be
1097 // consistent. The same applies for wasm_atomic_wait intrinsics too.
1098 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
1099 Infos.push_back(Elt: Info);
1100 return;
1101 case Intrinsic::wasm_memory_atomic_wait32:
1102 Info.opc = ISD::INTRINSIC_W_CHAIN;
1103 Info.memVT = MVT::i32;
1104 Info.ptrVal = I.getArgOperand(i: 0);
1105 Info.offset = 0;
1106 Info.align = Align(4);
1107 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
1108 Infos.push_back(Elt: Info);
1109 return;
1110 case Intrinsic::wasm_memory_atomic_wait64:
1111 Info.opc = ISD::INTRINSIC_W_CHAIN;
1112 Info.memVT = MVT::i64;
1113 Info.ptrVal = I.getArgOperand(i: 0);
1114 Info.offset = 0;
1115 Info.align = Align(8);
1116 Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
1117 Infos.push_back(Elt: Info);
1118 return;
1119 case Intrinsic::wasm_loadf16_f32:
1120 Info.opc = ISD::INTRINSIC_W_CHAIN;
1121 Info.memVT = MVT::f16;
1122 Info.ptrVal = I.getArgOperand(i: 0);
1123 Info.offset = 0;
1124 Info.align = Align(2);
1125 Info.flags = MachineMemOperand::MOLoad;
1126 Infos.push_back(Elt: Info);
1127 return;
1128 case Intrinsic::wasm_storef16_f32:
1129 Info.opc = ISD::INTRINSIC_VOID;
1130 Info.memVT = MVT::f16;
1131 Info.ptrVal = I.getArgOperand(i: 1);
1132 Info.offset = 0;
1133 Info.align = Align(2);
1134 Info.flags = MachineMemOperand::MOStore;
1135 Infos.push_back(Elt: Info);
1136 return;
1137 default:
1138 return;
1139 }
1140}
1141
1142void WebAssemblyTargetLowering::computeKnownBitsForTargetNode(
1143 const SDValue Op, KnownBits &Known, const APInt &DemandedElts,
1144 const SelectionDAG &DAG, unsigned Depth) const {
1145 switch (Op.getOpcode()) {
1146 default:
1147 break;
1148 case ISD::INTRINSIC_WO_CHAIN: {
1149 unsigned IntNo = Op.getConstantOperandVal(i: 0);
1150 switch (IntNo) {
1151 default:
1152 break;
1153 case Intrinsic::wasm_bitmask: {
1154 unsigned BitWidth = Known.getBitWidth();
1155 EVT VT = Op.getOperand(i: 1).getSimpleValueType();
1156 unsigned PossibleBits = VT.getVectorNumElements();
1157 APInt ZeroMask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - PossibleBits);
1158 Known.Zero |= ZeroMask;
1159 break;
1160 }
1161 }
1162 break;
1163 }
1164 case WebAssemblyISD::EXTEND_LOW_U:
1165 case WebAssemblyISD::EXTEND_HIGH_U: {
1166 // We know the high half, of each destination vector element, will be zero.
1167 SDValue SrcOp = Op.getOperand(i: 0);
1168 EVT VT = SrcOp.getSimpleValueType();
1169 unsigned BitWidth = Known.getBitWidth();
1170 if (VT == MVT::v8i8 || VT == MVT::v16i8) {
1171 assert(BitWidth >= 8 && "Unexpected width!");
1172 APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 8);
1173 Known.Zero |= Mask;
1174 } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
1175 assert(BitWidth >= 16 && "Unexpected width!");
1176 APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 16);
1177 Known.Zero |= Mask;
1178 } else if (VT == MVT::v2i32 || VT == MVT::v4i32) {
1179 assert(BitWidth >= 32 && "Unexpected width!");
1180 APInt Mask = APInt::getHighBitsSet(numBits: BitWidth, hiBitsSet: BitWidth - 32);
1181 Known.Zero |= Mask;
1182 }
1183 break;
1184 }
1185 // For 128-bit addition if the upper bits are all zero then it's known that
1186 // the upper bits of the result will have all bits guaranteed zero except the
1187 // first.
1188 case WebAssemblyISD::I64_ADD128:
1189 if (Op.getResNo() == 1) {
1190 SDValue LHS_HI = Op.getOperand(i: 1);
1191 SDValue RHS_HI = Op.getOperand(i: 3);
1192 if (isNullConstant(V: LHS_HI) && isNullConstant(V: RHS_HI))
1193 Known.Zero.setBitsFrom(1);
1194 }
1195 break;
1196 }
1197}
1198
1199TargetLoweringBase::LegalizeTypeAction
1200WebAssemblyTargetLowering::getPreferredVectorAction(MVT VT) const {
1201 if (VT.isFixedLengthVector()) {
1202 MVT EltVT = VT.getVectorElementType();
1203 // We have legal vector types with these lane types, so widening the
1204 // vector would let us use some of the lanes directly without having to
1205 // extend or truncate values.
1206 if (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 ||
1207 EltVT == MVT::i64 || EltVT == MVT::f32 || EltVT == MVT::f64)
1208 return TypeWidenVector;
1209 }
1210
1211 return TargetLoweringBase::getPreferredVectorAction(VT);
1212}
1213
1214bool WebAssemblyTargetLowering::isFMAFasterThanFMulAndFAdd(
1215 const MachineFunction &MF, EVT VT) const {
1216 if (!Subtarget->hasFP16() || !VT.isVector())
1217 return false;
1218
1219 EVT ScalarVT = VT.getScalarType();
1220 if (!ScalarVT.isSimple())
1221 return false;
1222
1223 return ScalarVT.getSimpleVT().SimpleTy == MVT::f16;
1224}
1225
1226bool WebAssemblyTargetLowering::shouldSimplifyDemandedVectorElts(
1227 SDValue Op, const TargetLoweringOpt &TLO) const {
1228 // ISel process runs DAGCombiner after legalization; this step is called
1229 // SelectionDAG optimization phase. This post-legalization combining process
1230 // runs DAGCombiner on each node, and if there was a change to be made,
1231 // re-runs legalization again on it and its user nodes to make sure
1232 // everythiing is in a legalized state.
1233 //
1234 // The legalization calls lowering routines, and we do our custom lowering for
1235 // build_vectors (LowerBUILD_VECTOR), which converts undef vector elements
1236 // into zeros. But there is a set of routines in DAGCombiner that turns unused
1237 // (= not demanded) nodes into undef, among which SimplifyDemandedVectorElts
1238 // turns unused vector elements into undefs. But this routine does not work
1239 // with our custom LowerBUILD_VECTOR, which turns undefs into zeros. This
1240 // combination can result in a infinite loop, in which undefs are converted to
1241 // zeros in legalization and back to undefs in combining.
1242 //
1243 // So after DAG is legalized, we prevent SimplifyDemandedVectorElts from
1244 // running for build_vectors.
1245 if (Op.getOpcode() == ISD::BUILD_VECTOR && TLO.LegalOps && TLO.LegalTys)
1246 return false;
1247 return true;
1248}
1249
1250//===----------------------------------------------------------------------===//
1251// WebAssembly Lowering private implementation.
1252//===----------------------------------------------------------------------===//
1253
1254//===----------------------------------------------------------------------===//
1255// Lowering Code
1256//===----------------------------------------------------------------------===//
1257
1258static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) {
1259 MachineFunction &MF = DAG.getMachineFunction();
1260 DAG.getContext()->diagnose(
1261 DI: DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc()));
1262}
1263
1264// Test whether the given calling convention is supported.
1265static bool callingConvSupported(CallingConv::ID CallConv) {
1266 // We currently support the language-independent target-independent
1267 // conventions. We don't yet have a way to annotate calls with properties like
1268 // "cold", and we don't have any call-clobbered registers, so these are mostly
1269 // all handled the same.
1270 return CallConv == CallingConv::C || CallConv == CallingConv::Fast ||
1271 CallConv == CallingConv::Cold ||
1272 CallConv == CallingConv::PreserveMost ||
1273 CallConv == CallingConv::PreserveAll ||
1274 CallConv == CallingConv::CXX_FAST_TLS ||
1275 CallConv == CallingConv::WASM_EmscriptenInvoke ||
1276 CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail;
1277}
1278
1279SDValue
1280WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI,
1281 SmallVectorImpl<SDValue> &InVals) const {
1282 SelectionDAG &DAG = CLI.DAG;
1283 SDLoc DL = CLI.DL;
1284 SDValue Chain = CLI.Chain;
1285 SDValue Callee = CLI.Callee;
1286 MachineFunction &MF = DAG.getMachineFunction();
1287 auto Layout = MF.getDataLayout();
1288
1289 // A call through a funcref is expressed in IR as a call through the pointer
1290 // produced by the llvm.wasm.funcref.to_ptr intrinsic. Detect this here and
1291 // recover the underlying funcref value so the call can be lowered to a
1292 // table.set + call_indirect through the dedicated __funcref_call_table.
1293 bool IsFuncrefCall = false;
1294 if (Callee.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
1295 Callee.getConstantOperandVal(i: 0) == Intrinsic::wasm_funcref_to_ptr) {
1296 Callee = Callee.getOperand(i: 1);
1297 IsFuncrefCall = true;
1298 }
1299
1300 CallingConv::ID CallConv = CLI.CallConv;
1301 if (!callingConvSupported(CallConv))
1302 fail(DL, DAG,
1303 Msg: "WebAssembly doesn't support language-specific or target-specific "
1304 "calling conventions yet");
1305 if (CLI.IsPatchPoint)
1306 fail(DL, DAG, Msg: "WebAssembly doesn't support patch point yet");
1307
1308 if (CLI.IsTailCall) {
1309 auto NoTail = [&](const char *Msg) {
1310 if (CLI.CB && CLI.CB->isMustTailCall())
1311 fail(DL, DAG, Msg);
1312 CLI.IsTailCall = false;
1313 };
1314
1315 if (!Subtarget->hasTailCall())
1316 NoTail("WebAssembly 'tail-call' feature not enabled");
1317
1318 // Varargs calls cannot be tail calls because the buffer is on the stack
1319 if (CLI.IsVarArg)
1320 NoTail("WebAssembly does not support varargs tail calls");
1321
1322 // Do not tail call unless caller and callee return types match
1323 const Function &F = MF.getFunction();
1324 const TargetMachine &TM = getTargetMachine();
1325 Type *RetTy = F.getReturnType();
1326 SmallVector<MVT, 4> CallerRetTys;
1327 SmallVector<MVT, 4> CalleeRetTys;
1328 computeLegalValueVTs(F, TM, Ty: RetTy, ValueVTs&: CallerRetTys);
1329 computeLegalValueVTs(F, TM, Ty: CLI.RetTy, ValueVTs&: CalleeRetTys);
1330 bool TypesMatch = CallerRetTys.size() == CalleeRetTys.size() &&
1331 std::equal(first1: CallerRetTys.begin(), last1: CallerRetTys.end(),
1332 first2: CalleeRetTys.begin());
1333 if (!TypesMatch)
1334 NoTail("WebAssembly tail call requires caller and callee return types to "
1335 "match");
1336
1337 // If pointers to local stack values are passed, we cannot tail call
1338 if (CLI.CB) {
1339 for (auto &Arg : CLI.CB->args()) {
1340 Value *Val = Arg.get();
1341 // Trace the value back through pointer operations
1342 while (true) {
1343 Value *Src = Val->stripPointerCastsAndAliases();
1344 if (auto *GEP = dyn_cast<GetElementPtrInst>(Val: Src))
1345 Src = GEP->getPointerOperand();
1346 if (Val == Src)
1347 break;
1348 Val = Src;
1349 }
1350 if (isa<AllocaInst>(Val)) {
1351 NoTail(
1352 "WebAssembly does not support tail calling with stack arguments");
1353 break;
1354 }
1355 }
1356 }
1357 }
1358
1359 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1360 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1361 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1362
1363 // The generic code may have added an sret argument. If we're lowering an
1364 // invoke function, the ABI requires that the function pointer be the first
1365 // argument, so we may have to swap the arguments.
1366 if (CallConv == CallingConv::WASM_EmscriptenInvoke && Outs.size() >= 2 &&
1367 Outs[0].Flags.isSRet()) {
1368 std::swap(a&: Outs[0], b&: Outs[1]);
1369 std::swap(a&: OutVals[0], b&: OutVals[1]);
1370 }
1371
1372 bool HasSwiftSelfArg = false;
1373 bool HasSwiftErrorArg = false;
1374 bool HasSwiftAsyncArg = false;
1375 unsigned NumFixedArgs = 0;
1376 for (unsigned I = 0; I < Outs.size(); ++I) {
1377 const ISD::OutputArg &Out = Outs[I];
1378 SDValue &OutVal = OutVals[I];
1379 HasSwiftSelfArg |= Out.Flags.isSwiftSelf();
1380 HasSwiftErrorArg |= Out.Flags.isSwiftError();
1381 HasSwiftAsyncArg |= Out.Flags.isSwiftAsync();
1382 if (Out.Flags.isNest())
1383 fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments");
1384 if (Out.Flags.isInAlloca())
1385 fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments");
1386 if (Out.Flags.isInConsecutiveRegs())
1387 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments");
1388 if (Out.Flags.isInConsecutiveRegsLast())
1389 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments");
1390 if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) {
1391 auto &MFI = MF.getFrameInfo();
1392 int FI = MFI.CreateStackObject(Size: Out.Flags.getByValSize(),
1393 Alignment: Out.Flags.getNonZeroByValAlign(),
1394 /*isSS=*/isSpillSlot: false);
1395 SDValue SizeNode =
1396 DAG.getConstant(Val: Out.Flags.getByValSize(), DL, VT: MVT::i32);
1397 SDValue FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout));
1398 Align Alignment = Out.Flags.getNonZeroByValAlign();
1399 Chain = DAG.getMemcpy(Chain, dl: DL, Dst: FINode, Src: OutVal, Size: SizeNode, DstAlign: Alignment,
1400 SrcAlign: Alignment,
1401 /*isVolatile*/ isVol: false, /*AlwaysInline=*/false,
1402 /*CI=*/nullptr, OverrideTailCall: std::nullopt, DstPtrInfo: MachinePointerInfo(),
1403 SrcPtrInfo: MachinePointerInfo());
1404 OutVal = FINode;
1405 }
1406 // Count the number of fixed args *after* legalization.
1407 NumFixedArgs += !Out.Flags.isVarArg();
1408 }
1409
1410 bool IsVarArg = CLI.IsVarArg;
1411 auto PtrVT = getPointerTy(DL: Layout);
1412
1413 // For swiftcc and swifttailcc, emit additional swiftself, swifterror, and
1414 // (for swifttailcc) swiftasync arguments if there aren't. These additional
1415 // arguments are also added for callee signature. They are necessary to match
1416 // callee and caller signature for indirect call.
1417 if (CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail) {
1418 Type *PtrTy = PointerType::getUnqual(C&: *DAG.getContext());
1419 if (!HasSwiftSelfArg) {
1420 NumFixedArgs++;
1421 ISD::ArgFlagsTy Flags;
1422 Flags.setSwiftSelf();
1423 ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0);
1424 CLI.Outs.push_back(Elt: Arg);
1425 SDValue ArgVal = DAG.getUNDEF(VT: PtrVT);
1426 CLI.OutVals.push_back(Elt: ArgVal);
1427 }
1428 if (!HasSwiftErrorArg) {
1429 NumFixedArgs++;
1430 ISD::ArgFlagsTy Flags;
1431 Flags.setSwiftError();
1432 ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0);
1433 CLI.Outs.push_back(Elt: Arg);
1434 SDValue ArgVal = DAG.getUNDEF(VT: PtrVT);
1435 CLI.OutVals.push_back(Elt: ArgVal);
1436 }
1437 if (CallConv == CallingConv::SwiftTail && !HasSwiftAsyncArg) {
1438 NumFixedArgs++;
1439 ISD::ArgFlagsTy Flags;
1440 Flags.setSwiftAsync();
1441 ISD::OutputArg Arg(Flags, PtrVT, EVT(PtrVT), PtrTy, 0, 0);
1442 CLI.Outs.push_back(Elt: Arg);
1443 SDValue ArgVal = DAG.getUNDEF(VT: PtrVT);
1444 CLI.OutVals.push_back(Elt: ArgVal);
1445 }
1446 }
1447
1448 // Analyze operands of the call, assigning locations to each operand.
1449 SmallVector<CCValAssign, 16> ArgLocs;
1450 CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
1451
1452 if (IsVarArg) {
1453 // Outgoing non-fixed arguments are placed in a buffer. First
1454 // compute their offsets and the total amount of buffer space needed.
1455 for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) {
1456 const ISD::OutputArg &Out = Outs[I];
1457 SDValue &Arg = OutVals[I];
1458 EVT VT = Arg.getValueType();
1459 assert(VT != MVT::iPTR && "Legalized args should be concrete");
1460 Type *Ty = VT.getTypeForEVT(Context&: *DAG.getContext());
1461 Align Alignment =
1462 std::max(a: Out.Flags.getNonZeroOrigAlign(), b: Layout.getABITypeAlign(Ty));
1463 unsigned Offset =
1464 CCInfo.AllocateStack(Size: Layout.getTypeAllocSize(Ty), Alignment);
1465 CCInfo.addLoc(V: CCValAssign::getMem(ValNo: ArgLocs.size(), ValVT: VT.getSimpleVT(),
1466 Offset, LocVT: VT.getSimpleVT(),
1467 HTP: CCValAssign::Full));
1468 }
1469 }
1470
1471 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
1472
1473 SDValue FINode;
1474 if (IsVarArg && NumBytes) {
1475 // For non-fixed arguments, next emit stores to store the argument values
1476 // to the stack buffer at the offsets computed above.
1477 MaybeAlign StackAlign = Layout.getStackAlignment();
1478 assert(StackAlign && "data layout string is missing stack alignment");
1479 int FI = MF.getFrameInfo().CreateStackObject(Size: NumBytes, Alignment: *StackAlign,
1480 /*isSS=*/isSpillSlot: false);
1481 unsigned ValNo = 0;
1482 SmallVector<SDValue, 8> Chains;
1483 for (SDValue Arg : drop_begin(RangeOrContainer&: OutVals, N: NumFixedArgs)) {
1484 assert(ArgLocs[ValNo].getValNo() == ValNo &&
1485 "ArgLocs should remain in order and only hold varargs args");
1486 unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
1487 FINode = DAG.getFrameIndex(FI, VT: getPointerTy(DL: Layout));
1488 SDValue Add = DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: FINode,
1489 N2: DAG.getConstant(Val: Offset, DL, VT: PtrVT));
1490 Chains.push_back(
1491 Elt: DAG.getStore(Chain, dl: DL, Val: Arg, Ptr: Add,
1492 PtrInfo: MachinePointerInfo::getFixedStack(MF, FI, Offset)));
1493 }
1494 if (!Chains.empty())
1495 Chain = DAG.getNode(Opcode: ISD::TokenFactor, DL, VT: MVT::Other, Ops: Chains);
1496 } else if (IsVarArg) {
1497 FINode = DAG.getIntPtrConstant(Val: 0, DL);
1498 }
1499
1500 if (Callee->getOpcode() == ISD::GlobalAddress) {
1501 // If the callee is a GlobalAddress node (quite common, every direct call
1502 // is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress
1503 // doesn't at MO_GOT which is not needed for direct calls.
1504 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Val&: Callee);
1505 Callee = DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL,
1506 VT: getPointerTy(DL: DAG.getDataLayout()),
1507 offset: GA->getOffset());
1508 Callee = DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL,
1509 VT: getPointerTy(DL: DAG.getDataLayout()), Operand: Callee);
1510 }
1511
1512 // Compute the operands for the CALLn node.
1513 SmallVector<SDValue, 16> Ops;
1514 Ops.push_back(Elt: Chain);
1515 Ops.push_back(Elt: Callee);
1516
1517 // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
1518 // isn't reliable.
1519 Ops.append(in_start: OutVals.begin(),
1520 in_end: IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
1521 // Add a pointer to the vararg buffer.
1522 if (IsVarArg)
1523 Ops.push_back(Elt: FINode);
1524
1525 SmallVector<EVT, 8> InTys;
1526 for (const auto &In : Ins) {
1527 assert(!In.Flags.isByVal() && "byval is not valid for return values");
1528 assert(!In.Flags.isNest() && "nest is not valid for return values");
1529 if (In.Flags.isInAlloca())
1530 fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca return values");
1531 if (In.Flags.isInConsecutiveRegs())
1532 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs return values");
1533 if (In.Flags.isInConsecutiveRegsLast())
1534 fail(DL, DAG,
1535 Msg: "WebAssembly hasn't implemented cons regs last return values");
1536 // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in
1537 // registers.
1538 InTys.push_back(Elt: In.VT);
1539 }
1540
1541 // Lastly, if this is a call to a funcref we need to add an instruction
1542 // table.set to the chain and transform the call.
1543 if (IsFuncrefCall) {
1544 // In the absence of function references proposal where a funcref call is
1545 // lowered to call_ref, using reference types we generate a table.set to set
1546 // the funcref to a special table used solely for this purpose, followed by
1547 // a call_indirect. Here we just generate the table set, and return the
1548 // SDValue of the table.set so that LowerCall can finalize the lowering by
1549 // generating the call_indirect.
1550 SDValue Chain = Ops[0];
1551
1552 MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol(
1553 Ctx&: MF.getContext(), Subtarget);
1554 SDValue Sym = DAG.getMCSymbol(Sym: Table, VT: PtrVT);
1555 SDValue TableSlot = DAG.getConstant(Val: 0, DL, VT: MVT::i32);
1556 SDValue TableSetOps[] = {Chain, Sym, TableSlot, Callee};
1557 SDValue TableSet = DAG.getMemIntrinsicNode(
1558 Opcode: WebAssemblyISD::TABLE_SET, dl: DL, VTList: DAG.getVTList(VT: MVT::Other), Ops: TableSetOps,
1559 MemVT: MVT::funcref, PtrInfo: MachinePointerInfo(), Alignment: Align(1),
1560 Flags: MachineMemOperand::MOStore);
1561
1562 Ops[0] = TableSet; // The new chain is the TableSet itself
1563 }
1564
1565 if (CLI.IsTailCall) {
1566 // ret_calls do not return values to the current frame
1567 SDVTList NodeTys = DAG.getVTList(VT1: MVT::Other, VT2: MVT::Glue);
1568 return DAG.getNode(Opcode: WebAssemblyISD::RET_CALL, DL, VTList: NodeTys, Ops);
1569 }
1570
1571 InTys.push_back(Elt: MVT::Other);
1572 SDVTList InTyList = DAG.getVTList(VTs: InTys);
1573 SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::CALL, DL, VTList: InTyList, Ops);
1574
1575 for (size_t I = 0; I < Ins.size(); ++I)
1576 InVals.push_back(Elt: Res.getValue(R: I));
1577
1578 // Return the chain
1579 return Res.getValue(R: Ins.size());
1580}
1581
1582bool WebAssemblyTargetLowering::CanLowerReturn(
1583 CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
1584 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext & /*Context*/,
1585 const Type *RetTy) const {
1586 // WebAssembly can only handle returning tuples with multivalue enabled
1587 return WebAssembly::canLowerReturn(ResultSize: Outs.size(), Subtarget);
1588}
1589
1590SDValue WebAssemblyTargetLowering::LowerReturn(
1591 SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
1592 const SmallVectorImpl<ISD::OutputArg> &Outs,
1593 const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
1594 SelectionDAG &DAG) const {
1595 assert(WebAssembly::canLowerReturn(Outs.size(), Subtarget) &&
1596 "MVP WebAssembly can only return up to one value");
1597 if (!callingConvSupported(CallConv))
1598 fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions");
1599
1600 SmallVector<SDValue, 4> RetOps(1, Chain);
1601 RetOps.append(in_start: OutVals.begin(), in_end: OutVals.end());
1602 Chain = DAG.getNode(Opcode: WebAssemblyISD::RETURN, DL, VT: MVT::Other, Ops: RetOps);
1603
1604 // Record the number and types of the return values.
1605 for (const ISD::OutputArg &Out : Outs) {
1606 assert(!Out.Flags.isByVal() && "byval is not valid for return values");
1607 assert(!Out.Flags.isNest() && "nest is not valid for return values");
1608 assert(!Out.Flags.isVarArg() && "non-fixed return value is not valid");
1609 if (Out.Flags.isInAlloca())
1610 fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca results");
1611 if (Out.Flags.isInConsecutiveRegs())
1612 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs results");
1613 if (Out.Flags.isInConsecutiveRegsLast())
1614 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last results");
1615 }
1616
1617 return Chain;
1618}
1619
1620SDValue WebAssemblyTargetLowering::LowerFormalArguments(
1621 SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
1622 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
1623 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1624 if (!callingConvSupported(CallConv))
1625 fail(DL, DAG, Msg: "WebAssembly doesn't support non-C calling conventions");
1626
1627 MachineFunction &MF = DAG.getMachineFunction();
1628 auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>();
1629
1630 // Set up the incoming ARGUMENTS value, which serves to represent the liveness
1631 // of the incoming values before they're represented by virtual registers.
1632 MF.getRegInfo().addLiveIn(Reg: WebAssembly::ARGUMENTS);
1633
1634 bool HasSwiftErrorArg = false;
1635 bool HasSwiftSelfArg = false;
1636 bool HasSwiftAsyncArg = false;
1637 for (const ISD::InputArg &In : Ins) {
1638 HasSwiftSelfArg |= In.Flags.isSwiftSelf();
1639 HasSwiftErrorArg |= In.Flags.isSwiftError();
1640 HasSwiftAsyncArg |= In.Flags.isSwiftAsync();
1641 if (In.Flags.isInAlloca())
1642 fail(DL, DAG, Msg: "WebAssembly hasn't implemented inalloca arguments");
1643 if (In.Flags.isNest())
1644 fail(DL, DAG, Msg: "WebAssembly hasn't implemented nest arguments");
1645 if (In.Flags.isInConsecutiveRegs())
1646 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs arguments");
1647 if (In.Flags.isInConsecutiveRegsLast())
1648 fail(DL, DAG, Msg: "WebAssembly hasn't implemented cons regs last arguments");
1649 // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in
1650 // registers.
1651 InVals.push_back(Elt: In.Used ? DAG.getNode(Opcode: WebAssemblyISD::ARGUMENT, DL, VT: In.VT,
1652 Operand: DAG.getTargetConstant(Val: InVals.size(),
1653 DL, VT: MVT::i32))
1654 : DAG.getUNDEF(VT: In.VT));
1655
1656 // Record the number and types of arguments.
1657 MFI->addParam(VT: In.VT);
1658 }
1659
1660 // For swiftcc and swifttailcc, emit additional swiftself, swifterror, and
1661 // (for swifttailcc) swiftasync arguments if there aren't. These additional
1662 // arguments are also added for callee signature. They are necessary to match
1663 // callee and caller signature for indirect call.
1664 auto PtrVT = getPointerTy(DL: MF.getDataLayout());
1665 if (CallConv == CallingConv::Swift || CallConv == CallingConv::SwiftTail) {
1666 if (!HasSwiftSelfArg) {
1667 MFI->addParam(VT: PtrVT);
1668 }
1669 if (!HasSwiftErrorArg) {
1670 MFI->addParam(VT: PtrVT);
1671 }
1672 if (CallConv == CallingConv::SwiftTail && !HasSwiftAsyncArg) {
1673 MFI->addParam(VT: PtrVT);
1674 }
1675 }
1676 // Varargs are copied into a buffer allocated by the caller, and a pointer to
1677 // the buffer is passed as an argument.
1678 if (IsVarArg) {
1679 MVT PtrVT = getPointerTy(DL: MF.getDataLayout());
1680 Register VarargVreg =
1681 MF.getRegInfo().createVirtualRegister(RegClass: getRegClassFor(VT: PtrVT));
1682 MFI->setVarargBufferVreg(VarargVreg);
1683 Chain = DAG.getCopyToReg(
1684 Chain, dl: DL, Reg: VarargVreg,
1685 N: DAG.getNode(Opcode: WebAssemblyISD::ARGUMENT, DL, VT: PtrVT,
1686 Operand: DAG.getTargetConstant(Val: Ins.size(), DL, VT: MVT::i32)));
1687 MFI->addParam(VT: PtrVT);
1688 }
1689
1690 // Record the number and types of arguments and results.
1691 SmallVector<MVT, 4> Params;
1692 SmallVector<MVT, 4> Results;
1693 computeSignatureVTs(Ty: MF.getFunction().getFunctionType(), TargetFunc: &MF.getFunction(),
1694 ContextFunc: MF.getFunction(), TM: DAG.getTarget(), Params, Results);
1695 for (MVT VT : Results)
1696 MFI->addResult(VT);
1697 // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify
1698 // the param logic here with ComputeSignatureVTs
1699 assert(MFI->getParams().size() == Params.size() &&
1700 std::equal(MFI->getParams().begin(), MFI->getParams().end(),
1701 Params.begin()));
1702
1703 return Chain;
1704}
1705
1706void WebAssemblyTargetLowering::ReplaceNodeResults(
1707 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
1708 switch (N->getOpcode()) {
1709 case ISD::SIGN_EXTEND_INREG:
1710 // Do not add any results, signifying that N should not be custom lowered
1711 // after all. This happens because simd128 turns on custom lowering for
1712 // SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an
1713 // illegal type.
1714 break;
1715 case ISD::ANY_EXTEND_VECTOR_INREG:
1716 case ISD::SIGN_EXTEND_VECTOR_INREG:
1717 case ISD::ZERO_EXTEND_VECTOR_INREG:
1718 // Do not add any results, signifying that N should not be custom lowered.
1719 // EXTEND_VECTOR_INREG is implemented for some vectors, but not all.
1720 break;
1721 case ISD::FP_ROUND: {
1722 EVT VT = N->getValueType(ResNo: 0);
1723 SDValue Src = N->getOperand(Num: 0);
1724 if (VT == MVT::v4f16 && Src.getValueType() == MVT::v4f32) {
1725 Results.push_back(
1726 Elt: DAG.getNode(Opcode: WebAssemblyISD::DEMOTE_ZERO, DL: SDLoc(N), VT: MVT::v8f16, Operand: Src));
1727 }
1728 break;
1729 }
1730 case ISD::ADD:
1731 case ISD::SUB:
1732 Results.push_back(Elt: Replace128Op(N, DAG));
1733 break;
1734 default:
1735 llvm_unreachable(
1736 "ReplaceNodeResults not implemented for this op for WebAssembly!");
1737 }
1738}
1739
1740//===----------------------------------------------------------------------===//
1741// Custom lowering hooks.
1742//===----------------------------------------------------------------------===//
1743
1744SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
1745 SelectionDAG &DAG) const {
1746 SDLoc DL(Op);
1747 switch (Op.getOpcode()) {
1748 default:
1749 llvm_unreachable("unimplemented operation lowering");
1750 return SDValue();
1751 case ISD::FrameIndex:
1752 return LowerFrameIndex(Op, DAG);
1753 case ISD::GlobalAddress:
1754 return LowerGlobalAddress(Op, DAG);
1755 case ISD::GlobalTLSAddress:
1756 return LowerGlobalTLSAddress(Op, DAG);
1757 case ISD::ExternalSymbol:
1758 return LowerExternalSymbol(Op, DAG);
1759 case ISD::JumpTable:
1760 return LowerJumpTable(Op, DAG);
1761 case ISD::BR_JT:
1762 return LowerBR_JT(Op, DAG);
1763 case ISD::VASTART:
1764 return LowerVASTART(Op, DAG);
1765 case ISD::BlockAddress:
1766 case ISD::BRIND:
1767 fail(DL, DAG, Msg: "WebAssembly hasn't implemented computed gotos");
1768 return SDValue();
1769 case ISD::RETURNADDR:
1770 return LowerRETURNADDR(Op, DAG);
1771 case ISD::FRAMEADDR:
1772 return LowerFRAMEADDR(Op, DAG);
1773 case ISD::CopyToReg:
1774 return LowerCopyToReg(Op, DAG);
1775 case ISD::EXTRACT_VECTOR_ELT:
1776 case ISD::INSERT_VECTOR_ELT:
1777 return LowerAccessVectorElement(Op, DAG);
1778 case ISD::INTRINSIC_VOID:
1779 case ISD::INTRINSIC_WO_CHAIN:
1780 case ISD::INTRINSIC_W_CHAIN:
1781 return LowerIntrinsic(Op, DAG);
1782 case ISD::SIGN_EXTEND_INREG:
1783 return LowerSIGN_EXTEND_INREG(Op, DAG);
1784 case ISD::ZERO_EXTEND_VECTOR_INREG:
1785 case ISD::SIGN_EXTEND_VECTOR_INREG:
1786 case ISD::ANY_EXTEND_VECTOR_INREG:
1787 return LowerEXTEND_VECTOR_INREG(Op, DAG);
1788 case ISD::BUILD_VECTOR:
1789 return LowerBUILD_VECTOR(Op, DAG);
1790 case ISD::VECTOR_SHUFFLE:
1791 return LowerVECTOR_SHUFFLE(Op, DAG);
1792 case ISD::SETCC:
1793 return LowerSETCC(Op, DAG);
1794 case ISD::SHL:
1795 case ISD::SRA:
1796 case ISD::SRL:
1797 return LowerShift(Op, DAG);
1798 case ISD::FP_TO_SINT_SAT:
1799 case ISD::FP_TO_UINT_SAT:
1800 return LowerFP_TO_INT_SAT(Op, DAG);
1801 case ISD::FMINNUM:
1802 case ISD::FMINIMUMNUM:
1803 return LowerFMIN(Op, DAG);
1804 case ISD::FMAXNUM:
1805 case ISD::FMAXIMUMNUM:
1806 return LowerFMAX(Op, DAG);
1807 case ISD::LOAD:
1808 return LowerLoad(Op, DAG);
1809 case ISD::STORE:
1810 return LowerStore(Op, DAG);
1811 case ISD::CTPOP:
1812 case ISD::CTLZ:
1813 case ISD::CTTZ:
1814 return DAG.UnrollVectorOp(N: Op.getNode());
1815 case ISD::CLEAR_CACHE:
1816 // Report this as a diagnostic rather than aborting, like the other
1817 // unsupported features in this target. Pass the chain through so that
1818 // codegen can reach the point where the diagnostic is emitted.
1819 fail(DL: SDLoc(Op), DAG, Msg: "llvm.clear_cache is not supported on wasm");
1820 return Op.getOperand(i: 0);
1821 case ISD::SMUL_LOHI:
1822 case ISD::UMUL_LOHI:
1823 return LowerMUL_LOHI(Op, DAG);
1824 case ISD::UADDO:
1825 return LowerUADDO(Op, DAG);
1826 }
1827}
1828
1829static bool IsWebAssemblyGlobal(SDValue Op) {
1830 if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op))
1831 return WebAssembly::isWasmVarAddressSpace(AS: GA->getAddressSpace());
1832
1833 return false;
1834}
1835
1836static std::optional<unsigned> IsWebAssemblyLocal(SDValue Op,
1837 SelectionDAG &DAG) {
1838 const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Val&: Op);
1839 if (!FI)
1840 return std::nullopt;
1841
1842 auto &MF = DAG.getMachineFunction();
1843 return WebAssemblyFrameLowering::getLocalForStackObject(MF, FrameIndex: FI->getIndex());
1844}
1845
1846SDValue WebAssemblyTargetLowering::LowerStore(SDValue Op,
1847 SelectionDAG &DAG) const {
1848 SDLoc DL(Op);
1849 StoreSDNode *SN = cast<StoreSDNode>(Val: Op.getNode());
1850 const SDValue &Value = SN->getValue();
1851 const SDValue &Base = SN->getBasePtr();
1852 const SDValue &Offset = SN->getOffset();
1853
1854 if (IsWebAssemblyGlobal(Op: Base)) {
1855 if (!Offset->isUndef())
1856 report_fatal_error(reason: "unexpected offset when storing to webassembly global",
1857 gen_crash_diag: false);
1858
1859 SDVTList Tys = DAG.getVTList(VT: MVT::Other);
1860 SDValue Ops[] = {SN->getChain(), Value, Base};
1861 return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_SET, dl: DL, VTList: Tys, Ops,
1862 MemVT: SN->getMemoryVT(), MMO: SN->getMemOperand());
1863 }
1864
1865 if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) {
1866 if (!Offset->isUndef())
1867 report_fatal_error(reason: "unexpected offset when storing to webassembly local",
1868 gen_crash_diag: false);
1869
1870 SDValue Idx = DAG.getTargetConstant(Val: *Local, DL: Base, VT: MVT::i32);
1871 SDVTList Tys = DAG.getVTList(VT: MVT::Other); // The chain.
1872 SDValue Ops[] = {SN->getChain(), Idx, Value};
1873 return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_SET, DL, VTList: Tys, Ops);
1874 }
1875
1876 if (WebAssembly::isWasmVarAddressSpace(AS: SN->getAddressSpace()))
1877 report_fatal_error(
1878 reason: "Encountered an unlowerable store to the wasm_var address space",
1879 gen_crash_diag: false);
1880
1881 return Op;
1882}
1883
1884SDValue WebAssemblyTargetLowering::LowerLoad(SDValue Op,
1885 SelectionDAG &DAG) const {
1886 SDLoc DL(Op);
1887 LoadSDNode *LN = cast<LoadSDNode>(Val: Op.getNode());
1888 const SDValue &Base = LN->getBasePtr();
1889 const SDValue &Offset = LN->getOffset();
1890
1891 if (IsWebAssemblyGlobal(Op: Base)) {
1892 if (!Offset->isUndef())
1893 report_fatal_error(
1894 reason: "unexpected offset when loading from webassembly global", gen_crash_diag: false);
1895
1896 SDVTList Tys = DAG.getVTList(VT1: LN->getValueType(ResNo: 0), VT2: MVT::Other);
1897 SDValue Ops[] = {LN->getChain(), Base};
1898 return DAG.getMemIntrinsicNode(Opcode: WebAssemblyISD::GLOBAL_GET, dl: DL, VTList: Tys, Ops,
1899 MemVT: LN->getMemoryVT(), MMO: LN->getMemOperand());
1900 }
1901
1902 if (std::optional<unsigned> Local = IsWebAssemblyLocal(Op: Base, DAG)) {
1903 if (!Offset->isUndef())
1904 report_fatal_error(
1905 reason: "unexpected offset when loading from webassembly local", gen_crash_diag: false);
1906
1907 SDValue Idx = DAG.getTargetConstant(Val: *Local, DL: Base, VT: MVT::i32);
1908 EVT LocalVT = LN->getValueType(ResNo: 0);
1909 return DAG.getNode(Opcode: WebAssemblyISD::LOCAL_GET, DL, ResultTys: {LocalVT, MVT::Other},
1910 Ops: {LN->getChain(), Idx});
1911 }
1912
1913 if (WebAssembly::isWasmVarAddressSpace(AS: LN->getAddressSpace()))
1914 report_fatal_error(
1915 reason: "Encountered an unlowerable load from the wasm_var address space",
1916 gen_crash_diag: false);
1917
1918 return Op;
1919}
1920
1921SDValue WebAssemblyTargetLowering::LowerMUL_LOHI(SDValue Op,
1922 SelectionDAG &DAG) const {
1923 assert(Subtarget->hasWideArithmetic());
1924 assert(Op.getValueType() == MVT::i64);
1925 SDLoc DL(Op);
1926 unsigned Opcode;
1927 switch (Op.getOpcode()) {
1928 case ISD::UMUL_LOHI:
1929 Opcode = WebAssemblyISD::I64_MUL_WIDE_U;
1930 break;
1931 case ISD::SMUL_LOHI:
1932 Opcode = WebAssemblyISD::I64_MUL_WIDE_S;
1933 break;
1934 default:
1935 llvm_unreachable("unexpected opcode");
1936 }
1937 SDValue LHS = Op.getOperand(i: 0);
1938 SDValue RHS = Op.getOperand(i: 1);
1939 SDValue Lo =
1940 DAG.getNode(Opcode, DL, VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), N1: LHS, N2: RHS);
1941 SDValue Hi(Lo.getNode(), 1);
1942 SDValue Ops[] = {Lo, Hi};
1943 return DAG.getMergeValues(Ops, dl: DL);
1944}
1945
1946// Lowers `UADDO` intrinsics to an `i64.add128` instruction when it's enabled.
1947//
1948// This enables generating a single wasm instruction for this operation where
1949// the upper half of both operands are constant zeros. The upper half of the
1950// result is then whether the overflow happened.
1951SDValue WebAssemblyTargetLowering::LowerUADDO(SDValue Op,
1952 SelectionDAG &DAG) const {
1953 assert(Subtarget->hasWideArithmetic());
1954 assert(Op.getValueType() == MVT::i64);
1955 assert(Op.getOpcode() == ISD::UADDO);
1956 SDLoc DL(Op);
1957 SDValue LHS = Op.getOperand(i: 0);
1958 SDValue RHS = Op.getOperand(i: 1);
1959 SDValue Zero = DAG.getConstant(Val: 0, DL, VT: MVT::i64);
1960 SDValue Result =
1961 DAG.getNode(Opcode: WebAssemblyISD::I64_ADD128, DL,
1962 VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64), N1: LHS, N2: Zero, N3: RHS, N4: Zero);
1963 SDValue CarryI64(Result.getNode(), 1);
1964 SDValue CarryI32 = DAG.getNode(Opcode: ISD::TRUNCATE, DL, VT: MVT::i32, Operand: CarryI64);
1965 SDValue Ops[] = {Result, CarryI32};
1966 return DAG.getMergeValues(Ops, dl: DL);
1967}
1968
1969SDValue WebAssemblyTargetLowering::Replace128Op(SDNode *N,
1970 SelectionDAG &DAG) const {
1971 assert(Subtarget->hasWideArithmetic());
1972 assert(N->getValueType(0) == MVT::i128);
1973 SDLoc DL(N);
1974 unsigned Opcode;
1975 switch (N->getOpcode()) {
1976 case ISD::ADD:
1977 Opcode = WebAssemblyISD::I64_ADD128;
1978 break;
1979 case ISD::SUB:
1980 Opcode = WebAssemblyISD::I64_SUB128;
1981 break;
1982 default:
1983 llvm_unreachable("unexpected opcode");
1984 }
1985 SDValue LHS = N->getOperand(Num: 0);
1986 SDValue RHS = N->getOperand(Num: 1);
1987
1988 SDValue C0 = DAG.getConstant(Val: 0, DL, VT: MVT::i64);
1989 SDValue C1 = DAG.getConstant(Val: 1, DL, VT: MVT::i64);
1990 SDValue LHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: LHS, N2: C0);
1991 SDValue LHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: LHS, N2: C1);
1992 SDValue RHS_0 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: RHS, N2: C0);
1993 SDValue RHS_1 = DAG.getNode(Opcode: ISD::EXTRACT_ELEMENT, DL, VT: MVT::i64, N1: RHS, N2: C1);
1994 SDValue Result_LO = DAG.getNode(Opcode, DL, VTList: DAG.getVTList(VT1: MVT::i64, VT2: MVT::i64),
1995 N1: LHS_0, N2: LHS_1, N3: RHS_0, N4: RHS_1);
1996 SDValue Result_HI(Result_LO.getNode(), 1);
1997 return DAG.getNode(Opcode: ISD::BUILD_PAIR, DL, VTList: N->getVTList(), N1: Result_LO, N2: Result_HI);
1998}
1999
2000SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op,
2001 SelectionDAG &DAG) const {
2002 SDValue Src = Op.getOperand(i: 2);
2003 if (isa<FrameIndexSDNode>(Val: Src.getNode())) {
2004 // CopyToReg nodes don't support FrameIndex operands. Other targets select
2005 // the FI to some LEA-like instruction, but since we don't have that, we
2006 // need to insert some kind of instruction that can take an FI operand and
2007 // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
2008 // local.copy between Op and its FI operand.
2009 SDValue Chain = Op.getOperand(i: 0);
2010 SDLoc DL(Op);
2011 Register Reg = cast<RegisterSDNode>(Val: Op.getOperand(i: 1))->getReg();
2012 EVT VT = Src.getValueType();
2013 SDValue Copy(DAG.getMachineNode(Opcode: VT == MVT::i32 ? WebAssembly::COPY_I32
2014 : WebAssembly::COPY_I64,
2015 dl: DL, VT, Op1: Src),
2016 0);
2017 return Op.getNode()->getNumValues() == 1
2018 ? DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy)
2019 : DAG.getCopyToReg(Chain, dl: DL, Reg, N: Copy,
2020 Glue: Op.getNumOperands() == 4 ? Op.getOperand(i: 3)
2021 : SDValue());
2022 }
2023 return SDValue();
2024}
2025
2026SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op,
2027 SelectionDAG &DAG) const {
2028 int FI = cast<FrameIndexSDNode>(Val&: Op)->getIndex();
2029 return DAG.getTargetFrameIndex(FI, VT: Op.getValueType());
2030}
2031
2032SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op,
2033 SelectionDAG &DAG) const {
2034 SDLoc DL(Op);
2035
2036 if (!Subtarget->getTargetTriple().isOSEmscripten()) {
2037 fail(DL, DAG,
2038 Msg: "Non-Emscripten WebAssembly hasn't implemented "
2039 "__builtin_return_address");
2040 return SDValue();
2041 }
2042
2043 unsigned Depth = Op.getConstantOperandVal(i: 0);
2044 MakeLibCallOptions CallOptions;
2045 return makeLibCall(DAG, LC: RTLIB::RETURN_ADDRESS, RetVT: Op.getValueType(),
2046 Ops: {DAG.getConstant(Val: Depth, DL, VT: MVT::i32)}, CallOptions, dl: DL)
2047 .first;
2048}
2049
2050SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op,
2051 SelectionDAG &DAG) const {
2052 // Non-zero depths are not supported by WebAssembly currently. Use the
2053 // legalizer's default expansion, which is to return 0 (what this function is
2054 // documented to do).
2055 if (Op.getConstantOperandVal(i: 0) > 0)
2056 return SDValue();
2057
2058 DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true);
2059 EVT VT = Op.getValueType();
2060 Register FP =
2061 Subtarget->getRegisterInfo()->getFrameRegister(MF: DAG.getMachineFunction());
2062 return DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: SDLoc(Op), Reg: FP, VT);
2063}
2064
2065SDValue
2066WebAssemblyTargetLowering::LowerGlobalTLSAddress(SDValue Op,
2067 SelectionDAG &DAG) const {
2068 SDLoc DL(Op);
2069 const auto *GA = cast<GlobalAddressSDNode>(Val&: Op);
2070
2071 MachineFunction &MF = DAG.getMachineFunction();
2072 if (!MF.getSubtarget<WebAssemblySubtarget>().hasBulkMemory())
2073 report_fatal_error(reason: "cannot use thread-local storage without bulk memory",
2074 gen_crash_diag: false);
2075
2076 const GlobalValue *GV = GA->getGlobal();
2077
2078 // Currently only Emscripten supports dynamic linking with threads. Therefore,
2079 // on other targets, if we have thread-local storage, only the local-exec
2080 // model is possible.
2081 auto model = Subtarget->getTargetTriple().isOSEmscripten()
2082 ? GV->getThreadLocalMode()
2083 : GlobalValue::LocalExecTLSModel;
2084
2085 // Unsupported TLS modes
2086 assert(model != GlobalValue::NotThreadLocal);
2087 assert(model != GlobalValue::InitialExecTLSModel);
2088
2089 if (model == GlobalValue::LocalExecTLSModel ||
2090 model == GlobalValue::LocalDynamicTLSModel ||
2091 (model == GlobalValue::GeneralDynamicTLSModel &&
2092 getTargetMachine().shouldAssumeDSOLocal(GV))) {
2093 // For DSO-local TLS variables we use offset from __tls_base, or
2094 // __wasm_get_tls_base() if using libcall thread context.
2095
2096 MVT PtrVT = getPointerTy(DL: DAG.getDataLayout());
2097 SDValue BaseAddr(WebAssembly::getTLSBase(DAG, DL, Subtarget), 0);
2098
2099 SDValue TLSOffset = DAG.getTargetGlobalAddress(
2100 GV, DL, VT: PtrVT, offset: GA->getOffset(), TargetFlags: WebAssemblyII::MO_TLS_BASE_REL);
2101 SDValue SymOffset =
2102 DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: TLSOffset);
2103
2104 return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymOffset);
2105 }
2106
2107 assert(model == GlobalValue::GeneralDynamicTLSModel);
2108
2109 EVT VT = Op.getValueType();
2110 return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT,
2111 Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT,
2112 offset: GA->getOffset(),
2113 TargetFlags: WebAssemblyII::MO_GOT_TLS));
2114}
2115
2116SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
2117 SelectionDAG &DAG) const {
2118 SDLoc DL(Op);
2119 const auto *GA = cast<GlobalAddressSDNode>(Val&: Op);
2120 EVT VT = Op.getValueType();
2121 assert(GA->getTargetFlags() == 0 &&
2122 "Unexpected target flags on generic GlobalAddressSDNode");
2123 if (!WebAssembly::isValidAddressSpace(AS: GA->getAddressSpace()))
2124 fail(DL, DAG, Msg: "Invalid address space for WebAssembly target");
2125
2126 unsigned OperandFlags = 0;
2127 const GlobalValue *GV = GA->getGlobal();
2128 // Since WebAssembly tables cannot yet be shared across modules, we don't
2129 // need special treatment for tables in PIC mode.
2130 if (isPositionIndependent() &&
2131 !WebAssembly::isWebAssemblyTableType(Ty: GV->getValueType())) {
2132 if (getTargetMachine().shouldAssumeDSOLocal(GV)) {
2133 MachineFunction &MF = DAG.getMachineFunction();
2134 MVT PtrVT = getPointerTy(DL: MF.getDataLayout());
2135 const char *BaseName;
2136 if (GV->getValueType()->isFunctionTy()) {
2137 BaseName = MF.createExternalSymbolName(Name: "__table_base");
2138 OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL;
2139 } else {
2140 BaseName = MF.createExternalSymbolName(Name: "__memory_base");
2141 OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL;
2142 }
2143 SDValue BaseAddr =
2144 DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT,
2145 Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT));
2146
2147 SDValue SymAddr = DAG.getNode(
2148 Opcode: WebAssemblyISD::WrapperREL, DL, VT,
2149 Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT, offset: GA->getOffset(),
2150 TargetFlags: OperandFlags));
2151
2152 return DAG.getNode(Opcode: ISD::ADD, DL, VT, N1: BaseAddr, N2: SymAddr);
2153 }
2154 OperandFlags = WebAssemblyII::MO_GOT;
2155 }
2156
2157 return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT,
2158 Operand: DAG.getTargetGlobalAddress(GV: GA->getGlobal(), DL, VT,
2159 offset: GA->getOffset(), TargetFlags: OperandFlags));
2160}
2161
2162SDValue
2163WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op,
2164 SelectionDAG &DAG) const {
2165 SDLoc DL(Op);
2166 const auto *ES = cast<ExternalSymbolSDNode>(Val&: Op);
2167 EVT VT = Op.getValueType();
2168 assert(ES->getTargetFlags() == 0 &&
2169 "Unexpected target flags on generic ExternalSymbolSDNode");
2170 return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT,
2171 Operand: DAG.getTargetExternalSymbol(Sym: ES->getSymbol(), VT));
2172}
2173
2174SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op,
2175 SelectionDAG &DAG) const {
2176 // There's no need for a Wrapper node because we always incorporate a jump
2177 // table operand into a BR_TABLE instruction, rather than ever
2178 // materializing it in a register.
2179 const JumpTableSDNode *JT = cast<JumpTableSDNode>(Val&: Op);
2180 return DAG.getTargetJumpTable(JTI: JT->getIndex(), VT: Op.getValueType(),
2181 TargetFlags: JT->getTargetFlags());
2182}
2183
2184SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
2185 SelectionDAG &DAG) const {
2186 SDLoc DL(Op);
2187 SDValue Chain = Op.getOperand(i: 0);
2188 const auto *JT = cast<JumpTableSDNode>(Val: Op.getOperand(i: 1));
2189 SDValue Index = Op.getOperand(i: 2);
2190 assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags");
2191
2192 SmallVector<SDValue, 8> Ops;
2193 Ops.push_back(Elt: Chain);
2194 Ops.push_back(Elt: Index);
2195
2196 MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
2197 const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs;
2198
2199 // Add an operand for each case.
2200 for (auto *MBB : MBBs)
2201 Ops.push_back(Elt: DAG.getBasicBlock(MBB));
2202
2203 // Add the first MBB as a dummy default target for now. This will be replaced
2204 // with the proper default target (and the preceding range check eliminated)
2205 // if possible by WebAssemblyFixBrTableDefaults.
2206 Ops.push_back(Elt: DAG.getBasicBlock(MBB: *MBBs.begin()));
2207 return DAG.getNode(Opcode: WebAssemblyISD::BR_TABLE, DL, VT: MVT::Other, Ops);
2208}
2209
2210SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
2211 SelectionDAG &DAG) const {
2212 SDLoc DL(Op);
2213 EVT PtrVT = getPointerTy(DL: DAG.getMachineFunction().getDataLayout());
2214
2215 auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>();
2216 const Value *SV = cast<SrcValueSDNode>(Val: Op.getOperand(i: 2))->getValue();
2217
2218 SDValue ArgN = DAG.getCopyFromReg(Chain: DAG.getEntryNode(), dl: DL,
2219 Reg: MFI->getVarargBufferVreg(), VT: PtrVT);
2220 return DAG.getStore(Chain: Op.getOperand(i: 0), dl: DL, Val: ArgN, Ptr: Op.getOperand(i: 1),
2221 PtrInfo: MachinePointerInfo(SV));
2222}
2223
2224SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op,
2225 SelectionDAG &DAG) const {
2226 MachineFunction &MF = DAG.getMachineFunction();
2227 unsigned IntNo;
2228 switch (Op.getOpcode()) {
2229 case ISD::INTRINSIC_VOID:
2230 case ISD::INTRINSIC_W_CHAIN:
2231 IntNo = Op.getConstantOperandVal(i: 1);
2232 break;
2233 case ISD::INTRINSIC_WO_CHAIN:
2234 IntNo = Op.getConstantOperandVal(i: 0);
2235 break;
2236 default:
2237 llvm_unreachable("Invalid intrinsic");
2238 }
2239 SDLoc DL(Op);
2240
2241 switch (IntNo) {
2242 default:
2243 return SDValue(); // Don't custom lower most intrinsics.
2244
2245 case Intrinsic::wasm_lsda: {
2246 auto PtrVT = getPointerTy(DL: MF.getDataLayout());
2247 const char *SymName = MF.createExternalSymbolName(
2248 Name: "GCC_except_table" + std::to_string(val: MF.getFunctionNumber()));
2249 if (isPositionIndependent()) {
2250 SDValue Node = DAG.getTargetExternalSymbol(
2251 Sym: SymName, VT: PtrVT, TargetFlags: WebAssemblyII::MO_MEMORY_BASE_REL);
2252 const char *BaseName = MF.createExternalSymbolName(Name: "__memory_base");
2253 SDValue BaseAddr =
2254 DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT,
2255 Operand: DAG.getTargetExternalSymbol(Sym: BaseName, VT: PtrVT));
2256 SDValue SymAddr =
2257 DAG.getNode(Opcode: WebAssemblyISD::WrapperREL, DL, VT: PtrVT, Operand: Node);
2258 return DAG.getNode(Opcode: ISD::ADD, DL, VT: PtrVT, N1: BaseAddr, N2: SymAddr);
2259 }
2260 SDValue Node = DAG.getTargetExternalSymbol(Sym: SymName, VT: PtrVT);
2261 return DAG.getNode(Opcode: WebAssemblyISD::Wrapper, DL, VT: PtrVT, Operand: Node);
2262 }
2263
2264 case Intrinsic::wasm_shuffle: {
2265 // Drop in-chain and replace undefs, but otherwise pass through unchanged
2266 SDValue Ops[18];
2267 size_t OpIdx = 0;
2268 Ops[OpIdx++] = Op.getOperand(i: 1);
2269 Ops[OpIdx++] = Op.getOperand(i: 2);
2270 while (OpIdx < 18) {
2271 const SDValue &MaskIdx = Op.getOperand(i: OpIdx + 1);
2272 if (MaskIdx.isUndef() || MaskIdx.getNode()->getAsZExtVal() >= 32) {
2273 bool isTarget = MaskIdx.getNode()->getOpcode() == ISD::TargetConstant;
2274 Ops[OpIdx++] = DAG.getConstant(Val: 0, DL, VT: MVT::i32, isTarget);
2275 } else {
2276 Ops[OpIdx++] = MaskIdx;
2277 }
2278 }
2279 return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops);
2280 }
2281
2282 case Intrinsic::wasm_funcref_to_ptr: {
2283 // llvm.wasm.funcref.to_ptr only has a defined lowering when its result
2284 // feeds directly into an indirect call. Reaching here means the pointer
2285 // escapes a direct call. We haven't implemented conversion of a funcref
2286 // into a real function pointer so we crash if we get here.
2287 fail(DL, DAG,
2288 Msg: "a funcref can only be converted to a pointer to be directly called; "
2289 "the resulting pointer cannot otherwise be used");
2290 return DAG.getPOISON(VT: Op.getValueType());
2291 }
2292
2293 case Intrinsic::thread_pointer: {
2294 return SDValue(WebAssembly::getTLSBase(DAG, DL, Subtarget), 0);
2295 }
2296 }
2297}
2298
2299SDValue
2300WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
2301 SelectionDAG &DAG) const {
2302 SDLoc DL(Op);
2303 // If sign extension operations are disabled, allow sext_inreg only if operand
2304 // is a vector extract of an i8 or i16 lane. SIMD does not depend on sign
2305 // extension operations, but allowing sext_inreg in this context lets us have
2306 // simple patterns to select extract_lane_s instructions. Expanding sext_inreg
2307 // everywhere would be simpler in this file, but would necessitate large and
2308 // brittle patterns to undo the expansion and select extract_lane_s
2309 // instructions.
2310 assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128());
2311 if (Op.getOperand(i: 0).getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2312 return SDValue();
2313
2314 const SDValue &Extract = Op.getOperand(i: 0);
2315 MVT VecT = Extract.getOperand(i: 0).getSimpleValueType();
2316 if (VecT.getVectorElementType().getSizeInBits() > 32)
2317 return SDValue();
2318 MVT ExtractedLaneT =
2319 cast<VTSDNode>(Val: Op.getOperand(i: 1).getNode())->getVT().getSimpleVT();
2320 MVT ExtractedVecT =
2321 MVT::getVectorVT(VT: ExtractedLaneT, NumElements: 128 / ExtractedLaneT.getSizeInBits());
2322 if (ExtractedVecT == VecT)
2323 return Op;
2324
2325 // Bitcast vector to appropriate type to ensure ISel pattern coverage
2326 const SDNode *Index = Extract.getOperand(i: 1).getNode();
2327 if (!isa<ConstantSDNode>(Val: Index))
2328 return SDValue();
2329 unsigned IndexVal = Index->getAsZExtVal();
2330 unsigned Scale =
2331 ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements();
2332 assert(Scale > 1);
2333 SDValue NewIndex =
2334 DAG.getConstant(Val: IndexVal * Scale, DL, VT: Index->getValueType(ResNo: 0));
2335 SDValue NewExtract = DAG.getNode(
2336 Opcode: ISD::EXTRACT_VECTOR_ELT, DL, VT: Extract.getValueType(),
2337 N1: DAG.getBitcast(VT: ExtractedVecT, V: Extract.getOperand(i: 0)), N2: NewIndex);
2338 return DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: Op.getValueType(), N1: NewExtract,
2339 N2: Op.getOperand(i: 1));
2340}
2341
2342static SDValue GetExtendHigh(SDValue Op, unsigned UserOpc, EVT VT,
2343 SelectionDAG &DAG) {
2344 SDValue Source = peekThroughBitcasts(V: Op);
2345 if (Source.getOpcode() != ISD::VECTOR_SHUFFLE)
2346 return SDValue();
2347
2348 assert((UserOpc == WebAssemblyISD::EXTEND_LOW_U ||
2349 UserOpc == WebAssemblyISD::EXTEND_LOW_S) &&
2350 "expected extend_low");
2351 auto *Shuffle = cast<ShuffleVectorSDNode>(Val: Source.getNode());
2352
2353 ArrayRef<int> Mask = Shuffle->getMask();
2354 // Look for a shuffle which moves from the high half to the low half.
2355 size_t FirstIdx = Mask.size() / 2;
2356 for (size_t i = 0; i < Mask.size() / 2; ++i) {
2357 if (Mask[i] != static_cast<int>(FirstIdx + i)) {
2358 return SDValue();
2359 }
2360 }
2361
2362 SDLoc DL(Op);
2363 unsigned Opc = UserOpc == WebAssemblyISD::EXTEND_LOW_S
2364 ? WebAssemblyISD::EXTEND_HIGH_S
2365 : WebAssemblyISD::EXTEND_HIGH_U;
2366 SDValue ShuffleSrc = Shuffle->getOperand(Num: 0);
2367 if (Op.getOpcode() == ISD::BITCAST)
2368 ShuffleSrc = DAG.getBitcast(VT: Op.getValueType(), V: ShuffleSrc);
2369
2370 return DAG.getNode(Opcode: Opc, DL, VT, Operand: ShuffleSrc);
2371}
2372
2373SDValue
2374WebAssemblyTargetLowering::LowerEXTEND_VECTOR_INREG(SDValue Op,
2375 SelectionDAG &DAG) const {
2376 SDLoc DL(Op);
2377 EVT VT = Op.getValueType();
2378 SDValue Src = Op.getOperand(i: 0);
2379 EVT SrcVT = Src.getValueType();
2380
2381 if (SrcVT.getVectorElementType() == MVT::i1 ||
2382 SrcVT.getVectorElementType() == MVT::i64)
2383 return SDValue();
2384
2385 assert(VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits() == 0 &&
2386 "Unexpected extension factor.");
2387 unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits();
2388
2389 if (Scale != 2 && Scale != 4 && Scale != 8)
2390 return SDValue();
2391
2392 unsigned Ext;
2393 switch (Op.getOpcode()) {
2394 default:
2395 llvm_unreachable("unexpected opcode");
2396 case ISD::ANY_EXTEND_VECTOR_INREG:
2397 case ISD::ZERO_EXTEND_VECTOR_INREG:
2398 Ext = WebAssemblyISD::EXTEND_LOW_U;
2399 break;
2400 case ISD::SIGN_EXTEND_VECTOR_INREG:
2401 Ext = WebAssemblyISD::EXTEND_LOW_S;
2402 break;
2403 }
2404
2405 if (Scale == 2) {
2406 // See if we can use EXTEND_HIGH.
2407 if (auto ExtendHigh = GetExtendHigh(Op: Op.getOperand(i: 0), UserOpc: Ext, VT, DAG))
2408 return ExtendHigh;
2409 }
2410
2411 SDValue Ret = Src;
2412 while (Scale != 1) {
2413 Ret = DAG.getNode(Opcode: Ext, DL,
2414 VT: Ret.getValueType()
2415 .widenIntegerVectorElementType(Context&: *DAG.getContext())
2416 .getHalfNumVectorElementsVT(Context&: *DAG.getContext()),
2417 Operand: Ret);
2418 Scale /= 2;
2419 }
2420 assert(Ret.getValueType() == VT);
2421 return Ret;
2422}
2423
2424static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
2425 SDLoc DL(Op);
2426 if (Op.getValueType() != MVT::v2f64 && Op.getValueType() != MVT::v4f32)
2427 return SDValue();
2428
2429 auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec,
2430 unsigned &Index) -> bool {
2431 switch (Op.getOpcode()) {
2432 case ISD::SINT_TO_FP:
2433 Opcode = WebAssemblyISD::CONVERT_LOW_S;
2434 break;
2435 case ISD::UINT_TO_FP:
2436 Opcode = WebAssemblyISD::CONVERT_LOW_U;
2437 break;
2438 case ISD::FP_EXTEND:
2439 case ISD::FP16_TO_FP:
2440 Opcode = WebAssemblyISD::PROMOTE_LOW;
2441 break;
2442 default:
2443 return false;
2444 }
2445
2446 auto ExtractVector = Op.getOperand(i: 0);
2447 if (ExtractVector.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2448 return false;
2449
2450 if (!isa<ConstantSDNode>(Val: ExtractVector.getOperand(i: 1).getNode()))
2451 return false;
2452
2453 SrcVec = ExtractVector.getOperand(i: 0);
2454 Index = ExtractVector.getConstantOperandVal(i: 1);
2455 return true;
2456 };
2457
2458 unsigned NumLanes = Op.getValueType() == MVT::v2f64 ? 2 : 4;
2459 unsigned FirstOpcode = 0, SecondOpcode = 0, ThirdOpcode = 0, FourthOpcode = 0;
2460 unsigned FirstIndex = 0, SecondIndex = 0, ThirdIndex = 0, FourthIndex = 0;
2461 SDValue FirstSrcVec, SecondSrcVec, ThirdSrcVec, FourthSrcVec;
2462
2463 if (!GetConvertedLane(Op.getOperand(i: 0), FirstOpcode, FirstSrcVec,
2464 FirstIndex) ||
2465 !GetConvertedLane(Op.getOperand(i: 1), SecondOpcode, SecondSrcVec,
2466 SecondIndex))
2467 return SDValue();
2468
2469 // If we're converting to v4f32, check the third and fourth lanes, too.
2470 if (NumLanes == 4 && (!GetConvertedLane(Op.getOperand(i: 2), ThirdOpcode,
2471 ThirdSrcVec, ThirdIndex) ||
2472 !GetConvertedLane(Op.getOperand(i: 3), FourthOpcode,
2473 FourthSrcVec, FourthIndex)))
2474 return SDValue();
2475
2476 if (FirstOpcode != SecondOpcode)
2477 return SDValue();
2478
2479 // TODO Add an optimization similar to the v2f64 below for shuffling the
2480 // vectors when the lanes are in the wrong order or come from different src
2481 // vectors.
2482 if (NumLanes == 4 &&
2483 (FirstOpcode != ThirdOpcode || FirstOpcode != FourthOpcode ||
2484 FirstSrcVec != SecondSrcVec || FirstSrcVec != ThirdSrcVec ||
2485 FirstSrcVec != FourthSrcVec || FirstIndex != 0 || SecondIndex != 1 ||
2486 ThirdIndex != 2 || FourthIndex != 3))
2487 return SDValue();
2488
2489 MVT ExpectedSrcVT;
2490 switch (FirstOpcode) {
2491 case WebAssemblyISD::CONVERT_LOW_S:
2492 case WebAssemblyISD::CONVERT_LOW_U:
2493 ExpectedSrcVT = MVT::v4i32;
2494 break;
2495 case WebAssemblyISD::PROMOTE_LOW:
2496 ExpectedSrcVT = NumLanes == 2 ? MVT::v4f32 : MVT::v8i16;
2497 break;
2498 }
2499 if (FirstSrcVec.getValueType() != ExpectedSrcVT)
2500 return SDValue();
2501
2502 auto Src = FirstSrcVec;
2503 if (NumLanes == 2 &&
2504 (FirstIndex != 0 || SecondIndex != 1 || FirstSrcVec != SecondSrcVec)) {
2505 // Shuffle the source vector so that the converted lanes are the low lanes.
2506 Src = DAG.getVectorShuffle(VT: ExpectedSrcVT, dl: DL, N1: FirstSrcVec, N2: SecondSrcVec,
2507 Mask: {static_cast<int>(FirstIndex),
2508 static_cast<int>(SecondIndex) + 4, -1, -1});
2509 }
2510 return DAG.getNode(Opcode: FirstOpcode, DL, VT: NumLanes == 2 ? MVT::v2f64 : MVT::v4f32,
2511 Operand: Src);
2512}
2513
2514SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,
2515 SelectionDAG &DAG) const {
2516 MVT VT = Op.getSimpleValueType();
2517 if (VT == MVT::v8f16) {
2518 // BUILD_VECTOR can't handle FP16 operands since Wasm doesn't have a scalar
2519 // FP16 type, so cast them to I16s.
2520 MVT IVT = VT.changeVectorElementType(EltVT: MVT::i16);
2521 SmallVector<SDValue, 8> NewOps;
2522 for (unsigned I = 0, E = Op.getNumOperands(); I < E; ++I)
2523 NewOps.push_back(Elt: DAG.getBitcast(VT: MVT::i16, V: Op.getOperand(i: I)));
2524 SDValue Res = DAG.getNode(Opcode: ISD::BUILD_VECTOR, DL: SDLoc(), VT: IVT, Ops: NewOps);
2525 return DAG.getBitcast(VT, V: Res);
2526 }
2527
2528 if (auto ConvertLow = LowerConvertLow(Op, DAG))
2529 return ConvertLow;
2530
2531 SDLoc DL(Op);
2532 const EVT VecT = Op.getValueType();
2533 const EVT LaneT = Op.getOperand(i: 0).getValueType();
2534 const size_t Lanes = Op.getNumOperands();
2535 bool CanSwizzle = VecT == MVT::v16i8;
2536
2537 // BUILD_VECTORs are lowered to the instruction that initializes the highest
2538 // possible number of lanes at once followed by a sequence of replace_lane
2539 // instructions to individually initialize any remaining lanes.
2540
2541 // TODO: Tune this. For example, lanewise swizzling is very expensive, so
2542 // swizzled lanes should be given greater weight.
2543
2544 // TODO: Investigate looping rather than always extracting/replacing specific
2545 // lanes to fill gaps.
2546
2547 auto IsConstant = [](const SDValue &V) {
2548 return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP;
2549 };
2550
2551 // Returns the source vector and index vector pair if they exist. Checks for:
2552 // (extract_vector_elt
2553 // $src,
2554 // (sign_extend_inreg (extract_vector_elt $indices, $i))
2555 // )
2556 auto GetSwizzleSrcs = [](size_t I, const SDValue &Lane) {
2557 auto Bail = std::make_pair(x: SDValue(), y: SDValue());
2558 if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2559 return Bail;
2560 const SDValue &SwizzleSrc = Lane->getOperand(Num: 0);
2561 const SDValue &IndexExt = Lane->getOperand(Num: 1);
2562 if (IndexExt->getOpcode() != ISD::SIGN_EXTEND_INREG)
2563 return Bail;
2564 const SDValue &Index = IndexExt->getOperand(Num: 0);
2565 if (Index->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2566 return Bail;
2567 const SDValue &SwizzleIndices = Index->getOperand(Num: 0);
2568 if (SwizzleSrc.getValueType() != MVT::v16i8 ||
2569 SwizzleIndices.getValueType() != MVT::v16i8 ||
2570 Index->getOperand(Num: 1)->getOpcode() != ISD::Constant ||
2571 Index->getConstantOperandVal(Num: 1) != I)
2572 return Bail;
2573 return std::make_pair(x: SwizzleSrc, y: SwizzleIndices);
2574 };
2575
2576 // If the lane is extracted from another vector at a constant index, return
2577 // that vector. The source vector must not have more lanes than the dest
2578 // because the shufflevector indices are in terms of the destination lanes and
2579 // would not be able to address the smaller individual source lanes.
2580 auto GetShuffleSrc = [&](const SDValue &Lane) {
2581 if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2582 return SDValue();
2583 if (!isa<ConstantSDNode>(Val: Lane->getOperand(Num: 1).getNode()))
2584 return SDValue();
2585 if (Lane->getOperand(Num: 0).getValueType().getVectorNumElements() >
2586 VecT.getVectorNumElements())
2587 return SDValue();
2588 return Lane->getOperand(Num: 0);
2589 };
2590
2591 using ValueEntry = std::pair<SDValue, size_t>;
2592 SmallVector<ValueEntry, 16> SplatValueCounts;
2593
2594 using SwizzleEntry = std::pair<std::pair<SDValue, SDValue>, size_t>;
2595 SmallVector<SwizzleEntry, 16> SwizzleCounts;
2596
2597 using ShuffleEntry = std::pair<SDValue, size_t>;
2598 SmallVector<ShuffleEntry, 16> ShuffleCounts;
2599
2600 auto AddCount = [](auto &Counts, const auto &Val) {
2601 auto CountIt =
2602 llvm::find_if(Counts, [&Val](auto E) { return E.first == Val; });
2603 if (CountIt == Counts.end()) {
2604 Counts.emplace_back(Val, 1);
2605 } else {
2606 CountIt->second++;
2607 }
2608 };
2609
2610 auto GetMostCommon = [](auto &Counts) {
2611 auto CommonIt = llvm::max_element(Counts, llvm::less_second());
2612 assert(CommonIt != Counts.end() && "Unexpected all-undef build_vector");
2613 return *CommonIt;
2614 };
2615
2616 size_t NumConstantLanes = 0;
2617
2618 // Count eligible lanes for each type of vector creation op
2619 for (size_t I = 0; I < Lanes; ++I) {
2620 const SDValue &Lane = Op->getOperand(Num: I);
2621 if (Lane.isUndef())
2622 continue;
2623
2624 AddCount(SplatValueCounts, Lane);
2625
2626 if (IsConstant(Lane))
2627 NumConstantLanes++;
2628 if (auto ShuffleSrc = GetShuffleSrc(Lane))
2629 AddCount(ShuffleCounts, ShuffleSrc);
2630 if (CanSwizzle) {
2631 auto SwizzleSrcs = GetSwizzleSrcs(I, Lane);
2632 if (SwizzleSrcs.first)
2633 AddCount(SwizzleCounts, SwizzleSrcs);
2634 }
2635 }
2636
2637 SDValue SplatValue;
2638 size_t NumSplatLanes;
2639 std::tie(args&: SplatValue, args&: NumSplatLanes) = GetMostCommon(SplatValueCounts);
2640
2641 SDValue SwizzleSrc;
2642 SDValue SwizzleIndices;
2643 size_t NumSwizzleLanes = 0;
2644 if (SwizzleCounts.size())
2645 std::forward_as_tuple(args: std::tie(args&: SwizzleSrc, args&: SwizzleIndices),
2646 args&: NumSwizzleLanes) = GetMostCommon(SwizzleCounts);
2647
2648 // Shuffles can draw from up to two vectors, so find the two most common
2649 // sources.
2650 SDValue ShuffleSrc1, ShuffleSrc2;
2651 size_t NumShuffleLanes = 0;
2652 if (ShuffleCounts.size()) {
2653 std::tie(args&: ShuffleSrc1, args&: NumShuffleLanes) = GetMostCommon(ShuffleCounts);
2654 llvm::erase_if(C&: ShuffleCounts,
2655 P: [&](const auto &Pair) { return Pair.first == ShuffleSrc1; });
2656 }
2657 if (ShuffleCounts.size()) {
2658 size_t AdditionalShuffleLanes;
2659 std::tie(args&: ShuffleSrc2, args&: AdditionalShuffleLanes) =
2660 GetMostCommon(ShuffleCounts);
2661 NumShuffleLanes += AdditionalShuffleLanes;
2662 }
2663
2664 // Predicate returning true if the lane is properly initialized by the
2665 // original instruction
2666 std::function<bool(size_t, const SDValue &)> IsLaneConstructed;
2667 SDValue Result;
2668 // Prefer swizzles over shuffles over vector consts over splats
2669 if (NumSwizzleLanes >= NumShuffleLanes &&
2670 NumSwizzleLanes >= NumConstantLanes && NumSwizzleLanes >= NumSplatLanes) {
2671 Result = DAG.getNode(Opcode: WebAssemblyISD::SWIZZLE, DL, VT: VecT, N1: SwizzleSrc,
2672 N2: SwizzleIndices);
2673 auto Swizzled = std::make_pair(x&: SwizzleSrc, y&: SwizzleIndices);
2674 IsLaneConstructed = [&, Swizzled](size_t I, const SDValue &Lane) {
2675 return Swizzled == GetSwizzleSrcs(I, Lane);
2676 };
2677 } else if (NumShuffleLanes >= NumConstantLanes &&
2678 NumShuffleLanes >= NumSplatLanes) {
2679 size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits() / 8;
2680 size_t DestLaneCount = VecT.getVectorNumElements();
2681 size_t Scale1 = 1;
2682 size_t Scale2 = 1;
2683 SDValue Src1 = ShuffleSrc1;
2684 SDValue Src2 = ShuffleSrc2 ? ShuffleSrc2 : DAG.getUNDEF(VT: VecT);
2685 if (Src1.getValueType() != VecT) {
2686 size_t LaneSize =
2687 Src1.getValueType().getVectorElementType().getFixedSizeInBits() / 8;
2688 assert(LaneSize > DestLaneSize);
2689 Scale1 = LaneSize / DestLaneSize;
2690 Src1 = DAG.getBitcast(VT: VecT, V: Src1);
2691 }
2692 if (Src2.getValueType() != VecT) {
2693 size_t LaneSize =
2694 Src2.getValueType().getVectorElementType().getFixedSizeInBits() / 8;
2695 assert(LaneSize > DestLaneSize);
2696 Scale2 = LaneSize / DestLaneSize;
2697 Src2 = DAG.getBitcast(VT: VecT, V: Src2);
2698 }
2699
2700 int Mask[16];
2701 assert(DestLaneCount <= 16);
2702 for (size_t I = 0; I < DestLaneCount; ++I) {
2703 const SDValue &Lane = Op->getOperand(Num: I);
2704 SDValue Src = GetShuffleSrc(Lane);
2705 if (Src == ShuffleSrc1) {
2706 Mask[I] = Lane->getConstantOperandVal(Num: 1) * Scale1;
2707 } else if (Src && Src == ShuffleSrc2) {
2708 Mask[I] = DestLaneCount + Lane->getConstantOperandVal(Num: 1) * Scale2;
2709 } else {
2710 Mask[I] = -1;
2711 }
2712 }
2713 ArrayRef<int> MaskRef(Mask, DestLaneCount);
2714 Result = DAG.getVectorShuffle(VT: VecT, dl: DL, N1: Src1, N2: Src2, Mask: MaskRef);
2715 IsLaneConstructed = [&](size_t, const SDValue &Lane) {
2716 auto Src = GetShuffleSrc(Lane);
2717 return Src == ShuffleSrc1 || (Src && Src == ShuffleSrc2);
2718 };
2719 } else if (NumConstantLanes >= NumSplatLanes) {
2720 SmallVector<SDValue, 16> ConstLanes;
2721 for (const SDValue &Lane : Op->op_values()) {
2722 if (IsConstant(Lane)) {
2723 // Values may need to be fixed so that they will sign extend to be
2724 // within the expected range during ISel. Check whether the value is in
2725 // bounds based on the lane bit width and if it is out of bounds, lop
2726 // off the extra bits.
2727 uint64_t LaneBits = 128 / Lanes;
2728 if (auto *Const = dyn_cast<ConstantSDNode>(Val: Lane.getNode())) {
2729 ConstLanes.push_back(Elt: DAG.getConstant(
2730 Val: Const->getAPIntValue().trunc(width: LaneBits).getZExtValue(),
2731 DL: SDLoc(Lane), VT: LaneT));
2732 } else {
2733 ConstLanes.push_back(Elt: Lane);
2734 }
2735 } else if (LaneT.isFloatingPoint()) {
2736 ConstLanes.push_back(Elt: DAG.getConstantFP(Val: 0, DL, VT: LaneT));
2737 } else {
2738 ConstLanes.push_back(Elt: DAG.getConstant(Val: 0, DL, VT: LaneT));
2739 }
2740 }
2741 Result = DAG.getBuildVector(VT: VecT, DL, Ops: ConstLanes);
2742 IsLaneConstructed = [&IsConstant](size_t _, const SDValue &Lane) {
2743 return IsConstant(Lane);
2744 };
2745 } else {
2746 size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits();
2747 if (NumSplatLanes == 1 && Op->getOperand(Num: 0) == SplatValue &&
2748 (DestLaneSize == 32 || DestLaneSize == 64)) {
2749 // Could be selected to load_zero.
2750 Result = DAG.getNode(Opcode: ISD::SCALAR_TO_VECTOR, DL, VT: VecT, Operand: SplatValue);
2751 } else {
2752 // Use a splat (which might be selected as a load splat)
2753 Result = DAG.getSplatBuildVector(VT: VecT, DL, Op: SplatValue);
2754 }
2755 IsLaneConstructed = [&SplatValue](size_t _, const SDValue &Lane) {
2756 return Lane == SplatValue;
2757 };
2758 }
2759
2760 assert(Result);
2761 assert(IsLaneConstructed);
2762
2763 // Add replace_lane instructions for any unhandled values
2764 for (size_t I = 0; I < Lanes; ++I) {
2765 const SDValue &Lane = Op->getOperand(Num: I);
2766 if (!Lane.isUndef() && !IsLaneConstructed(I, Lane))
2767 Result = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL, VT: VecT, N1: Result, N2: Lane,
2768 N3: DAG.getConstant(Val: I, DL, VT: MVT::i32));
2769 }
2770
2771 return Result;
2772}
2773
2774SDValue
2775WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
2776 SelectionDAG &DAG) const {
2777 SDLoc DL(Op);
2778 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Val: Op.getNode())->getMask();
2779 MVT VecType = Op.getOperand(i: 0).getSimpleValueType();
2780 assert(VecType.is128BitVector() && "Unexpected shuffle vector type");
2781 size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8;
2782
2783 // Space for two vector args and sixteen mask indices
2784 SDValue Ops[18];
2785 size_t OpIdx = 0;
2786 Ops[OpIdx++] = Op.getOperand(i: 0);
2787 Ops[OpIdx++] = Op.getOperand(i: 1);
2788
2789 // Expand mask indices to byte indices and materialize them as operands
2790 for (int M : Mask) {
2791 for (size_t J = 0; J < LaneBytes; ++J) {
2792 // Lower undefs (represented by -1 in mask) to {0..J}, which use a
2793 // whole lane of vector input, to allow further reduction at VM. E.g.
2794 // match an 8x16 byte shuffle to an equivalent cheaper 32x4 shuffle.
2795 uint64_t ByteIndex = M == -1 ? J : (uint64_t)M * LaneBytes + J;
2796 Ops[OpIdx++] = DAG.getConstant(Val: ByteIndex, DL, VT: MVT::i32);
2797 }
2798 }
2799
2800 return DAG.getNode(Opcode: WebAssemblyISD::SHUFFLE, DL, VT: Op.getValueType(), Ops);
2801}
2802
2803SDValue WebAssemblyTargetLowering::LowerSETCC(SDValue Op,
2804 SelectionDAG &DAG) const {
2805 SDLoc DL(Op);
2806 // The legalizer does not know how to expand the unsupported comparison modes
2807 // of i64x2 vectors, so we manually unroll them here.
2808 assert(Op->getOperand(0)->getSimpleValueType(0) == MVT::v2i64);
2809 SmallVector<SDValue, 2> LHS, RHS;
2810 DAG.ExtractVectorElements(Op: Op->getOperand(Num: 0), Args&: LHS);
2811 DAG.ExtractVectorElements(Op: Op->getOperand(Num: 1), Args&: RHS);
2812 const SDValue &CC = Op->getOperand(Num: 2);
2813 auto MakeLane = [&](unsigned I) {
2814 return DAG.getNode(Opcode: ISD::SELECT_CC, DL, VT: MVT::i64, N1: LHS[I], N2: RHS[I],
2815 N3: DAG.getConstant(Val: uint64_t(-1), DL, VT: MVT::i64),
2816 N4: DAG.getConstant(Val: uint64_t(0), DL, VT: MVT::i64), N5: CC);
2817 };
2818 return DAG.getBuildVector(VT: Op->getValueType(ResNo: 0), DL,
2819 Ops: {MakeLane(0), MakeLane(1)});
2820}
2821
2822SDValue
2823WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op,
2824 SelectionDAG &DAG) const {
2825 if (Op.getOpcode() == ISD::INSERT_VECTOR_ELT &&
2826 Op.getValueType() == MVT::v8f16) {
2827 // INSERT_VECTOR_ELT can't handle FP16 operands since Wasm doesn't have a
2828 // scalar FP16 type, so cast them to I16s.
2829 SDLoc DL(Op);
2830 SDValue IntVector = DAG.getBitcast(VT: MVT::v8i16, V: Op.getOperand(i: 0));
2831 SDValue IntElement = DAG.getBitcast(VT: MVT::i16, V: Op.getOperand(i: 1));
2832 SDValue Inserted = DAG.getNode(Opcode: ISD::INSERT_VECTOR_ELT, DL, VT: MVT::v8i16,
2833 N1: IntVector, N2: IntElement, N3: Op.getOperand(i: 2));
2834 return DAG.getBitcast(VT: MVT::v8f16, V: Inserted);
2835 }
2836
2837 // Allow constant lane indices, expand variable lane indices
2838 SDNode *IdxNode = Op.getOperand(i: Op.getNumOperands() - 1).getNode();
2839 if (isa<ConstantSDNode>(Val: IdxNode)) {
2840 // Ensure the index type is i32 to match the tablegen patterns
2841 uint64_t Idx = IdxNode->getAsZExtVal();
2842 SmallVector<SDValue, 3> Ops(Op.getNode()->ops());
2843 Ops[Op.getNumOperands() - 1] =
2844 DAG.getConstant(Val: Idx, DL: SDLoc(IdxNode), VT: MVT::i32);
2845 return DAG.getNode(Opcode: Op.getOpcode(), DL: SDLoc(Op), VT: Op.getValueType(), Ops);
2846 }
2847 // Perform default expansion
2848 return SDValue();
2849}
2850
2851static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) {
2852 EVT LaneT = Op.getSimpleValueType().getVectorElementType();
2853 // 32-bit and 64-bit unrolled shifts will have proper semantics
2854 if (LaneT.bitsGE(VT: MVT::i32))
2855 return DAG.UnrollVectorOp(N: Op.getNode());
2856 // Otherwise mask the shift value to get proper semantics from 32-bit shift
2857 SDLoc DL(Op);
2858 size_t NumLanes = Op.getSimpleValueType().getVectorNumElements();
2859 SDValue Mask = DAG.getConstant(Val: LaneT.getSizeInBits() - 1, DL, VT: MVT::i32);
2860 unsigned ShiftOpcode = Op.getOpcode();
2861 SmallVector<SDValue, 16> ShiftedElements;
2862 DAG.ExtractVectorElements(Op: Op.getOperand(i: 0), Args&: ShiftedElements, Start: 0, Count: 0, EltVT: MVT::i32);
2863 SmallVector<SDValue, 16> ShiftElements;
2864 DAG.ExtractVectorElements(Op: Op.getOperand(i: 1), Args&: ShiftElements, Start: 0, Count: 0, EltVT: MVT::i32);
2865 SmallVector<SDValue, 16> UnrolledOps;
2866 for (size_t i = 0; i < NumLanes; ++i) {
2867 SDValue MaskedShiftValue =
2868 DAG.getNode(Opcode: ISD::AND, DL, VT: MVT::i32, N1: ShiftElements[i], N2: Mask);
2869 SDValue ShiftedValue = ShiftedElements[i];
2870 if (ShiftOpcode == ISD::SRA)
2871 ShiftedValue = DAG.getNode(Opcode: ISD::SIGN_EXTEND_INREG, DL, VT: MVT::i32,
2872 N1: ShiftedValue, N2: DAG.getValueType(LaneT));
2873 UnrolledOps.push_back(
2874 Elt: DAG.getNode(Opcode: ShiftOpcode, DL, VT: MVT::i32, N1: ShiftedValue, N2: MaskedShiftValue));
2875 }
2876 return DAG.getBuildVector(VT: Op.getValueType(), DL, Ops: UnrolledOps);
2877}
2878
2879SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op,
2880 SelectionDAG &DAG) const {
2881 SDLoc DL(Op);
2882 // Only manually lower vector shifts
2883 assert(Op.getSimpleValueType().isVector());
2884
2885 uint64_t LaneBits = Op.getValueType().getScalarSizeInBits();
2886 auto ShiftVal = Op.getOperand(i: 1);
2887
2888 // Try to skip bitmask operation since it is implied inside shift instruction
2889 auto SkipImpliedMask = [](SDValue MaskOp, uint64_t MaskBits) {
2890 if (MaskOp.getOpcode() != ISD::AND)
2891 return MaskOp;
2892 SDValue LHS = MaskOp.getOperand(i: 0);
2893 SDValue RHS = MaskOp.getOperand(i: 1);
2894 if (MaskOp.getValueType().isVector()) {
2895 APInt MaskVal;
2896 if (!ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal))
2897 std::swap(a&: LHS, b&: RHS);
2898
2899 if (ISD::isConstantSplatVector(N: RHS.getNode(), SplatValue&: MaskVal) &&
2900 MaskVal == MaskBits)
2901 MaskOp = LHS;
2902 } else {
2903 if (!isa<ConstantSDNode>(Val: RHS.getNode()))
2904 std::swap(a&: LHS, b&: RHS);
2905
2906 auto ConstantRHS = dyn_cast<ConstantSDNode>(Val: RHS.getNode());
2907 if (ConstantRHS && ConstantRHS->getAPIntValue() == MaskBits)
2908 MaskOp = LHS;
2909 }
2910
2911 return MaskOp;
2912 };
2913
2914 // Skip vector and operation
2915 ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1);
2916 ShiftVal = DAG.getSplatValue(V: ShiftVal);
2917 if (!ShiftVal)
2918 return unrollVectorShift(Op, DAG);
2919
2920 // Skip scalar and operation
2921 ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1);
2922 // Use anyext because none of the high bits can affect the shift
2923 ShiftVal = DAG.getAnyExtOrTrunc(Op: ShiftVal, DL, VT: MVT::i32);
2924
2925 unsigned Opcode;
2926 switch (Op.getOpcode()) {
2927 case ISD::SHL:
2928 Opcode = WebAssemblyISD::VEC_SHL;
2929 break;
2930 case ISD::SRA:
2931 Opcode = WebAssemblyISD::VEC_SHR_S;
2932 break;
2933 case ISD::SRL:
2934 Opcode = WebAssemblyISD::VEC_SHR_U;
2935 break;
2936 default:
2937 llvm_unreachable("unexpected opcode");
2938 }
2939
2940 return DAG.getNode(Opcode, DL, VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: ShiftVal);
2941}
2942
2943SDValue WebAssemblyTargetLowering::LowerFP_TO_INT_SAT(SDValue Op,
2944 SelectionDAG &DAG) const {
2945 EVT ResT = Op.getValueType();
2946 EVT SatVT = cast<VTSDNode>(Val: Op.getOperand(i: 1))->getVT();
2947
2948 if ((ResT == MVT::i32 || ResT == MVT::i64) &&
2949 (SatVT == MVT::i32 || SatVT == MVT::i64))
2950 return Op;
2951
2952 if (ResT == MVT::v4i32 && SatVT == MVT::i32)
2953 return Op;
2954
2955 if (ResT == MVT::v8i16 && SatVT == MVT::i16)
2956 return Op;
2957
2958 return SDValue();
2959}
2960
2961static bool HasNoSignedZerosOrNaNs(SDValue Op, SelectionDAG &DAG) {
2962 return (Op->getFlags().hasNoNaNs() ||
2963 (DAG.isKnownNeverNaN(Op: Op->getOperand(Num: 0)) &&
2964 DAG.isKnownNeverNaN(Op: Op->getOperand(Num: 1)))) &&
2965 (Op->getFlags().hasNoSignedZeros() ||
2966 DAG.isKnownNeverLogicalZero(Op: Op->getOperand(Num: 0)) ||
2967 DAG.isKnownNeverLogicalZero(Op: Op->getOperand(Num: 1)));
2968}
2969
2970SDValue WebAssemblyTargetLowering::LowerFMIN(SDValue Op,
2971 SelectionDAG &DAG) const {
2972 if (Subtarget->hasRelaxedSIMD() && HasNoSignedZerosOrNaNs(Op, DAG)) {
2973 return DAG.getNode(Opcode: WebAssemblyISD::RELAXED_FMIN, DL: SDLoc(Op),
2974 VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: Op.getOperand(i: 1));
2975 }
2976 return SDValue();
2977}
2978
2979SDValue WebAssemblyTargetLowering::LowerFMAX(SDValue Op,
2980 SelectionDAG &DAG) const {
2981 if (Subtarget->hasRelaxedSIMD() && HasNoSignedZerosOrNaNs(Op, DAG)) {
2982 return DAG.getNode(Opcode: WebAssemblyISD::RELAXED_FMAX, DL: SDLoc(Op),
2983 VT: Op.getValueType(), N1: Op.getOperand(i: 0), N2: Op.getOperand(i: 1));
2984 }
2985 return SDValue();
2986}
2987
2988//===----------------------------------------------------------------------===//
2989// Custom DAG combine hooks
2990//===----------------------------------------------------------------------===//
2991static SDValue
2992performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
2993 auto &DAG = DCI.DAG;
2994 auto Shuffle = cast<ShuffleVectorSDNode>(Val: N);
2995
2996 // Hoist vector bitcasts that don't change the number of lanes out of unary
2997 // shuffles, where they are less likely to get in the way of other combines.
2998 // (shuffle (vNxT1 (bitcast (vNxT0 x))), undef, mask) ->
2999 // (vNxT1 (bitcast (vNxT0 (shuffle x, undef, mask))))
3000 SDValue Bitcast = N->getOperand(Num: 0);
3001 if (Bitcast.getOpcode() != ISD::BITCAST)
3002 return SDValue();
3003 if (!N->getOperand(Num: 1).isUndef())
3004 return SDValue();
3005 SDValue CastOp = Bitcast.getOperand(i: 0);
3006 EVT SrcType = CastOp.getValueType();
3007 EVT DstType = Bitcast.getValueType();
3008 if (!SrcType.is128BitVector() ||
3009 SrcType.getVectorNumElements() != DstType.getVectorNumElements())
3010 return SDValue();
3011 SDValue NewShuffle = DAG.getVectorShuffle(
3012 VT: SrcType, dl: SDLoc(N), N1: CastOp, N2: DAG.getUNDEF(VT: SrcType), Mask: Shuffle->getMask());
3013 return DAG.getBitcast(VT: DstType, V: NewShuffle);
3014}
3015
3016/// Convert ({u,s}itofp vec) --> ({u,s}itofp ({s,z}ext vec)) so it doesn't get
3017/// split up into scalar instructions during legalization, and the vector
3018/// extending instructions are selected in performVectorExtendCombine below.
3019static SDValue
3020performVectorExtendToFPCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
3021 const WebAssemblySubtarget *Subtarget) {
3022 auto &DAG = DCI.DAG;
3023 assert(N->getOpcode() == ISD::UINT_TO_FP ||
3024 N->getOpcode() == ISD::SINT_TO_FP);
3025
3026 EVT InVT = N->getOperand(Num: 0)->getValueType(ResNo: 0);
3027 EVT ResVT = N->getValueType(ResNo: 0);
3028 MVT ExtVT;
3029 if (ResVT == MVT::v4f32 && (InVT == MVT::v4i16 || InVT == MVT::v4i8))
3030 ExtVT = MVT::v4i32;
3031 else if (ResVT == MVT::v2f64 && (InVT == MVT::v2i16 || InVT == MVT::v2i8))
3032 ExtVT = MVT::v2i32;
3033 else if (Subtarget->hasFP16() && ResVT == MVT::v8f16 && InVT == MVT::v8i8)
3034 ExtVT = MVT::v8i16;
3035 else
3036 return SDValue();
3037
3038 unsigned Op =
3039 N->getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
3040 SDValue Conv = DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ExtVT, Operand: N->getOperand(Num: 0));
3041 return DAG.getNode(Opcode: N->getOpcode(), DL: SDLoc(N), VT: ResVT, Operand: Conv);
3042}
3043
3044static SDValue
3045performVectorNonNegToFPCombine(SDNode *N,
3046 TargetLowering::DAGCombinerInfo &DCI) {
3047 auto &DAG = DCI.DAG;
3048
3049 SDNodeFlags Flags = N->getFlags();
3050 SDValue Op0 = N->getOperand(Num: 0);
3051 EVT VT = N->getValueType(ResNo: 0);
3052
3053 // Optimize uitofp to sitofp when the sign bit is known to be zero.
3054 // Depending on the target (runtime) backend, this might be performance
3055 // neutral (e.g. AArch64) or a significant improvement (e.g. x86_64).
3056 if (VT.isVector() && (Flags.hasNonNeg() || DAG.SignBitIsZero(Op: Op0))) {
3057 return DAG.getNode(Opcode: ISD::SINT_TO_FP, DL: SDLoc(N), VT, Operand: Op0);
3058 }
3059
3060 return SDValue();
3061}
3062
3063static SDValue
3064performVectorExtendCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
3065 auto &DAG = DCI.DAG;
3066 assert(N->getOpcode() == ISD::SIGN_EXTEND ||
3067 N->getOpcode() == ISD::ZERO_EXTEND);
3068
3069 EVT ResVT = N->getValueType(ResNo: 0);
3070 bool IsSext = N->getOpcode() == ISD::SIGN_EXTEND;
3071 SDLoc DL(N);
3072
3073 if (ResVT == MVT::v16i32 && N->getOperand(Num: 0)->getValueType(ResNo: 0) == MVT::v16i8) {
3074 // Use a tree of extend low/high to split and extend the input in two
3075 // layers to avoid doing several shuffles and even more extends.
3076 unsigned LowOp =
3077 IsSext ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U;
3078 unsigned HighOp =
3079 IsSext ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U;
3080 SDValue Input = N->getOperand(Num: 0);
3081 SDValue LowHalf = DAG.getNode(Opcode: LowOp, DL, VT: MVT::v8i16, Operand: Input);
3082 SDValue HighHalf = DAG.getNode(Opcode: HighOp, DL, VT: MVT::v8i16, Operand: Input);
3083 SDValue Subvectors[] = {
3084 DAG.getNode(Opcode: LowOp, DL, VT: MVT::v4i32, Operand: LowHalf),
3085 DAG.getNode(Opcode: HighOp, DL, VT: MVT::v4i32, Operand: LowHalf),
3086 DAG.getNode(Opcode: LowOp, DL, VT: MVT::v4i32, Operand: HighHalf),
3087 DAG.getNode(Opcode: HighOp, DL, VT: MVT::v4i32, Operand: HighHalf),
3088 };
3089 return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: ResVT, Ops: Subvectors);
3090 }
3091
3092 // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if
3093 // possible before the extract_subvector can be expanded.
3094 auto Extract = N->getOperand(Num: 0);
3095 if (Extract.getOpcode() != ISD::EXTRACT_SUBVECTOR)
3096 return SDValue();
3097 auto Source = Extract.getOperand(i: 0);
3098 auto *IndexNode = dyn_cast<ConstantSDNode>(Val: Extract.getOperand(i: 1));
3099 if (IndexNode == nullptr)
3100 return SDValue();
3101 auto Index = IndexNode->getZExtValue();
3102
3103 // Only v8i8, v4i16, and v2i32 extracts can be widened, and only if the
3104 // extracted subvector is the low or high half of its source.
3105 if (ResVT == MVT::v8i16) {
3106 if (Extract.getValueType() != MVT::v8i8 ||
3107 Source.getValueType() != MVT::v16i8 || (Index != 0 && Index != 8))
3108 return SDValue();
3109 } else if (ResVT == MVT::v4i32) {
3110 if (Extract.getValueType() != MVT::v4i16 ||
3111 Source.getValueType() != MVT::v8i16 || (Index != 0 && Index != 4))
3112 return SDValue();
3113 } else if (ResVT == MVT::v2i64) {
3114 if (Extract.getValueType() != MVT::v2i32 ||
3115 Source.getValueType() != MVT::v4i32 || (Index != 0 && Index != 2))
3116 return SDValue();
3117 } else {
3118 return SDValue();
3119 }
3120
3121 bool IsLow = Index == 0;
3122
3123 unsigned Op = IsSext ? (IsLow ? WebAssemblyISD::EXTEND_LOW_S
3124 : WebAssemblyISD::EXTEND_HIGH_S)
3125 : (IsLow ? WebAssemblyISD::EXTEND_LOW_U
3126 : WebAssemblyISD::EXTEND_HIGH_U);
3127
3128 return DAG.getNode(Opcode: Op, DL, VT: ResVT, Operand: Source);
3129}
3130
3131static SDValue
3132performVectorTruncZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
3133 auto &DAG = DCI.DAG;
3134
3135 auto GetWasmConversionOp = [](unsigned Op) {
3136 switch (Op) {
3137 case ISD::FP_TO_SINT_SAT:
3138 return WebAssemblyISD::TRUNC_SAT_ZERO_S;
3139 case ISD::FP_TO_UINT_SAT:
3140 return WebAssemblyISD::TRUNC_SAT_ZERO_U;
3141 case ISD::FP_ROUND:
3142 return WebAssemblyISD::DEMOTE_ZERO;
3143 }
3144 llvm_unreachable("unexpected op");
3145 };
3146
3147 auto IsZeroSplat = [](SDValue SplatVal) {
3148 auto *Splat = dyn_cast<BuildVectorSDNode>(Val: SplatVal.getNode());
3149 APInt SplatValue, SplatUndef;
3150 unsigned SplatBitSize;
3151 bool HasAnyUndefs;
3152 // Endianness doesn't matter in this context because we are looking for
3153 // an all-zero value.
3154 return Splat &&
3155 Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
3156 HasAnyUndefs) &&
3157 SplatValue == 0;
3158 };
3159
3160 if (N->getOpcode() == ISD::CONCAT_VECTORS) {
3161 // Combine this:
3162 //
3163 // (concat_vectors (v2i32 (fp_to_{s,u}int_sat $x, 32)), (v2i32 (splat 0)))
3164 //
3165 // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x).
3166 //
3167 // Or this:
3168 //
3169 // (concat_vectors ({v2f32, v4f16} (fp_round ({v2f64, v4f32} $x))),
3170 // ({v2f32, v4f16} (splat 0)))
3171 //
3172 // into ({f32x4, f16x8}.demote_zero_{f64x2, f32x4} $x).
3173 EVT ResVT;
3174 EVT ExpectedConversionType;
3175 auto Conversion = N->getOperand(Num: 0);
3176 auto ConversionOp = Conversion.getOpcode();
3177 switch (ConversionOp) {
3178 case ISD::FP_TO_SINT_SAT:
3179 case ISD::FP_TO_UINT_SAT:
3180 ResVT = MVT::v4i32;
3181 ExpectedConversionType = MVT::v2i32;
3182 break;
3183 case ISD::FP_ROUND:
3184 if (Conversion.getValueType() == MVT::v2f32) {
3185 ResVT = MVT::v4f32;
3186 ExpectedConversionType = MVT::v2f32;
3187 } else if (Conversion.getValueType() == MVT::v4f16) {
3188 ResVT = MVT::v8f16;
3189 ExpectedConversionType = MVT::v4f16;
3190 } else {
3191 return SDValue();
3192 }
3193 break;
3194 default:
3195 return SDValue();
3196 }
3197
3198 if (N->getValueType(ResNo: 0) != ResVT)
3199 return SDValue();
3200
3201 if (Conversion.getValueType() != ExpectedConversionType)
3202 return SDValue();
3203
3204 auto Source = Conversion.getOperand(i: 0);
3205 if (!((Source.getValueType() == MVT::v2f64 && ResVT == MVT::v4f32) ||
3206 (Source.getValueType() == MVT::v2f64 && ResVT == MVT::v4i32) ||
3207 (Source.getValueType() == MVT::v4f32 && ResVT == MVT::v8f16)))
3208 return SDValue();
3209
3210 if (!IsZeroSplat(N->getOperand(Num: 1)) ||
3211 N->getOperand(Num: 1).getValueType() != ExpectedConversionType)
3212 return SDValue();
3213
3214 unsigned Op = GetWasmConversionOp(ConversionOp);
3215 return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Source);
3216 }
3217
3218 // Combine this:
3219 //
3220 // (fp_to_{s,u}int_sat (concat_vectors $x, (v2f64 (splat 0))), 32)
3221 //
3222 // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x).
3223 //
3224 // Or this:
3225 //
3226 // ({v4f32, v8f16} (fp_round (concat_vectors $x,
3227 // ({v2f64, v4f32} (splat 0)))))
3228 //
3229 // into ({f32x4, f16x8}.demote_zero_{f64x2, f32x4} $x).
3230 EVT ResVT;
3231 auto ConversionOp = N->getOpcode();
3232 switch (ConversionOp) {
3233 case ISD::FP_TO_SINT_SAT:
3234 case ISD::FP_TO_UINT_SAT:
3235 ResVT = MVT::v4i32;
3236 break;
3237 case ISD::FP_ROUND:
3238 ResVT = N->getValueType(ResNo: 0);
3239 break;
3240 default:
3241 llvm_unreachable("unexpected op");
3242 }
3243
3244 if (N->getValueType(ResNo: 0) != ResVT)
3245 return SDValue();
3246
3247 auto Concat = N->getOperand(Num: 0);
3248 if (Concat.getOpcode() != ISD::CONCAT_VECTORS)
3249 return SDValue();
3250 EVT ConcatVT = Concat.getValueType();
3251 EVT SourceVT = Concat.getOperand(i: 0).getValueType();
3252
3253 if (!IsZeroSplat(Concat.getOperand(i: 1)))
3254 return SDValue();
3255
3256 if (ConversionOp == ISD::FP_ROUND) {
3257 bool IsF64ToF32 =
3258 ConcatVT == MVT::v4f64 && SourceVT == MVT::v2f64 && ResVT == MVT::v4f32;
3259 bool IsF32ToF16 =
3260 ConcatVT == MVT::v8f32 && SourceVT == MVT::v4f32 && ResVT == MVT::v8f16;
3261 if (!(IsF64ToF32 || IsF32ToF16))
3262 return SDValue();
3263 } else {
3264 if (ConcatVT != MVT::v4f64 || SourceVT != MVT::v2f64 || ResVT != MVT::v4i32)
3265 return SDValue();
3266 }
3267
3268 unsigned Op = GetWasmConversionOp(ConversionOp);
3269 return DAG.getNode(Opcode: Op, DL: SDLoc(N), VT: ResVT, Operand: Concat.getOperand(i: 0));
3270}
3271
3272// Helper to extract VectorWidth bits from Vec, starting from IdxVal.
3273static SDValue extractSubVector(SDValue Vec, unsigned IdxVal, SelectionDAG &DAG,
3274 const SDLoc &DL, unsigned VectorWidth) {
3275 EVT VT = Vec.getValueType();
3276 EVT ElVT = VT.getVectorElementType();
3277 unsigned Factor = VT.getSizeInBits() / VectorWidth;
3278 EVT ResultVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: ElVT,
3279 NumElements: VT.getVectorNumElements() / Factor);
3280
3281 // Extract the relevant VectorWidth bits. Generate an EXTRACT_SUBVECTOR
3282 unsigned ElemsPerChunk = VectorWidth / ElVT.getSizeInBits();
3283 assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
3284
3285 // This is the index of the first element of the VectorWidth-bit chunk
3286 // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
3287 IdxVal &= ~(ElemsPerChunk - 1);
3288
3289 // If the input is a buildvector just emit a smaller one.
3290 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
3291 return DAG.getBuildVector(VT: ResultVT, DL,
3292 Ops: Vec->ops().slice(N: IdxVal, M: ElemsPerChunk));
3293
3294 SDValue VecIdx = DAG.getIntPtrConstant(Val: IdxVal, DL);
3295 return DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: ResultVT, N1: Vec, N2: VecIdx);
3296}
3297
3298// Helper to recursively truncate vector elements in half with NARROW_U. DstVT
3299// is the expected destination value type after recursion. In is the initial
3300// input. Note that the input should have enough leading zero bits to prevent
3301// NARROW_U from saturating results.
3302static SDValue truncateVectorWithNARROW(EVT DstVT, SDValue In, const SDLoc &DL,
3303 SelectionDAG &DAG) {
3304 EVT SrcVT = In.getValueType();
3305
3306 // No truncation required, we might get here due to recursive calls.
3307 if (SrcVT == DstVT)
3308 return In;
3309
3310 unsigned SrcSizeInBits = SrcVT.getSizeInBits();
3311 unsigned NumElems = SrcVT.getVectorNumElements();
3312 if (!isPowerOf2_32(Value: NumElems))
3313 return SDValue();
3314 assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation");
3315 assert(SrcSizeInBits > DstVT.getSizeInBits() && "Illegal truncation");
3316
3317 LLVMContext &Ctx = *DAG.getContext();
3318 EVT PackedSVT = EVT::getIntegerVT(Context&: Ctx, BitWidth: SrcVT.getScalarSizeInBits() / 2);
3319
3320 // Narrow to the largest type possible:
3321 // vXi64/vXi32 -> i16x8.narrow_i32x4_u and vXi16 -> i8x16.narrow_i16x8_u.
3322 EVT InVT = MVT::i16, OutVT = MVT::i8;
3323 if (SrcVT.getScalarSizeInBits() > 16) {
3324 InVT = MVT::i32;
3325 OutVT = MVT::i16;
3326 }
3327 unsigned SubSizeInBits = SrcSizeInBits / 2;
3328 InVT = EVT::getVectorVT(Context&: Ctx, VT: InVT, NumElements: SubSizeInBits / InVT.getSizeInBits());
3329 OutVT = EVT::getVectorVT(Context&: Ctx, VT: OutVT, NumElements: SubSizeInBits / OutVT.getSizeInBits());
3330
3331 // Split lower/upper subvectors.
3332 SDValue Lo = extractSubVector(Vec: In, IdxVal: 0, DAG, DL, VectorWidth: SubSizeInBits);
3333 SDValue Hi = extractSubVector(Vec: In, IdxVal: NumElems / 2, DAG, DL, VectorWidth: SubSizeInBits);
3334
3335 // 256bit -> 128bit truncate - Narrow lower/upper 128-bit subvectors.
3336 if (SrcVT.is256BitVector() && DstVT.is128BitVector()) {
3337 Lo = DAG.getBitcast(VT: InVT, V: Lo);
3338 Hi = DAG.getBitcast(VT: InVT, V: Hi);
3339 SDValue Res = DAG.getNode(Opcode: WebAssemblyISD::NARROW_U, DL, VT: OutVT, N1: Lo, N2: Hi);
3340 return DAG.getBitcast(VT: DstVT, V: Res);
3341 }
3342
3343 // Recursively narrow lower/upper subvectors, concat result and narrow again.
3344 EVT PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems / 2);
3345 Lo = truncateVectorWithNARROW(DstVT: PackedVT, In: Lo, DL, DAG);
3346 Hi = truncateVectorWithNARROW(DstVT: PackedVT, In: Hi, DL, DAG);
3347
3348 PackedVT = EVT::getVectorVT(Context&: Ctx, VT: PackedSVT, NumElements: NumElems);
3349 SDValue Res = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: PackedVT, N1: Lo, N2: Hi);
3350 return truncateVectorWithNARROW(DstVT, In: Res, DL, DAG);
3351}
3352
3353static SDValue performTruncateCombine(SDNode *N,
3354 TargetLowering::DAGCombinerInfo &DCI) {
3355 auto &DAG = DCI.DAG;
3356
3357 SDValue In = N->getOperand(Num: 0);
3358 EVT InVT = In.getValueType();
3359 if (!InVT.isSimple())
3360 return SDValue();
3361
3362 EVT OutVT = N->getValueType(ResNo: 0);
3363 if (!OutVT.isVector())
3364 return SDValue();
3365
3366 EVT OutSVT = OutVT.getVectorElementType();
3367 EVT InSVT = InVT.getVectorElementType();
3368 // Currently only cover truncate to v16i8 or v8i16.
3369 if (!((InSVT == MVT::i16 || InSVT == MVT::i32 || InSVT == MVT::i64) &&
3370 (OutSVT == MVT::i8 || OutSVT == MVT::i16) && OutVT.is128BitVector()))
3371 return SDValue();
3372
3373 SDLoc DL(N);
3374 APInt Mask = APInt::getLowBitsSet(numBits: InVT.getScalarSizeInBits(),
3375 loBitsSet: OutVT.getScalarSizeInBits());
3376 In = DAG.getNode(Opcode: ISD::AND, DL, VT: InVT, N1: In, N2: DAG.getConstant(Val: Mask, DL, VT: InVT));
3377 return truncateVectorWithNARROW(DstVT: OutVT, In, DL, DAG);
3378}
3379
3380static SDValue performBitcastCombine(SDNode *N,
3381 TargetLowering::DAGCombinerInfo &DCI) {
3382 using namespace llvm::SDPatternMatch;
3383 auto &DAG = DCI.DAG;
3384 SDLoc DL(N);
3385 SDValue Src = N->getOperand(Num: 0);
3386 EVT VT = N->getValueType(ResNo: 0);
3387 EVT SrcVT = Src.getValueType();
3388
3389 if (!(DCI.isBeforeLegalize() && VT.isScalarInteger() &&
3390 SrcVT.isFixedLengthVectorOf(EltVT: MVT::i1)))
3391 return SDValue();
3392
3393 unsigned NumElts = SrcVT.getVectorNumElements();
3394 EVT Width = MVT::getIntegerVT(BitWidth: 128 / NumElts);
3395
3396 // bitcast <N x i1> to iN, where N = 2, 4, 8, 16 (legal)
3397 // ==> bitmask
3398 if (NumElts == 2 || NumElts == 4 || NumElts == 8 || NumElts == 16) {
3399 return DAG.getZExtOrTrunc(
3400 Op: DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32,
3401 Ops: {DAG.getConstant(Val: Intrinsic::wasm_bitmask, DL, VT: MVT::i32),
3402 DAG.getSExtOrTrunc(Op: N->getOperand(Num: 0), DL,
3403 VT: SrcVT.changeVectorElementType(
3404 Context&: *DAG.getContext(), EltVT: Width))}),
3405 DL, VT);
3406 }
3407
3408 // bitcast <N x i1>(setcc ...) to concat iN, where N = 32 and 64 (illegal)
3409 if (NumElts == 32 || NumElts == 64) {
3410 SDValue Concat, SetCCVector;
3411 ISD::CondCode SetCond;
3412
3413 if (!sd_match(N, P: m_BitCast(Op: m_c_SetCC(LHS: m_Value(N&: Concat), RHS: m_Value(N&: SetCCVector),
3414 CC: m_CondCode(CC&: SetCond)))))
3415 return SDValue();
3416 if (Concat.getOpcode() != ISD::CONCAT_VECTORS)
3417 return SDValue();
3418
3419 // Reconstruct the wide bitmask from each CONCAT_VECTORS operand.
3420 // Derive the per-chunk mask/integer types from the actual operand type
3421 // instead of hardcoding v16i1 / i16 for every chunk.
3422 EVT ConcatOperandVT = Concat.getOperand(i: 0).getValueType();
3423 unsigned ConcatOperandNumElts = ConcatOperandVT.getVectorNumElements();
3424
3425 EVT ConcatOperandMaskVT =
3426 EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1,
3427 EC: ElementCount::getFixed(MinVal: ConcatOperandNumElts));
3428 EVT ConcatOperandBitmaskVT =
3429 EVT::getIntegerVT(Context&: *DAG.getContext(), BitWidth: ConcatOperandNumElts);
3430 EVT ReturnVT = N->getValueType(ResNo: 0);
3431 SDValue ReconstructedBitmask = DAG.getConstant(Val: 0, DL, VT: ReturnVT);
3432 // Example:
3433 // v32i16 = concat(v8i16, v8i16, v8i16, v8i16)
3434 // -> v8i1 + v8i1 + v8i1 + v8i1
3435 // -> i8 + i8 + i8 + i8
3436 // -> reconstructed i32 bitmask
3437 for (size_t I = 0; I < Concat->ops().size(); ++I) {
3438 SDValue ConcatOperand = Concat.getOperand(i: I);
3439 assert(ConcatOperand.getValueType() == ConcatOperandVT &&
3440 "concat_vectors operands must have the same type");
3441
3442 SDValue SetCCVectorOperand =
3443 extractSubVector(Vec: SetCCVector, IdxVal: I * ConcatOperandNumElts, DAG, DL, VectorWidth: 128);
3444 if (!SetCCVectorOperand ||
3445 SetCCVectorOperand.getValueType() != ConcatOperandVT)
3446 return SDValue();
3447
3448 // Build the per-chunk mask using the correct chunk type:
3449 // v16i8 -> v16i1 -> i16
3450 // v8i16 -> v8i1 -> i8
3451 // v4i32 -> v4i1 -> i4
3452 // v2i64 -> v2i1 -> i2
3453 SDValue ConcatOperandMask = DAG.getSetCC(
3454 DL, VT: ConcatOperandMaskVT, LHS: ConcatOperand, RHS: SetCCVectorOperand, Cond: SetCond);
3455 SDValue ConcatOperandBitmask =
3456 DAG.getBitcast(VT: ConcatOperandBitmaskVT, V: ConcatOperandMask);
3457 SDValue ExtendedConcatOperandBitmask =
3458 DAG.getZExtOrTrunc(Op: ConcatOperandBitmask, DL, VT: ReturnVT);
3459
3460 // Shift the previously reconstructed bits to make room for this chunk.
3461 if (I != 0) {
3462 ReconstructedBitmask = DAG.getNode(
3463 Opcode: ISD::SHL, DL, VT: ReturnVT, N1: ReconstructedBitmask,
3464 N2: DAG.getShiftAmountConstant(Val: ConcatOperandNumElts, VT: ReturnVT, DL));
3465 }
3466
3467 // Merge disjoint partial bitmasks with OR.
3468 ReconstructedBitmask =
3469 DAG.getNode(Opcode: ISD::OR, DL, VT: ReturnVT, N1: ReconstructedBitmask,
3470 N2: ExtendedConcatOperandBitmask);
3471 }
3472
3473 return ReconstructedBitmask;
3474 }
3475
3476 return SDValue();
3477}
3478
3479static SDValue performBitmaskCombine(SDNode *N, SelectionDAG &DAG) {
3480 // bitmask (setcc <X>, 0, setlt) => bitmask X
3481 assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN);
3482 using namespace llvm::SDPatternMatch;
3483
3484 if (N->getConstantOperandVal(Num: 0) != Intrinsic::wasm_bitmask)
3485 return SDValue();
3486
3487 SDValue LHS;
3488 if (!sd_match(N: N->getOperand(Num: 1), P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(),
3489 CC: m_SpecificCondCode(CC: ISD::SETLT))))
3490 return SDValue();
3491
3492 SDLoc DL(N);
3493 return DAG.getNode(
3494 Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: N->getValueType(ResNo: 0),
3495 Ops: {DAG.getConstant(Val: Intrinsic::wasm_bitmask, DL, VT: MVT::i32), LHS});
3496}
3497
3498static SDValue performAnyAllCombine(SDNode *N, SelectionDAG &DAG) {
3499 // any_true (setcc <X>, 0, eq) => (not (all_true X))
3500 // all_true (setcc <X>, 0, eq) => (not (any_true X))
3501 // any_true (setcc <X>, 0, ne) => (any_true X)
3502 // all_true (setcc <X>, 0, ne) => (all_true X)
3503 assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN);
3504 using namespace llvm::SDPatternMatch;
3505
3506 SDValue LHS;
3507 if (N->getNumOperands() < 2 ||
3508 !sd_match(N: N->getOperand(Num: 1),
3509 P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(), CC: m_CondCode())))
3510 return SDValue();
3511 EVT LT = LHS.getValueType();
3512 if (LT.getScalarSizeInBits() > 128 / LT.getVectorNumElements())
3513 return SDValue();
3514
3515 auto CombineSetCC = [&N, &DAG](Intrinsic::WASMIntrinsics InPre,
3516 ISD::CondCode SetType,
3517 Intrinsic::WASMIntrinsics InPost) {
3518 if (N->getConstantOperandVal(Num: 0) != InPre)
3519 return SDValue();
3520
3521 SDValue LHS;
3522 if (!sd_match(N: N->getOperand(Num: 1), P: m_c_SetCC(LHS: m_Value(N&: LHS), RHS: m_Zero(),
3523 CC: m_SpecificCondCode(CC: SetType))))
3524 return SDValue();
3525
3526 SDLoc DL(N);
3527 SDValue Ret = DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32,
3528 Ops: {DAG.getConstant(Val: InPost, DL, VT: MVT::i32), LHS});
3529 if (SetType == ISD::SETEQ)
3530 Ret = DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Ret,
3531 N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32));
3532 return DAG.getZExtOrTrunc(Op: Ret, DL, VT: N->getValueType(ResNo: 0));
3533 };
3534
3535 if (SDValue AnyTrueEQ = CombineSetCC(Intrinsic::wasm_anytrue, ISD::SETEQ,
3536 Intrinsic::wasm_alltrue))
3537 return AnyTrueEQ;
3538 if (SDValue AllTrueEQ = CombineSetCC(Intrinsic::wasm_alltrue, ISD::SETEQ,
3539 Intrinsic::wasm_anytrue))
3540 return AllTrueEQ;
3541 if (SDValue AnyTrueNE = CombineSetCC(Intrinsic::wasm_anytrue, ISD::SETNE,
3542 Intrinsic::wasm_anytrue))
3543 return AnyTrueNE;
3544 if (SDValue AllTrueNE = CombineSetCC(Intrinsic::wasm_alltrue, ISD::SETNE,
3545 Intrinsic::wasm_alltrue))
3546 return AllTrueNE;
3547
3548 return SDValue();
3549}
3550
3551struct MaskReduceInfo {
3552 Intrinsic::ID IID;
3553 unsigned WideCombineOpcode;
3554 bool Invert;
3555};
3556
3557static SDValue combineSmallMaskReduction(SDNode *N, EVT FromVT,
3558 unsigned NumElts,
3559 const MaskReduceInfo &Info,
3560 SelectionDAG &DAG) {
3561 EVT VecVT = FromVT.changeVectorElementType(Context&: *DAG.getContext(),
3562 EltVT: MVT::getIntegerVT(BitWidth: 128 / NumElts));
3563 assert(VecVT.getSizeInBits() == 128 &&
3564 "mask reduction should be widened to a 128-bit vector");
3565
3566 SDLoc DL(N);
3567 SDValue Mask = N->getOperand(Num: 0)->getOperand(Num: 0);
3568 SDValue Ret = DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32,
3569 Ops: {DAG.getConstant(Val: Info.IID, DL, VT: MVT::i32),
3570 DAG.getSExtOrTrunc(Op: Mask, DL, VT: VecVT)});
3571 if (Info.Invert)
3572 Ret = DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Ret,
3573 N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32));
3574 return DAG.getZExtOrTrunc(Op: Ret, DL, VT: N->getValueType(ResNo: 0));
3575}
3576
3577static SDValue combineWideMaskReduction(SDNode *N, SDValue Mask, EVT MaskVT,
3578 unsigned NumElts,
3579 const MaskReduceInfo &Info,
3580 SelectionDAG &DAG) {
3581 assert((NumElts == 32 || NumElts == 64) &&
3582 "combineWideMaskReduction is only for wide masks");
3583 assert(MaskVT.isFixedLengthVector() &&
3584 MaskVT.getVectorElementType() == MVT::i1);
3585 SDLoc DL(N);
3586 unsigned ChunkElts = 16;
3587 EVT ChunkMaskVT = EVT::getVectorVT(Context&: *DAG.getContext(), VT: MVT::i1,
3588 EC: ElementCount::getFixed(MinVal: ChunkElts));
3589 EVT LegalVecVT = ChunkMaskVT.changeVectorElementType(
3590 Context&: *DAG.getContext(), EltVT: MVT::getIntegerVT(BitWidth: 128 / ChunkElts));
3591
3592 SmallVector<SDValue, 4> ChunkResults;
3593 // Split the wide mask into v16i1 chunks and reduce each chunk separately.
3594 // For example:
3595 // v32i1: [0..15] [16..31]
3596 // | |
3597 // v v
3598 // chunk0 chunk1
3599 //
3600 // v64i1: [0..15] [16..31] [32..47] [48..63]
3601 // | | | |
3602 // v v v v
3603 // chunk0 chunk1 chunk2 chunk3
3604 //
3605 // each chunk:
3606 // v16i1 -> v16i8 -> wasm_anytrue/alltrue -> i32 0/1
3607 for (unsigned I = 0; I < NumElts; I += ChunkElts) {
3608 SDValue ChunkMask = DAG.getNode(Opcode: ISD::EXTRACT_SUBVECTOR, DL, VT: ChunkMaskVT,
3609 N1: Mask, N2: DAG.getVectorIdxConstant(Val: I, DL));
3610 SDValue LegalMask = DAG.getSExtOrTrunc(Op: ChunkMask, DL, VT: LegalVecVT);
3611 SDValue Reduced =
3612 DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32,
3613 N1: DAG.getConstant(Val: Info.IID, DL, VT: MVT::i32), N2: LegalMask);
3614 ChunkResults.push_back(Elt: Reduced);
3615 }
3616
3617 SDValue Acc = ChunkResults[0];
3618 for (unsigned I = 1; I < ChunkResults.size(); ++I)
3619 Acc =
3620 DAG.getNode(Opcode: Info.WideCombineOpcode, DL, VT: MVT::i32, N1: Acc, N2: ChunkResults[I]);
3621
3622 if (Info.Invert)
3623 Acc = DAG.getNode(Opcode: ISD::XOR, DL, VT: MVT::i32, N1: Acc,
3624 N2: DAG.getConstant(Val: 1, DL, VT: MVT::i32));
3625
3626 return DAG.getZExtOrTrunc(Op: Acc, DL, VT: N->getValueType(ResNo: 0));
3627}
3628
3629static std::optional<MaskReduceInfo> classifyMaskReduction(SDNode *N) {
3630 auto *C = dyn_cast<ConstantSDNode>(Val: N->getOperand(Num: 1));
3631 if (!C)
3632 return std::nullopt;
3633
3634 ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get();
3635
3636 // setcc (bitcast mask), 0, ne -> any_true(mask)
3637 if (C->isZero() && CC == ISD::SETNE)
3638 return MaskReduceInfo{.IID: Intrinsic::wasm_anytrue, .WideCombineOpcode: ISD::OR, .Invert: false};
3639
3640 // setcc (bitcast mask), 0, eq -> !any_true(mask)
3641 if (C->isZero() && CC == ISD::SETEQ)
3642 return MaskReduceInfo{.IID: Intrinsic::wasm_anytrue, .WideCombineOpcode: ISD::OR, .Invert: true};
3643
3644 // setcc (bitcast mask), -1, eq -> all_true(mask)
3645 if (C->isAllOnes() && CC == ISD::SETEQ)
3646 return MaskReduceInfo{.IID: Intrinsic::wasm_alltrue, .WideCombineOpcode: ISD::AND, .Invert: false};
3647
3648 // setcc (bitcast mask), -1, ne -> !all_true(mask)
3649 if (C->isAllOnes() && CC == ISD::SETNE)
3650 return MaskReduceInfo{.IID: Intrinsic::wasm_alltrue, .WideCombineOpcode: ISD::AND, .Invert: true};
3651
3652 return std::nullopt;
3653}
3654
3655/// Try to convert a i128 comparison to a v16i8 comparison before type
3656/// legalization splits it up into chunks
3657static SDValue
3658combineVectorSizedSetCCEquality(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
3659 const WebAssemblySubtarget *Subtarget) {
3660
3661 SDLoc DL(N);
3662 SDValue X = N->getOperand(Num: 0);
3663 SDValue Y = N->getOperand(Num: 1);
3664 EVT VT = N->getValueType(ResNo: 0);
3665 EVT OpVT = X.getValueType();
3666
3667 SelectionDAG &DAG = DCI.DAG;
3668 if (DCI.DAG.getMachineFunction().getFunction().hasFnAttribute(
3669 Kind: Attribute::NoImplicitFloat))
3670 return SDValue();
3671
3672 ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 2))->get();
3673 // We're looking for an oversized integer equality comparison with SIMD
3674 if (!OpVT.isScalarInteger() || !OpVT.isByteSized() || OpVT != MVT::i128 ||
3675 !Subtarget->hasSIMD128() || !isIntEqualitySetCC(Code: CC))
3676 return SDValue();
3677
3678 // Don't perform this combine if constructing the vector will be expensive.
3679 auto IsVectorBitCastCheap = [](SDValue X) {
3680 X = peekThroughBitcasts(V: X);
3681 return isa<ConstantSDNode>(Val: X) || X.getOpcode() == ISD::LOAD;
3682 };
3683
3684 if (!IsVectorBitCastCheap(X) || !IsVectorBitCastCheap(Y))
3685 return SDValue();
3686
3687 SDValue VecX = DAG.getBitcast(VT: MVT::v16i8, V: X);
3688 SDValue VecY = DAG.getBitcast(VT: MVT::v16i8, V: Y);
3689 SDValue Cmp = DAG.getSetCC(DL, VT: MVT::v16i8, LHS: VecX, RHS: VecY, Cond: CC);
3690
3691 SDValue Intr =
3692 DAG.getNode(Opcode: ISD::INTRINSIC_WO_CHAIN, DL, VT: MVT::i32,
3693 Ops: {DAG.getConstant(Val: CC == ISD::SETEQ ? Intrinsic::wasm_alltrue
3694 : Intrinsic::wasm_anytrue,
3695 DL, VT: MVT::i32),
3696 Cmp});
3697
3698 return DAG.getSetCC(DL, VT, LHS: Intr, RHS: DAG.getConstant(Val: 0, DL, VT: MVT::i32),
3699 Cond: ISD::SETNE);
3700}
3701
3702static SDValue performSETCCCombine(SDNode *N,
3703 TargetLowering::DAGCombinerInfo &DCI,
3704 const WebAssemblySubtarget *Subtarget) {
3705 if (!DCI.isBeforeLegalize())
3706 return SDValue();
3707
3708 EVT VT = N->getValueType(ResNo: 0);
3709 if (!VT.isScalarInteger())
3710 return SDValue();
3711
3712 if (SDValue V = combineVectorSizedSetCCEquality(N, DCI, Subtarget))
3713 return V;
3714
3715 SDValue LHS = N->getOperand(Num: 0);
3716 if (LHS->getOpcode() != ISD::BITCAST)
3717 return SDValue();
3718
3719 EVT FromVT = LHS->getOperand(Num: 0).getValueType();
3720 if (!FromVT.isFixedLengthVectorOf(EltVT: MVT::i1))
3721 return SDValue();
3722
3723 unsigned NumElts = FromVT.getVectorNumElements();
3724 auto Info = classifyMaskReduction(N);
3725 if (!Info)
3726 return SDValue();
3727
3728 auto &DAG = DCI.DAG;
3729 if (NumElts == 2 || NumElts == 4 || NumElts == 8 || NumElts == 16)
3730 return combineSmallMaskReduction(N, FromVT, NumElts, Info: *Info, DAG);
3731
3732 if (NumElts == 32 || NumElts == 64)
3733 return combineWideMaskReduction(N, Mask: LHS.getOperand(i: 0), MaskVT: FromVT, NumElts,
3734 Info: *Info, DAG);
3735
3736 return SDValue();
3737}
3738
3739static SDValue TryWideExtMulCombine(SDNode *N, SelectionDAG &DAG) {
3740 EVT VT = N->getValueType(ResNo: 0);
3741 if (VT != MVT::v8i32 && VT != MVT::v16i32)
3742 return SDValue();
3743
3744 // Mul with extending inputs.
3745 SDValue LHS = N->getOperand(Num: 0);
3746 SDValue RHS = N->getOperand(Num: 1);
3747 if (LHS.getOpcode() != RHS.getOpcode())
3748 return SDValue();
3749
3750 if (LHS.getOpcode() != ISD::SIGN_EXTEND &&
3751 LHS.getOpcode() != ISD::ZERO_EXTEND)
3752 return SDValue();
3753
3754 if (LHS->getOperand(Num: 0).getValueType() != RHS->getOperand(Num: 0).getValueType())
3755 return SDValue();
3756
3757 EVT FromVT = LHS->getOperand(Num: 0).getValueType();
3758 EVT EltTy = FromVT.getVectorElementType();
3759 if (EltTy != MVT::i8)
3760 return SDValue();
3761
3762 // For an input DAG that looks like this
3763 // %a = input_type
3764 // %b = input_type
3765 // %lhs = extend %a to output_type
3766 // %rhs = extend %b to output_type
3767 // %mul = mul %lhs, %rhs
3768
3769 // input_type | output_type | instructions
3770 // v16i8 | v16i32 | %low = i16x8.extmul_low_i8x16_ %a, %b
3771 // | | %high = i16x8.extmul_high_i8x16_, %a, %b
3772 // | | %low_low = i32x4.ext_low_i16x8_ %low
3773 // | | %low_high = i32x4.ext_high_i16x8_ %low
3774 // | | %high_low = i32x4.ext_low_i16x8_ %high
3775 // | | %high_high = i32x4.ext_high_i16x8_ %high
3776 // | | %res = concat_vector(...)
3777 // v8i8 | v8i32 | %low = i16x8.extmul_low_i8x16_ %a, %b
3778 // | | %low_low = i32x4.ext_low_i16x8_ %low
3779 // | | %low_high = i32x4.ext_high_i16x8_ %low
3780 // | | %res = concat_vector(%low_low, %low_high)
3781
3782 SDLoc DL(N);
3783 unsigned NumElts = VT.getVectorNumElements();
3784 SDValue ExtendInLHS = LHS->getOperand(Num: 0);
3785 SDValue ExtendInRHS = RHS->getOperand(Num: 0);
3786 bool IsSigned = LHS->getOpcode() == ISD::SIGN_EXTEND;
3787 unsigned ExtendLowOpc =
3788 IsSigned ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U;
3789 unsigned ExtendHighOpc =
3790 IsSigned ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U;
3791
3792 auto GetExtendLow = [&DAG, &DL, &ExtendLowOpc](EVT VT, SDValue Op) {
3793 return DAG.getNode(Opcode: ExtendLowOpc, DL, VT, Operand: Op);
3794 };
3795 auto GetExtendHigh = [&DAG, &DL, &ExtendHighOpc](EVT VT, SDValue Op) {
3796 return DAG.getNode(Opcode: ExtendHighOpc, DL, VT, Operand: Op);
3797 };
3798
3799 if (NumElts == 16) {
3800 SDValue LowLHS = GetExtendLow(MVT::v8i16, ExtendInLHS);
3801 SDValue LowRHS = GetExtendLow(MVT::v8i16, ExtendInRHS);
3802 SDValue MulLow = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: LowLHS, N2: LowRHS);
3803 SDValue HighLHS = GetExtendHigh(MVT::v8i16, ExtendInLHS);
3804 SDValue HighRHS = GetExtendHigh(MVT::v8i16, ExtendInRHS);
3805 SDValue MulHigh = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: HighLHS, N2: HighRHS);
3806 SDValue SubVectors[] = {
3807 GetExtendLow(MVT::v4i32, MulLow),
3808 GetExtendHigh(MVT::v4i32, MulLow),
3809 GetExtendLow(MVT::v4i32, MulHigh),
3810 GetExtendHigh(MVT::v4i32, MulHigh),
3811 };
3812 return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, Ops: SubVectors);
3813 } else {
3814 assert(NumElts == 8);
3815 SDValue LowLHS = DAG.getNode(Opcode: LHS->getOpcode(), DL, VT: MVT::v8i16, Operand: ExtendInLHS);
3816 SDValue LowRHS = DAG.getNode(Opcode: RHS->getOpcode(), DL, VT: MVT::v8i16, Operand: ExtendInRHS);
3817 SDValue MulLow = DAG.getNode(Opcode: ISD::MUL, DL, VT: MVT::v8i16, N1: LowLHS, N2: LowRHS);
3818 SDValue Lo = GetExtendLow(MVT::v4i32, MulLow);
3819 SDValue Hi = GetExtendHigh(MVT::v4i32, MulLow);
3820 return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, N1: Lo, N2: Hi);
3821 }
3822 return SDValue();
3823}
3824
3825static SDValue performMulCombine(SDNode *N,
3826 TargetLowering::DAGCombinerInfo &DCI) {
3827 assert(N->getOpcode() == ISD::MUL);
3828 EVT VT = N->getValueType(ResNo: 0);
3829 if (!VT.isVector())
3830 return SDValue();
3831
3832 if (auto Res = TryWideExtMulCombine(N, DAG&: DCI.DAG))
3833 return Res;
3834
3835 // We don't natively support v16i8 or v8i8 mul, but we do support v8i16. So,
3836 // extend them to v8i16.
3837 if (VT != MVT::v8i8 && VT != MVT::v16i8)
3838 return SDValue();
3839
3840 SDLoc DL(N);
3841 SelectionDAG &DAG = DCI.DAG;
3842 SDValue LHS = N->getOperand(Num: 0);
3843 SDValue RHS = N->getOperand(Num: 1);
3844 EVT MulVT = MVT::v8i16;
3845
3846 if (VT == MVT::v8i8) {
3847 SDValue PromotedLHS = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: MVT::v16i8, N1: LHS,
3848 N2: DAG.getUNDEF(VT: MVT::v8i8));
3849 SDValue PromotedRHS = DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: MVT::v16i8, N1: RHS,
3850 N2: DAG.getUNDEF(VT: MVT::v8i8));
3851 SDValue LowLHS =
3852 DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: PromotedLHS);
3853 SDValue LowRHS =
3854 DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: PromotedRHS);
3855 SDValue MulLow = DAG.getBitcast(
3856 VT: MVT::v16i8, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: LowLHS, N2: LowRHS));
3857 // Take the low byte of each lane.
3858 SDValue Shuffle = DAG.getVectorShuffle(
3859 VT: MVT::v16i8, dl: DL, N1: MulLow, N2: DAG.getUNDEF(VT: MVT::v16i8),
3860 Mask: {0, 2, 4, 6, 8, 10, 12, 14, -1, -1, -1, -1, -1, -1, -1, -1});
3861 return extractSubVector(Vec: Shuffle, IdxVal: 0, DAG, DL, VectorWidth: 64);
3862 } else {
3863 assert(VT == MVT::v16i8 && "Expected v16i8");
3864 SDValue LowLHS = DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: LHS);
3865 SDValue LowRHS = DAG.getNode(Opcode: WebAssemblyISD::EXTEND_LOW_U, DL, VT: MulVT, Operand: RHS);
3866 SDValue HighLHS =
3867 DAG.getNode(Opcode: WebAssemblyISD::EXTEND_HIGH_U, DL, VT: MulVT, Operand: LHS);
3868 SDValue HighRHS =
3869 DAG.getNode(Opcode: WebAssemblyISD::EXTEND_HIGH_U, DL, VT: MulVT, Operand: RHS);
3870
3871 SDValue MulLow =
3872 DAG.getBitcast(VT, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: LowLHS, N2: LowRHS));
3873 SDValue MulHigh =
3874 DAG.getBitcast(VT, V: DAG.getNode(Opcode: ISD::MUL, DL, VT: MulVT, N1: HighLHS, N2: HighRHS));
3875
3876 // Take the low byte of each lane.
3877 return DAG.getVectorShuffle(
3878 VT, dl: DL, N1: MulLow, N2: MulHigh,
3879 Mask: {0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30});
3880 }
3881}
3882
3883SDValue DoubleVectorWidth(SDValue In, unsigned RequiredNumElems,
3884 SelectionDAG &DAG) {
3885 SDLoc DL(In);
3886 LLVMContext &Ctx = *DAG.getContext();
3887 EVT InVT = In.getValueType();
3888 unsigned NumElems = InVT.getVectorNumElements() * 2;
3889 EVT OutVT = EVT::getVectorVT(Context&: Ctx, VT: InVT.getVectorElementType(), NumElements: NumElems);
3890 SDValue Concat =
3891 DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT: OutVT, N1: In, N2: DAG.getPOISON(VT: InVT));
3892 if (NumElems < RequiredNumElems) {
3893 return DoubleVectorWidth(In: Concat, RequiredNumElems, DAG);
3894 }
3895 return Concat;
3896}
3897
3898SDValue performConvertFPCombine(SDNode *N, SelectionDAG &DAG) {
3899 EVT OutVT = N->getValueType(ResNo: 0);
3900 if (!OutVT.isVector())
3901 return SDValue();
3902
3903 EVT OutElTy = OutVT.getVectorElementType();
3904 if (OutElTy != MVT::i8 && OutElTy != MVT::i16)
3905 return SDValue();
3906
3907 unsigned NumElems = OutVT.getVectorNumElements();
3908 if (!isPowerOf2_32(Value: NumElems))
3909 return SDValue();
3910
3911 EVT FPVT = N->getOperand(Num: 0)->getValueType(ResNo: 0);
3912 if (FPVT.getVectorElementType() != MVT::f32)
3913 return SDValue();
3914
3915 SDLoc DL(N);
3916
3917 // First, convert to i32.
3918 LLVMContext &Ctx = *DAG.getContext();
3919 EVT IntVT = EVT::getVectorVT(Context&: Ctx, VT: MVT::i32, NumElements: NumElems);
3920 SDValue ToInt = DAG.getNode(Opcode: N->getOpcode(), DL, VT: IntVT, Operand: N->getOperand(Num: 0));
3921 APInt Mask = APInt::getLowBitsSet(numBits: IntVT.getScalarSizeInBits(),
3922 loBitsSet: OutVT.getScalarSizeInBits());
3923 // Mask out the top MSBs.
3924 SDValue Masked =
3925 DAG.getNode(Opcode: ISD::AND, DL, VT: IntVT, N1: ToInt, N2: DAG.getConstant(Val: Mask, DL, VT: IntVT));
3926
3927 if (OutVT.getSizeInBits() < 128) {
3928 // Create a wide enough vector that we can use narrow.
3929 EVT NarrowedVT = OutElTy == MVT::i8 ? MVT::v16i8 : MVT::v8i16;
3930 unsigned NumRequiredElems = NarrowedVT.getVectorNumElements();
3931 SDValue WideVector = DoubleVectorWidth(In: Masked, RequiredNumElems: NumRequiredElems, DAG);
3932 SDValue Trunc = truncateVectorWithNARROW(DstVT: NarrowedVT, In: WideVector, DL, DAG);
3933 return DAG.getBitcast(
3934 VT: OutVT, V: extractSubVector(Vec: Trunc, IdxVal: 0, DAG, DL, VectorWidth: OutVT.getSizeInBits()));
3935 } else {
3936 return truncateVectorWithNARROW(DstVT: OutVT, In: Masked, DL, DAG);
3937 }
3938 return SDValue();
3939}
3940
3941// Wide vector shift operations such as v8i32 with sign-extended
3942// operands cause Type Legalizer crashes because the target-specific
3943// extension nodes cannot be directly mapped to the 256-bit size.
3944//
3945// To resolve the crash and optimize performance, we intercept the
3946// illegal v8i32 shift in DAGCombine. We convert the shift amounts
3947// into multipliers and manually split the vector into two v4i32 halves.
3948//
3949// Before: t1: v8i32 = shl (sign_extend v8i16), const_vec
3950// After : t2: v4i32 = mul (ext_low_s v8i16), (ext_low_s narrow_vec)
3951// t3: v4i32 = mul (ext_high_s v8i16), (ext_high_s narrow_vec)
3952// t4: v8i32 = concat_vectors t2, t3
3953static SDValue performShiftCombine(SDNode *N,
3954 TargetLowering::DAGCombinerInfo &DCI) {
3955 SelectionDAG &DAG = DCI.DAG;
3956 assert(N->getOpcode() == ISD::SHL);
3957 EVT VT = N->getValueType(ResNo: 0);
3958 if (VT != MVT::v8i32)
3959 return SDValue();
3960
3961 SDValue LHS = N->getOperand(Num: 0);
3962 SDValue RHS = N->getOperand(Num: 1);
3963 unsigned ExtOpc = LHS.getOpcode();
3964 if (ExtOpc != ISD::SIGN_EXTEND && ExtOpc != ISD::ZERO_EXTEND)
3965 return SDValue();
3966
3967 if (RHS.getOpcode() != ISD::BUILD_VECTOR)
3968 return SDValue();
3969
3970 SDLoc DL(N);
3971 SDValue ExtendIn = LHS.getOperand(i: 0);
3972 EVT FromVT = ExtendIn.getValueType();
3973 if (FromVT != MVT::v8i16)
3974 return SDValue();
3975
3976 unsigned NumElts = VT.getVectorNumElements();
3977 unsigned BitWidth = FromVT.getScalarSizeInBits();
3978 bool IsSigned = (ExtOpc == ISD::SIGN_EXTEND);
3979 unsigned MaxValidShift = IsSigned ? (BitWidth - 1) : BitWidth;
3980 SmallVector<SDValue, 16> MulConsts;
3981 for (unsigned I = 0; I < NumElts; ++I) {
3982 auto *C = dyn_cast<ConstantSDNode>(Val: RHS.getOperand(i: I));
3983 if (!C)
3984 return SDValue();
3985
3986 const APInt &ShiftAmt = C->getAPIntValue();
3987 if (ShiftAmt.uge(RHS: MaxValidShift))
3988 return SDValue();
3989
3990 APInt MulAmt = APInt::getOneBitSet(numBits: BitWidth, BitNo: ShiftAmt.getZExtValue());
3991 MulConsts.push_back(Elt: DAG.getConstant(Val: MulAmt, DL, VT: FromVT.getScalarType(),
3992 /*isTarget=*/false, /*isOpaque=*/true));
3993 }
3994
3995 SDValue NarrowConst = DAG.getBuildVector(VT: FromVT, DL, Ops: MulConsts);
3996 unsigned ExtLowOpc =
3997 IsSigned ? WebAssemblyISD::EXTEND_LOW_S : WebAssemblyISD::EXTEND_LOW_U;
3998 unsigned ExtHighOpc =
3999 IsSigned ? WebAssemblyISD::EXTEND_HIGH_S : WebAssemblyISD::EXTEND_HIGH_U;
4000
4001 EVT HalfVT = MVT::v4i32;
4002 SDValue LHSLo = DAG.getNode(Opcode: ExtLowOpc, DL, VT: HalfVT, Operand: ExtendIn);
4003 SDValue LHSHi = DAG.getNode(Opcode: ExtHighOpc, DL, VT: HalfVT, Operand: ExtendIn);
4004 SDValue RHSLo = DAG.getNode(Opcode: ExtLowOpc, DL, VT: HalfVT, Operand: NarrowConst);
4005 SDValue RHSHi = DAG.getNode(Opcode: ExtHighOpc, DL, VT: HalfVT, Operand: NarrowConst);
4006 SDValue MulLo = DAG.getNode(Opcode: ISD::MUL, DL, VT: HalfVT, N1: LHSLo, N2: RHSLo);
4007 SDValue MulHi = DAG.getNode(Opcode: ISD::MUL, DL, VT: HalfVT, N1: LHSHi, N2: RHSHi);
4008 return DAG.getNode(Opcode: ISD::CONCAT_VECTORS, DL, VT, N1: MulLo, N2: MulHi);
4009}
4010
4011static SDValue performMinMaxF128Combine(SDNode *N, SelectionDAG &DAG) {
4012 if (N->getValueType(ResNo: 0) != MVT::f128)
4013 return SDValue();
4014
4015 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
4016 switch (N->getOpcode()) {
4017 // wasi-libc and emscripten do not currently define fminimuml and fmaximuml.
4018 case ISD::FMINIMUM:
4019 case ISD::FMAXIMUM:
4020 return TLI.expandFMINIMUM_FMAXIMUM(N, DAG);
4021
4022 // wasi-libc and emscripten do not currently define fminimum_numl and
4023 // fmaximum_numl.
4024 case ISD::FMINIMUMNUM:
4025 case ISD::FMAXIMUMNUM:
4026 return TLI.expandFMINIMUMNUM_FMAXIMUMNUM(N, DAG);
4027
4028 default:
4029 return SDValue();
4030 }
4031}
4032
4033SDValue
4034WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N,
4035 DAGCombinerInfo &DCI) const {
4036 switch (N->getOpcode()) {
4037 default:
4038 return SDValue();
4039 case ISD::BITCAST:
4040 return performBitcastCombine(N, DCI);
4041 case ISD::SETCC:
4042 return performSETCCCombine(N, DCI, Subtarget);
4043 case ISD::VECTOR_SHUFFLE:
4044 return performVECTOR_SHUFFLECombine(N, DCI);
4045 case ISD::SIGN_EXTEND:
4046 case ISD::ZERO_EXTEND:
4047 return performVectorExtendCombine(N, DCI);
4048 case ISD::UINT_TO_FP:
4049 if (auto ExtCombine = performVectorExtendToFPCombine(N, DCI, Subtarget))
4050 return ExtCombine;
4051 return performVectorNonNegToFPCombine(N, DCI);
4052 case ISD::SINT_TO_FP:
4053 return performVectorExtendToFPCombine(N, DCI, Subtarget);
4054 case ISD::FP_TO_SINT_SAT:
4055 case ISD::FP_TO_UINT_SAT:
4056 case ISD::FP_ROUND:
4057 case ISD::CONCAT_VECTORS:
4058 return performVectorTruncZeroCombine(N, DCI);
4059 case ISD::FP_TO_SINT:
4060 case ISD::FP_TO_UINT:
4061 return performConvertFPCombine(N, DAG&: DCI.DAG);
4062 case ISD::TRUNCATE:
4063 return performTruncateCombine(N, DCI);
4064 case ISD::INTRINSIC_WO_CHAIN: {
4065 if (SDValue V = performBitmaskCombine(N, DAG&: DCI.DAG))
4066 return V;
4067 return performAnyAllCombine(N, DAG&: DCI.DAG);
4068 }
4069 case ISD::MUL:
4070 return performMulCombine(N, DCI);
4071 case ISD::SHL:
4072 return performShiftCombine(N, DCI);
4073 case ISD::FMINIMUM:
4074 case ISD::FMAXIMUM:
4075 case ISD::FMINIMUMNUM:
4076 case ISD::FMAXIMUMNUM:
4077 return performMinMaxF128Combine(N, DAG&: DCI.DAG);
4078 }
4079}
4080