xref: /freebsd/contrib/llvm-project/llvm/lib/Target/WebAssembly/WebAssemblyISelLowering.cpp (revision 0fca6ea1d4eea4c934cfff25ac9ee8ad6fe95583)
1 //=- WebAssemblyISelLowering.cpp - WebAssembly DAG Lowering Implementation -==//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 ///
9 /// \file
10 /// This file implements the WebAssemblyTargetLowering class.
11 ///
12 //===----------------------------------------------------------------------===//
13 
14 #include "WebAssemblyISelLowering.h"
15 #include "MCTargetDesc/WebAssemblyMCTargetDesc.h"
16 #include "Utils/WebAssemblyTypeUtilities.h"
17 #include "WebAssemblyMachineFunctionInfo.h"
18 #include "WebAssemblySubtarget.h"
19 #include "WebAssemblyTargetMachine.h"
20 #include "WebAssemblyUtilities.h"
21 #include "llvm/CodeGen/CallingConvLower.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineFunctionPass.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineJumpTableInfo.h"
26 #include "llvm/CodeGen/MachineModuleInfo.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/SelectionDAGNodes.h"
30 #include "llvm/IR/DiagnosticInfo.h"
31 #include "llvm/IR/DiagnosticPrinter.h"
32 #include "llvm/IR/Function.h"
33 #include "llvm/IR/Intrinsics.h"
34 #include "llvm/IR/IntrinsicsWebAssembly.h"
35 #include "llvm/IR/PatternMatch.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/Support/ErrorHandling.h"
38 #include "llvm/Support/KnownBits.h"
39 #include "llvm/Support/MathExtras.h"
40 #include "llvm/Support/raw_ostream.h"
41 #include "llvm/Target/TargetOptions.h"
42 using namespace llvm;
43 
44 #define DEBUG_TYPE "wasm-lower"
45 
WebAssemblyTargetLowering(const TargetMachine & TM,const WebAssemblySubtarget & STI)46 WebAssemblyTargetLowering::WebAssemblyTargetLowering(
47     const TargetMachine &TM, const WebAssemblySubtarget &STI)
48     : TargetLowering(TM), Subtarget(&STI) {
49   auto MVTPtr = Subtarget->hasAddr64() ? MVT::i64 : MVT::i32;
50 
51   // Booleans always contain 0 or 1.
52   setBooleanContents(ZeroOrOneBooleanContent);
53   // Except in SIMD vectors
54   setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
55   // We don't know the microarchitecture here, so just reduce register pressure.
56   setSchedulingPreference(Sched::RegPressure);
57   // Tell ISel that we have a stack pointer.
58   setStackPointerRegisterToSaveRestore(
59       Subtarget->hasAddr64() ? WebAssembly::SP64 : WebAssembly::SP32);
60   // Set up the register classes.
61   addRegisterClass(MVT::i32, &WebAssembly::I32RegClass);
62   addRegisterClass(MVT::i64, &WebAssembly::I64RegClass);
63   addRegisterClass(MVT::f32, &WebAssembly::F32RegClass);
64   addRegisterClass(MVT::f64, &WebAssembly::F64RegClass);
65   if (Subtarget->hasSIMD128()) {
66     addRegisterClass(MVT::v16i8, &WebAssembly::V128RegClass);
67     addRegisterClass(MVT::v8i16, &WebAssembly::V128RegClass);
68     addRegisterClass(MVT::v4i32, &WebAssembly::V128RegClass);
69     addRegisterClass(MVT::v4f32, &WebAssembly::V128RegClass);
70     addRegisterClass(MVT::v2i64, &WebAssembly::V128RegClass);
71     addRegisterClass(MVT::v2f64, &WebAssembly::V128RegClass);
72   }
73   if (Subtarget->hasHalfPrecision()) {
74     addRegisterClass(MVT::v8f16, &WebAssembly::V128RegClass);
75   }
76   if (Subtarget->hasReferenceTypes()) {
77     addRegisterClass(MVT::externref, &WebAssembly::EXTERNREFRegClass);
78     addRegisterClass(MVT::funcref, &WebAssembly::FUNCREFRegClass);
79     if (Subtarget->hasExceptionHandling()) {
80       addRegisterClass(MVT::exnref, &WebAssembly::EXNREFRegClass);
81     }
82   }
83   // Compute derived properties from the register classes.
84   computeRegisterProperties(Subtarget->getRegisterInfo());
85 
86   // Transform loads and stores to pointers in address space 1 to loads and
87   // stores to WebAssembly global variables, outside linear memory.
88   for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64}) {
89     setOperationAction(ISD::LOAD, T, Custom);
90     setOperationAction(ISD::STORE, T, Custom);
91   }
92   if (Subtarget->hasSIMD128()) {
93     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
94                    MVT::v2f64}) {
95       setOperationAction(ISD::LOAD, T, Custom);
96       setOperationAction(ISD::STORE, T, Custom);
97     }
98   }
99   if (Subtarget->hasReferenceTypes()) {
100     // We need custom load and store lowering for both externref, funcref and
101     // Other. The MVT::Other here represents tables of reference types.
102     for (auto T : {MVT::externref, MVT::funcref, MVT::Other}) {
103       setOperationAction(ISD::LOAD, T, Custom);
104       setOperationAction(ISD::STORE, T, Custom);
105     }
106   }
107 
108   setOperationAction(ISD::GlobalAddress, MVTPtr, Custom);
109   setOperationAction(ISD::GlobalTLSAddress, MVTPtr, Custom);
110   setOperationAction(ISD::ExternalSymbol, MVTPtr, Custom);
111   setOperationAction(ISD::JumpTable, MVTPtr, Custom);
112   setOperationAction(ISD::BlockAddress, MVTPtr, Custom);
113   setOperationAction(ISD::BRIND, MVT::Other, Custom);
114   setOperationAction(ISD::CLEAR_CACHE, MVT::Other, Custom);
115 
116   // Take the default expansion for va_arg, va_copy, and va_end. There is no
117   // default action for va_start, so we do that custom.
118   setOperationAction(ISD::VASTART, MVT::Other, Custom);
119   setOperationAction(ISD::VAARG, MVT::Other, Expand);
120   setOperationAction(ISD::VACOPY, MVT::Other, Expand);
121   setOperationAction(ISD::VAEND, MVT::Other, Expand);
122 
123   for (auto T : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
124     // Don't expand the floating-point types to constant pools.
125     setOperationAction(ISD::ConstantFP, T, Legal);
126     // Expand floating-point comparisons.
127     for (auto CC : {ISD::SETO, ISD::SETUO, ISD::SETUEQ, ISD::SETONE,
128                     ISD::SETULT, ISD::SETULE, ISD::SETUGT, ISD::SETUGE})
129       setCondCodeAction(CC, T, Expand);
130     // Expand floating-point library function operators.
131     for (auto Op :
132          {ISD::FSIN, ISD::FCOS, ISD::FSINCOS, ISD::FPOW, ISD::FREM, ISD::FMA})
133       setOperationAction(Op, T, Expand);
134     // Note supported floating-point library function operators that otherwise
135     // default to expand.
136     for (auto Op : {ISD::FCEIL, ISD::FFLOOR, ISD::FTRUNC, ISD::FNEARBYINT,
137                     ISD::FRINT, ISD::FROUNDEVEN})
138       setOperationAction(Op, T, Legal);
139     // Support minimum and maximum, which otherwise default to expand.
140     setOperationAction(ISD::FMINIMUM, T, Legal);
141     setOperationAction(ISD::FMAXIMUM, T, Legal);
142     // WebAssembly currently has no builtin f16 support.
143     setOperationAction(ISD::FP16_TO_FP, T, Expand);
144     setOperationAction(ISD::FP_TO_FP16, T, Expand);
145     setLoadExtAction(ISD::EXTLOAD, T, MVT::f16, Expand);
146     setTruncStoreAction(T, MVT::f16, Expand);
147   }
148 
149   if (Subtarget->hasHalfPrecision()) {
150     setOperationAction(ISD::FMINIMUM, MVT::v8f16, Legal);
151     setOperationAction(ISD::FMAXIMUM, MVT::v8f16, Legal);
152   }
153 
154   // Expand unavailable integer operations.
155   for (auto Op :
156        {ISD::BSWAP, ISD::SMUL_LOHI, ISD::UMUL_LOHI, ISD::MULHS, ISD::MULHU,
157         ISD::SDIVREM, ISD::UDIVREM, ISD::SHL_PARTS, ISD::SRA_PARTS,
158         ISD::SRL_PARTS, ISD::ADDC, ISD::ADDE, ISD::SUBC, ISD::SUBE}) {
159     for (auto T : {MVT::i32, MVT::i64})
160       setOperationAction(Op, T, Expand);
161     if (Subtarget->hasSIMD128())
162       for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
163         setOperationAction(Op, T, Expand);
164   }
165 
166   if (Subtarget->hasNontrappingFPToInt())
167     for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT})
168       for (auto T : {MVT::i32, MVT::i64})
169         setOperationAction(Op, T, Custom);
170 
171   // SIMD-specific configuration
172   if (Subtarget->hasSIMD128()) {
173     // Combine vector mask reductions into alltrue/anytrue
174     setTargetDAGCombine(ISD::SETCC);
175 
176     // Convert vector to integer bitcasts to bitmask
177     setTargetDAGCombine(ISD::BITCAST);
178 
179     // Hoist bitcasts out of shuffles
180     setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
181 
182     // Combine extends of extract_subvectors into widening ops
183     setTargetDAGCombine({ISD::SIGN_EXTEND, ISD::ZERO_EXTEND});
184 
185     // Combine int_to_fp or fp_extend of extract_vectors and vice versa into
186     // conversions ops
187     setTargetDAGCombine({ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_EXTEND,
188                          ISD::EXTRACT_SUBVECTOR});
189 
190     // Combine fp_to_{s,u}int_sat or fp_round of concat_vectors or vice versa
191     // into conversion ops
192     setTargetDAGCombine({ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT,
193                          ISD::FP_ROUND, ISD::CONCAT_VECTORS});
194 
195     setTargetDAGCombine(ISD::TRUNCATE);
196 
197     // Support saturating add for i8x16 and i16x8
198     for (auto Op : {ISD::SADDSAT, ISD::UADDSAT})
199       for (auto T : {MVT::v16i8, MVT::v8i16})
200         setOperationAction(Op, T, Legal);
201 
202     // Support integer abs
203     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
204       setOperationAction(ISD::ABS, T, Legal);
205 
206     // Custom lower BUILD_VECTORs to minimize number of replace_lanes
207     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
208                    MVT::v2f64})
209       setOperationAction(ISD::BUILD_VECTOR, T, Custom);
210 
211     // We have custom shuffle lowering to expose the shuffle mask
212     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
213                    MVT::v2f64})
214       setOperationAction(ISD::VECTOR_SHUFFLE, T, Custom);
215 
216     // Support splatting
217     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
218                    MVT::v2f64})
219       setOperationAction(ISD::SPLAT_VECTOR, T, Legal);
220 
221     // Custom lowering since wasm shifts must have a scalar shift amount
222     for (auto Op : {ISD::SHL, ISD::SRA, ISD::SRL})
223       for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
224         setOperationAction(Op, T, Custom);
225 
226     // Custom lower lane accesses to expand out variable indices
227     for (auto Op : {ISD::EXTRACT_VECTOR_ELT, ISD::INSERT_VECTOR_ELT})
228       for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
229                      MVT::v2f64})
230         setOperationAction(Op, T, Custom);
231 
232     // There is no i8x16.mul instruction
233     setOperationAction(ISD::MUL, MVT::v16i8, Expand);
234 
235     // There is no vector conditional select instruction
236     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v4f32, MVT::v2i64,
237                    MVT::v2f64})
238       setOperationAction(ISD::SELECT_CC, T, Expand);
239 
240     // Expand integer operations supported for scalars but not SIMD
241     for (auto Op :
242          {ISD::SDIV, ISD::UDIV, ISD::SREM, ISD::UREM, ISD::ROTL, ISD::ROTR})
243       for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64})
244         setOperationAction(Op, T, Expand);
245 
246     // But we do have integer min and max operations
247     for (auto Op : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
248       for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
249         setOperationAction(Op, T, Legal);
250 
251     // And we have popcnt for i8x16. It can be used to expand ctlz/cttz.
252     setOperationAction(ISD::CTPOP, MVT::v16i8, Legal);
253     setOperationAction(ISD::CTLZ, MVT::v16i8, Expand);
254     setOperationAction(ISD::CTTZ, MVT::v16i8, Expand);
255 
256     // Custom lower bit counting operations for other types to scalarize them.
257     for (auto Op : {ISD::CTLZ, ISD::CTTZ, ISD::CTPOP})
258       for (auto T : {MVT::v8i16, MVT::v4i32, MVT::v2i64})
259         setOperationAction(Op, T, Custom);
260 
261     // Expand float operations supported for scalars but not SIMD
262     for (auto Op : {ISD::FCOPYSIGN, ISD::FLOG, ISD::FLOG2, ISD::FLOG10,
263                     ISD::FEXP, ISD::FEXP2})
264       for (auto T : {MVT::v4f32, MVT::v2f64})
265         setOperationAction(Op, T, Expand);
266 
267     // Unsigned comparison operations are unavailable for i64x2 vectors.
268     for (auto CC : {ISD::SETUGT, ISD::SETUGE, ISD::SETULT, ISD::SETULE})
269       setCondCodeAction(CC, MVT::v2i64, Custom);
270 
271     // 64x2 conversions are not in the spec
272     for (auto Op :
273          {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT})
274       for (auto T : {MVT::v2i64, MVT::v2f64})
275         setOperationAction(Op, T, Expand);
276 
277     // But saturating fp_to_int converstions are
278     for (auto Op : {ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT})
279       setOperationAction(Op, MVT::v4i32, Custom);
280 
281     // Support vector extending
282     for (auto T : MVT::integer_fixedlen_vector_valuetypes()) {
283       setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, T, Custom);
284       setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, T, Custom);
285     }
286   }
287 
288   // As a special case, these operators use the type to mean the type to
289   // sign-extend from.
290   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
291   if (!Subtarget->hasSignExt()) {
292     // Sign extends are legal only when extending a vector extract
293     auto Action = Subtarget->hasSIMD128() ? Custom : Expand;
294     for (auto T : {MVT::i8, MVT::i16, MVT::i32})
295       setOperationAction(ISD::SIGN_EXTEND_INREG, T, Action);
296   }
297   for (auto T : MVT::integer_fixedlen_vector_valuetypes())
298     setOperationAction(ISD::SIGN_EXTEND_INREG, T, Expand);
299 
300   // Dynamic stack allocation: use the default expansion.
301   setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
302   setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
303   setOperationAction(ISD::DYNAMIC_STACKALLOC, MVTPtr, Expand);
304 
305   setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
306   setOperationAction(ISD::FrameIndex, MVT::i64, Custom);
307   setOperationAction(ISD::CopyToReg, MVT::Other, Custom);
308 
309   // Expand these forms; we pattern-match the forms that we can handle in isel.
310   for (auto T : {MVT::i32, MVT::i64, MVT::f32, MVT::f64})
311     for (auto Op : {ISD::BR_CC, ISD::SELECT_CC})
312       setOperationAction(Op, T, Expand);
313 
314   // We have custom switch handling.
315   setOperationAction(ISD::BR_JT, MVT::Other, Custom);
316 
317   // WebAssembly doesn't have:
318   //  - Floating-point extending loads.
319   //  - Floating-point truncating stores.
320   //  - i1 extending loads.
321   //  - truncating SIMD stores and most extending loads
322   setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
323   setTruncStoreAction(MVT::f64, MVT::f32, Expand);
324   for (auto T : MVT::integer_valuetypes())
325     for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
326       setLoadExtAction(Ext, T, MVT::i1, Promote);
327   if (Subtarget->hasSIMD128()) {
328     for (auto T : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v4f32,
329                    MVT::v2f64}) {
330       for (auto MemT : MVT::fixedlen_vector_valuetypes()) {
331         if (MVT(T) != MemT) {
332           setTruncStoreAction(T, MemT, Expand);
333           for (auto Ext : {ISD::EXTLOAD, ISD::ZEXTLOAD, ISD::SEXTLOAD})
334             setLoadExtAction(Ext, T, MemT, Expand);
335         }
336       }
337     }
338     // But some vector extending loads are legal
339     for (auto Ext : {ISD::EXTLOAD, ISD::SEXTLOAD, ISD::ZEXTLOAD}) {
340       setLoadExtAction(Ext, MVT::v8i16, MVT::v8i8, Legal);
341       setLoadExtAction(Ext, MVT::v4i32, MVT::v4i16, Legal);
342       setLoadExtAction(Ext, MVT::v2i64, MVT::v2i32, Legal);
343     }
344     setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Legal);
345   }
346 
347   // Don't do anything clever with build_pairs
348   setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
349 
350   // Trap lowers to wasm unreachable
351   setOperationAction(ISD::TRAP, MVT::Other, Legal);
352   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
353 
354   // Exception handling intrinsics
355   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
356   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
357   setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
358 
359   setMaxAtomicSizeInBitsSupported(64);
360 
361   // Override the __gnu_f2h_ieee/__gnu_h2f_ieee names so that the f32 name is
362   // consistent with the f64 and f128 names.
363   setLibcallName(RTLIB::FPEXT_F16_F32, "__extendhfsf2");
364   setLibcallName(RTLIB::FPROUND_F32_F16, "__truncsfhf2");
365 
366   // Define the emscripten name for return address helper.
367   // TODO: when implementing other Wasm backends, make this generic or only do
368   // this on emscripten depending on what they end up doing.
369   setLibcallName(RTLIB::RETURN_ADDRESS, "emscripten_return_address");
370 
371   // Always convert switches to br_tables unless there is only one case, which
372   // is equivalent to a simple branch. This reduces code size for wasm, and we
373   // defer possible jump table optimizations to the VM.
374   setMinimumJumpTableEntries(2);
375 }
376 
getPointerTy(const DataLayout & DL,uint32_t AS) const377 MVT WebAssemblyTargetLowering::getPointerTy(const DataLayout &DL,
378                                             uint32_t AS) const {
379   if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF)
380     return MVT::externref;
381   if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF)
382     return MVT::funcref;
383   return TargetLowering::getPointerTy(DL, AS);
384 }
385 
getPointerMemTy(const DataLayout & DL,uint32_t AS) const386 MVT WebAssemblyTargetLowering::getPointerMemTy(const DataLayout &DL,
387                                                uint32_t AS) const {
388   if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_EXTERNREF)
389     return MVT::externref;
390   if (AS == WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF)
391     return MVT::funcref;
392   return TargetLowering::getPointerMemTy(DL, AS);
393 }
394 
395 TargetLowering::AtomicExpansionKind
shouldExpandAtomicRMWInIR(AtomicRMWInst * AI) const396 WebAssemblyTargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
397   // We have wasm instructions for these
398   switch (AI->getOperation()) {
399   case AtomicRMWInst::Add:
400   case AtomicRMWInst::Sub:
401   case AtomicRMWInst::And:
402   case AtomicRMWInst::Or:
403   case AtomicRMWInst::Xor:
404   case AtomicRMWInst::Xchg:
405     return AtomicExpansionKind::None;
406   default:
407     break;
408   }
409   return AtomicExpansionKind::CmpXChg;
410 }
411 
shouldScalarizeBinop(SDValue VecOp) const412 bool WebAssemblyTargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
413   // Implementation copied from X86TargetLowering.
414   unsigned Opc = VecOp.getOpcode();
415 
416   // Assume target opcodes can't be scalarized.
417   // TODO - do we have any exceptions?
418   if (Opc >= ISD::BUILTIN_OP_END)
419     return false;
420 
421   // If the vector op is not supported, try to convert to scalar.
422   EVT VecVT = VecOp.getValueType();
423   if (!isOperationLegalOrCustomOrPromote(Opc, VecVT))
424     return true;
425 
426   // If the vector op is supported, but the scalar op is not, the transform may
427   // not be worthwhile.
428   EVT ScalarVT = VecVT.getScalarType();
429   return isOperationLegalOrCustomOrPromote(Opc, ScalarVT);
430 }
431 
createFastISel(FunctionLoweringInfo & FuncInfo,const TargetLibraryInfo * LibInfo) const432 FastISel *WebAssemblyTargetLowering::createFastISel(
433     FunctionLoweringInfo &FuncInfo, const TargetLibraryInfo *LibInfo) const {
434   return WebAssembly::createFastISel(FuncInfo, LibInfo);
435 }
436 
getScalarShiftAmountTy(const DataLayout &,EVT VT) const437 MVT WebAssemblyTargetLowering::getScalarShiftAmountTy(const DataLayout & /*DL*/,
438                                                       EVT VT) const {
439   unsigned BitWidth = NextPowerOf2(VT.getSizeInBits() - 1);
440   if (BitWidth > 1 && BitWidth < 8)
441     BitWidth = 8;
442 
443   if (BitWidth > 64) {
444     // The shift will be lowered to a libcall, and compiler-rt libcalls expect
445     // the count to be an i32.
446     BitWidth = 32;
447     assert(BitWidth >= Log2_32_Ceil(VT.getSizeInBits()) &&
448            "32-bit shift counts ought to be enough for anyone");
449   }
450 
451   MVT Result = MVT::getIntegerVT(BitWidth);
452   assert(Result != MVT::INVALID_SIMPLE_VALUE_TYPE &&
453          "Unable to represent scalar shift amount type");
454   return Result;
455 }
456 
457 // Lower an fp-to-int conversion operator from the LLVM opcode, which has an
458 // undefined result on invalid/overflow, to the WebAssembly opcode, which
459 // traps on invalid/overflow.
LowerFPToInt(MachineInstr & MI,DebugLoc DL,MachineBasicBlock * BB,const TargetInstrInfo & TII,bool IsUnsigned,bool Int64,bool Float64,unsigned LoweredOpcode)460 static MachineBasicBlock *LowerFPToInt(MachineInstr &MI, DebugLoc DL,
461                                        MachineBasicBlock *BB,
462                                        const TargetInstrInfo &TII,
463                                        bool IsUnsigned, bool Int64,
464                                        bool Float64, unsigned LoweredOpcode) {
465   MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
466 
467   Register OutReg = MI.getOperand(0).getReg();
468   Register InReg = MI.getOperand(1).getReg();
469 
470   unsigned Abs = Float64 ? WebAssembly::ABS_F64 : WebAssembly::ABS_F32;
471   unsigned FConst = Float64 ? WebAssembly::CONST_F64 : WebAssembly::CONST_F32;
472   unsigned LT = Float64 ? WebAssembly::LT_F64 : WebAssembly::LT_F32;
473   unsigned GE = Float64 ? WebAssembly::GE_F64 : WebAssembly::GE_F32;
474   unsigned IConst = Int64 ? WebAssembly::CONST_I64 : WebAssembly::CONST_I32;
475   unsigned Eqz = WebAssembly::EQZ_I32;
476   unsigned And = WebAssembly::AND_I32;
477   int64_t Limit = Int64 ? INT64_MIN : INT32_MIN;
478   int64_t Substitute = IsUnsigned ? 0 : Limit;
479   double CmpVal = IsUnsigned ? -(double)Limit * 2.0 : -(double)Limit;
480   auto &Context = BB->getParent()->getFunction().getContext();
481   Type *Ty = Float64 ? Type::getDoubleTy(Context) : Type::getFloatTy(Context);
482 
483   const BasicBlock *LLVMBB = BB->getBasicBlock();
484   MachineFunction *F = BB->getParent();
485   MachineBasicBlock *TrueMBB = F->CreateMachineBasicBlock(LLVMBB);
486   MachineBasicBlock *FalseMBB = F->CreateMachineBasicBlock(LLVMBB);
487   MachineBasicBlock *DoneMBB = F->CreateMachineBasicBlock(LLVMBB);
488 
489   MachineFunction::iterator It = ++BB->getIterator();
490   F->insert(It, FalseMBB);
491   F->insert(It, TrueMBB);
492   F->insert(It, DoneMBB);
493 
494   // Transfer the remainder of BB and its successor edges to DoneMBB.
495   DoneMBB->splice(DoneMBB->begin(), BB, std::next(MI.getIterator()), BB->end());
496   DoneMBB->transferSuccessorsAndUpdatePHIs(BB);
497 
498   BB->addSuccessor(TrueMBB);
499   BB->addSuccessor(FalseMBB);
500   TrueMBB->addSuccessor(DoneMBB);
501   FalseMBB->addSuccessor(DoneMBB);
502 
503   unsigned Tmp0, Tmp1, CmpReg, EqzReg, FalseReg, TrueReg;
504   Tmp0 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
505   Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
506   CmpReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
507   EqzReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
508   FalseReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
509   TrueReg = MRI.createVirtualRegister(MRI.getRegClass(OutReg));
510 
511   MI.eraseFromParent();
512   // For signed numbers, we can do a single comparison to determine whether
513   // fabs(x) is within range.
514   if (IsUnsigned) {
515     Tmp0 = InReg;
516   } else {
517     BuildMI(BB, DL, TII.get(Abs), Tmp0).addReg(InReg);
518   }
519   BuildMI(BB, DL, TII.get(FConst), Tmp1)
520       .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, CmpVal)));
521   BuildMI(BB, DL, TII.get(LT), CmpReg).addReg(Tmp0).addReg(Tmp1);
522 
523   // For unsigned numbers, we have to do a separate comparison with zero.
524   if (IsUnsigned) {
525     Tmp1 = MRI.createVirtualRegister(MRI.getRegClass(InReg));
526     Register SecondCmpReg =
527         MRI.createVirtualRegister(&WebAssembly::I32RegClass);
528     Register AndReg = MRI.createVirtualRegister(&WebAssembly::I32RegClass);
529     BuildMI(BB, DL, TII.get(FConst), Tmp1)
530         .addFPImm(cast<ConstantFP>(ConstantFP::get(Ty, 0.0)));
531     BuildMI(BB, DL, TII.get(GE), SecondCmpReg).addReg(Tmp0).addReg(Tmp1);
532     BuildMI(BB, DL, TII.get(And), AndReg).addReg(CmpReg).addReg(SecondCmpReg);
533     CmpReg = AndReg;
534   }
535 
536   BuildMI(BB, DL, TII.get(Eqz), EqzReg).addReg(CmpReg);
537 
538   // Create the CFG diamond to select between doing the conversion or using
539   // the substitute value.
540   BuildMI(BB, DL, TII.get(WebAssembly::BR_IF)).addMBB(TrueMBB).addReg(EqzReg);
541   BuildMI(FalseMBB, DL, TII.get(LoweredOpcode), FalseReg).addReg(InReg);
542   BuildMI(FalseMBB, DL, TII.get(WebAssembly::BR)).addMBB(DoneMBB);
543   BuildMI(TrueMBB, DL, TII.get(IConst), TrueReg).addImm(Substitute);
544   BuildMI(*DoneMBB, DoneMBB->begin(), DL, TII.get(TargetOpcode::PHI), OutReg)
545       .addReg(FalseReg)
546       .addMBB(FalseMBB)
547       .addReg(TrueReg)
548       .addMBB(TrueMBB);
549 
550   return DoneMBB;
551 }
552 
553 static MachineBasicBlock *
LowerCallResults(MachineInstr & CallResults,DebugLoc DL,MachineBasicBlock * BB,const WebAssemblySubtarget * Subtarget,const TargetInstrInfo & TII)554 LowerCallResults(MachineInstr &CallResults, DebugLoc DL, MachineBasicBlock *BB,
555                  const WebAssemblySubtarget *Subtarget,
556                  const TargetInstrInfo &TII) {
557   MachineInstr &CallParams = *CallResults.getPrevNode();
558   assert(CallParams.getOpcode() == WebAssembly::CALL_PARAMS);
559   assert(CallResults.getOpcode() == WebAssembly::CALL_RESULTS ||
560          CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS);
561 
562   bool IsIndirect =
563       CallParams.getOperand(0).isReg() || CallParams.getOperand(0).isFI();
564   bool IsRetCall = CallResults.getOpcode() == WebAssembly::RET_CALL_RESULTS;
565 
566   bool IsFuncrefCall = false;
567   if (IsIndirect && CallParams.getOperand(0).isReg()) {
568     Register Reg = CallParams.getOperand(0).getReg();
569     const MachineFunction *MF = BB->getParent();
570     const MachineRegisterInfo &MRI = MF->getRegInfo();
571     const TargetRegisterClass *TRC = MRI.getRegClass(Reg);
572     IsFuncrefCall = (TRC == &WebAssembly::FUNCREFRegClass);
573     assert(!IsFuncrefCall || Subtarget->hasReferenceTypes());
574   }
575 
576   unsigned CallOp;
577   if (IsIndirect && IsRetCall) {
578     CallOp = WebAssembly::RET_CALL_INDIRECT;
579   } else if (IsIndirect) {
580     CallOp = WebAssembly::CALL_INDIRECT;
581   } else if (IsRetCall) {
582     CallOp = WebAssembly::RET_CALL;
583   } else {
584     CallOp = WebAssembly::CALL;
585   }
586 
587   MachineFunction &MF = *BB->getParent();
588   const MCInstrDesc &MCID = TII.get(CallOp);
589   MachineInstrBuilder MIB(MF, MF.CreateMachineInstr(MCID, DL));
590 
591   // Move the function pointer to the end of the arguments for indirect calls
592   if (IsIndirect) {
593     auto FnPtr = CallParams.getOperand(0);
594     CallParams.removeOperand(0);
595 
596     // For funcrefs, call_indirect is done through __funcref_call_table and the
597     // funcref is always installed in slot 0 of the table, therefore instead of
598     // having the function pointer added at the end of the params list, a zero
599     // (the index in
600     // __funcref_call_table is added).
601     if (IsFuncrefCall) {
602       Register RegZero =
603           MF.getRegInfo().createVirtualRegister(&WebAssembly::I32RegClass);
604       MachineInstrBuilder MIBC0 =
605           BuildMI(MF, DL, TII.get(WebAssembly::CONST_I32), RegZero).addImm(0);
606 
607       BB->insert(CallResults.getIterator(), MIBC0);
608       MachineInstrBuilder(MF, CallParams).addReg(RegZero);
609     } else
610       CallParams.addOperand(FnPtr);
611   }
612 
613   for (auto Def : CallResults.defs())
614     MIB.add(Def);
615 
616   if (IsIndirect) {
617     // Placeholder for the type index.
618     MIB.addImm(0);
619     // The table into which this call_indirect indexes.
620     MCSymbolWasm *Table = IsFuncrefCall
621                               ? WebAssembly::getOrCreateFuncrefCallTableSymbol(
622                                     MF.getContext(), Subtarget)
623                               : WebAssembly::getOrCreateFunctionTableSymbol(
624                                     MF.getContext(), Subtarget);
625     if (Subtarget->hasReferenceTypes()) {
626       MIB.addSym(Table);
627     } else {
628       // For the MVP there is at most one table whose number is 0, but we can't
629       // write a table symbol or issue relocations.  Instead we just ensure the
630       // table is live and write a zero.
631       Table->setNoStrip();
632       MIB.addImm(0);
633     }
634   }
635 
636   for (auto Use : CallParams.uses())
637     MIB.add(Use);
638 
639   BB->insert(CallResults.getIterator(), MIB);
640   CallParams.eraseFromParent();
641   CallResults.eraseFromParent();
642 
643   // If this is a funcref call, to avoid hidden GC roots, we need to clear the
644   // table slot with ref.null upon call_indirect return.
645   //
646   // This generates the following code, which comes right after a call_indirect
647   // of a funcref:
648   //
649   //    i32.const 0
650   //    ref.null func
651   //    table.set __funcref_call_table
652   if (IsIndirect && IsFuncrefCall) {
653     MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol(
654         MF.getContext(), Subtarget);
655     Register RegZero =
656         MF.getRegInfo().createVirtualRegister(&WebAssembly::I32RegClass);
657     MachineInstr *Const0 =
658         BuildMI(MF, DL, TII.get(WebAssembly::CONST_I32), RegZero).addImm(0);
659     BB->insertAfter(MIB.getInstr()->getIterator(), Const0);
660 
661     Register RegFuncref =
662         MF.getRegInfo().createVirtualRegister(&WebAssembly::FUNCREFRegClass);
663     MachineInstr *RefNull =
664         BuildMI(MF, DL, TII.get(WebAssembly::REF_NULL_FUNCREF), RegFuncref);
665     BB->insertAfter(Const0->getIterator(), RefNull);
666 
667     MachineInstr *TableSet =
668         BuildMI(MF, DL, TII.get(WebAssembly::TABLE_SET_FUNCREF))
669             .addSym(Table)
670             .addReg(RegZero)
671             .addReg(RegFuncref);
672     BB->insertAfter(RefNull->getIterator(), TableSet);
673   }
674 
675   return BB;
676 }
677 
EmitInstrWithCustomInserter(MachineInstr & MI,MachineBasicBlock * BB) const678 MachineBasicBlock *WebAssemblyTargetLowering::EmitInstrWithCustomInserter(
679     MachineInstr &MI, MachineBasicBlock *BB) const {
680   const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
681   DebugLoc DL = MI.getDebugLoc();
682 
683   switch (MI.getOpcode()) {
684   default:
685     llvm_unreachable("Unexpected instr type to insert");
686   case WebAssembly::FP_TO_SINT_I32_F32:
687     return LowerFPToInt(MI, DL, BB, TII, false, false, false,
688                         WebAssembly::I32_TRUNC_S_F32);
689   case WebAssembly::FP_TO_UINT_I32_F32:
690     return LowerFPToInt(MI, DL, BB, TII, true, false, false,
691                         WebAssembly::I32_TRUNC_U_F32);
692   case WebAssembly::FP_TO_SINT_I64_F32:
693     return LowerFPToInt(MI, DL, BB, TII, false, true, false,
694                         WebAssembly::I64_TRUNC_S_F32);
695   case WebAssembly::FP_TO_UINT_I64_F32:
696     return LowerFPToInt(MI, DL, BB, TII, true, true, false,
697                         WebAssembly::I64_TRUNC_U_F32);
698   case WebAssembly::FP_TO_SINT_I32_F64:
699     return LowerFPToInt(MI, DL, BB, TII, false, false, true,
700                         WebAssembly::I32_TRUNC_S_F64);
701   case WebAssembly::FP_TO_UINT_I32_F64:
702     return LowerFPToInt(MI, DL, BB, TII, true, false, true,
703                         WebAssembly::I32_TRUNC_U_F64);
704   case WebAssembly::FP_TO_SINT_I64_F64:
705     return LowerFPToInt(MI, DL, BB, TII, false, true, true,
706                         WebAssembly::I64_TRUNC_S_F64);
707   case WebAssembly::FP_TO_UINT_I64_F64:
708     return LowerFPToInt(MI, DL, BB, TII, true, true, true,
709                         WebAssembly::I64_TRUNC_U_F64);
710   case WebAssembly::CALL_RESULTS:
711   case WebAssembly::RET_CALL_RESULTS:
712     return LowerCallResults(MI, DL, BB, Subtarget, TII);
713   }
714 }
715 
716 const char *
getTargetNodeName(unsigned Opcode) const717 WebAssemblyTargetLowering::getTargetNodeName(unsigned Opcode) const {
718   switch (static_cast<WebAssemblyISD::NodeType>(Opcode)) {
719   case WebAssemblyISD::FIRST_NUMBER:
720   case WebAssemblyISD::FIRST_MEM_OPCODE:
721     break;
722 #define HANDLE_NODETYPE(NODE)                                                  \
723   case WebAssemblyISD::NODE:                                                   \
724     return "WebAssemblyISD::" #NODE;
725 #define HANDLE_MEM_NODETYPE(NODE) HANDLE_NODETYPE(NODE)
726 #include "WebAssemblyISD.def"
727 #undef HANDLE_MEM_NODETYPE
728 #undef HANDLE_NODETYPE
729   }
730   return nullptr;
731 }
732 
733 std::pair<unsigned, const TargetRegisterClass *>
getRegForInlineAsmConstraint(const TargetRegisterInfo * TRI,StringRef Constraint,MVT VT) const734 WebAssemblyTargetLowering::getRegForInlineAsmConstraint(
735     const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
736   // First, see if this is a constraint that directly corresponds to a
737   // WebAssembly register class.
738   if (Constraint.size() == 1) {
739     switch (Constraint[0]) {
740     case 'r':
741       assert(VT != MVT::iPTR && "Pointer MVT not expected here");
742       if (Subtarget->hasSIMD128() && VT.isVector()) {
743         if (VT.getSizeInBits() == 128)
744           return std::make_pair(0U, &WebAssembly::V128RegClass);
745       }
746       if (VT.isInteger() && !VT.isVector()) {
747         if (VT.getSizeInBits() <= 32)
748           return std::make_pair(0U, &WebAssembly::I32RegClass);
749         if (VT.getSizeInBits() <= 64)
750           return std::make_pair(0U, &WebAssembly::I64RegClass);
751       }
752       if (VT.isFloatingPoint() && !VT.isVector()) {
753         switch (VT.getSizeInBits()) {
754         case 32:
755           return std::make_pair(0U, &WebAssembly::F32RegClass);
756         case 64:
757           return std::make_pair(0U, &WebAssembly::F64RegClass);
758         default:
759           break;
760         }
761       }
762       break;
763     default:
764       break;
765     }
766   }
767 
768   return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
769 }
770 
isCheapToSpeculateCttz(Type * Ty) const771 bool WebAssemblyTargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
772   // Assume ctz is a relatively cheap operation.
773   return true;
774 }
775 
isCheapToSpeculateCtlz(Type * Ty) const776 bool WebAssemblyTargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
777   // Assume clz is a relatively cheap operation.
778   return true;
779 }
780 
isLegalAddressingMode(const DataLayout & DL,const AddrMode & AM,Type * Ty,unsigned AS,Instruction * I) const781 bool WebAssemblyTargetLowering::isLegalAddressingMode(const DataLayout &DL,
782                                                       const AddrMode &AM,
783                                                       Type *Ty, unsigned AS,
784                                                       Instruction *I) const {
785   // WebAssembly offsets are added as unsigned without wrapping. The
786   // isLegalAddressingMode gives us no way to determine if wrapping could be
787   // happening, so we approximate this by accepting only non-negative offsets.
788   if (AM.BaseOffs < 0)
789     return false;
790 
791   // WebAssembly has no scale register operands.
792   if (AM.Scale != 0)
793     return false;
794 
795   // Everything else is legal.
796   return true;
797 }
798 
allowsMisalignedMemoryAccesses(EVT,unsigned,Align,MachineMemOperand::Flags,unsigned * Fast) const799 bool WebAssemblyTargetLowering::allowsMisalignedMemoryAccesses(
800     EVT /*VT*/, unsigned /*AddrSpace*/, Align /*Align*/,
801     MachineMemOperand::Flags /*Flags*/, unsigned *Fast) const {
802   // WebAssembly supports unaligned accesses, though it should be declared
803   // with the p2align attribute on loads and stores which do so, and there
804   // may be a performance impact. We tell LLVM they're "fast" because
805   // for the kinds of things that LLVM uses this for (merging adjacent stores
806   // of constants, etc.), WebAssembly implementations will either want the
807   // unaligned access or they'll split anyway.
808   if (Fast)
809     *Fast = 1;
810   return true;
811 }
812 
isIntDivCheap(EVT VT,AttributeList Attr) const813 bool WebAssemblyTargetLowering::isIntDivCheap(EVT VT,
814                                               AttributeList Attr) const {
815   // The current thinking is that wasm engines will perform this optimization,
816   // so we can save on code size.
817   return true;
818 }
819 
isVectorLoadExtDesirable(SDValue ExtVal) const820 bool WebAssemblyTargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
821   EVT ExtT = ExtVal.getValueType();
822   EVT MemT = cast<LoadSDNode>(ExtVal->getOperand(0))->getValueType(0);
823   return (ExtT == MVT::v8i16 && MemT == MVT::v8i8) ||
824          (ExtT == MVT::v4i32 && MemT == MVT::v4i16) ||
825          (ExtT == MVT::v2i64 && MemT == MVT::v2i32);
826 }
827 
isOffsetFoldingLegal(const GlobalAddressSDNode * GA) const828 bool WebAssemblyTargetLowering::isOffsetFoldingLegal(
829     const GlobalAddressSDNode *GA) const {
830   // Wasm doesn't support function addresses with offsets
831   const GlobalValue *GV = GA->getGlobal();
832   return isa<Function>(GV) ? false : TargetLowering::isOffsetFoldingLegal(GA);
833 }
834 
shouldSinkOperands(Instruction * I,SmallVectorImpl<Use * > & Ops) const835 bool WebAssemblyTargetLowering::shouldSinkOperands(
836     Instruction *I, SmallVectorImpl<Use *> &Ops) const {
837   using namespace llvm::PatternMatch;
838 
839   if (!I->getType()->isVectorTy() || !I->isShift())
840     return false;
841 
842   Value *V = I->getOperand(1);
843   // We dont need to sink constant splat.
844   if (dyn_cast<Constant>(V))
845     return false;
846 
847   if (match(V, m_Shuffle(m_InsertElt(m_Value(), m_Value(), m_ZeroInt()),
848                          m_Value(), m_ZeroMask()))) {
849     // Sink insert
850     Ops.push_back(&cast<Instruction>(V)->getOperandUse(0));
851     // Sink shuffle
852     Ops.push_back(&I->getOperandUse(1));
853     return true;
854   }
855 
856   return false;
857 }
858 
getSetCCResultType(const DataLayout & DL,LLVMContext & C,EVT VT) const859 EVT WebAssemblyTargetLowering::getSetCCResultType(const DataLayout &DL,
860                                                   LLVMContext &C,
861                                                   EVT VT) const {
862   if (VT.isVector())
863     return VT.changeVectorElementTypeToInteger();
864 
865   // So far, all branch instructions in Wasm take an I32 condition.
866   // The default TargetLowering::getSetCCResultType returns the pointer size,
867   // which would be useful to reduce instruction counts when testing
868   // against 64-bit pointers/values if at some point Wasm supports that.
869   return EVT::getIntegerVT(C, 32);
870 }
871 
getTgtMemIntrinsic(IntrinsicInfo & Info,const CallInst & I,MachineFunction & MF,unsigned Intrinsic) const872 bool WebAssemblyTargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
873                                                    const CallInst &I,
874                                                    MachineFunction &MF,
875                                                    unsigned Intrinsic) const {
876   switch (Intrinsic) {
877   case Intrinsic::wasm_memory_atomic_notify:
878     Info.opc = ISD::INTRINSIC_W_CHAIN;
879     Info.memVT = MVT::i32;
880     Info.ptrVal = I.getArgOperand(0);
881     Info.offset = 0;
882     Info.align = Align(4);
883     // atomic.notify instruction does not really load the memory specified with
884     // this argument, but MachineMemOperand should either be load or store, so
885     // we set this to a load.
886     // FIXME Volatile isn't really correct, but currently all LLVM atomic
887     // instructions are treated as volatiles in the backend, so we should be
888     // consistent. The same applies for wasm_atomic_wait intrinsics too.
889     Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
890     return true;
891   case Intrinsic::wasm_memory_atomic_wait32:
892     Info.opc = ISD::INTRINSIC_W_CHAIN;
893     Info.memVT = MVT::i32;
894     Info.ptrVal = I.getArgOperand(0);
895     Info.offset = 0;
896     Info.align = Align(4);
897     Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
898     return true;
899   case Intrinsic::wasm_memory_atomic_wait64:
900     Info.opc = ISD::INTRINSIC_W_CHAIN;
901     Info.memVT = MVT::i64;
902     Info.ptrVal = I.getArgOperand(0);
903     Info.offset = 0;
904     Info.align = Align(8);
905     Info.flags = MachineMemOperand::MOVolatile | MachineMemOperand::MOLoad;
906     return true;
907   case Intrinsic::wasm_loadf16_f32:
908     Info.opc = ISD::INTRINSIC_W_CHAIN;
909     Info.memVT = MVT::f16;
910     Info.ptrVal = I.getArgOperand(0);
911     Info.offset = 0;
912     Info.align = Align(2);
913     Info.flags = MachineMemOperand::MOLoad;
914     return true;
915   case Intrinsic::wasm_storef16_f32:
916     Info.opc = ISD::INTRINSIC_VOID;
917     Info.memVT = MVT::f16;
918     Info.ptrVal = I.getArgOperand(1);
919     Info.offset = 0;
920     Info.align = Align(2);
921     Info.flags = MachineMemOperand::MOStore;
922     return true;
923   default:
924     return false;
925   }
926 }
927 
computeKnownBitsForTargetNode(const SDValue Op,KnownBits & Known,const APInt & DemandedElts,const SelectionDAG & DAG,unsigned Depth) const928 void WebAssemblyTargetLowering::computeKnownBitsForTargetNode(
929     const SDValue Op, KnownBits &Known, const APInt &DemandedElts,
930     const SelectionDAG &DAG, unsigned Depth) const {
931   switch (Op.getOpcode()) {
932   default:
933     break;
934   case ISD::INTRINSIC_WO_CHAIN: {
935     unsigned IntNo = Op.getConstantOperandVal(0);
936     switch (IntNo) {
937     default:
938       break;
939     case Intrinsic::wasm_bitmask: {
940       unsigned BitWidth = Known.getBitWidth();
941       EVT VT = Op.getOperand(1).getSimpleValueType();
942       unsigned PossibleBits = VT.getVectorNumElements();
943       APInt ZeroMask = APInt::getHighBitsSet(BitWidth, BitWidth - PossibleBits);
944       Known.Zero |= ZeroMask;
945       break;
946     }
947     }
948   }
949   }
950 }
951 
952 TargetLoweringBase::LegalizeTypeAction
getPreferredVectorAction(MVT VT) const953 WebAssemblyTargetLowering::getPreferredVectorAction(MVT VT) const {
954   if (VT.isFixedLengthVector()) {
955     MVT EltVT = VT.getVectorElementType();
956     // We have legal vector types with these lane types, so widening the
957     // vector would let us use some of the lanes directly without having to
958     // extend or truncate values.
959     if (EltVT == MVT::i8 || EltVT == MVT::i16 || EltVT == MVT::i32 ||
960         EltVT == MVT::i64 || EltVT == MVT::f32 || EltVT == MVT::f64)
961       return TypeWidenVector;
962   }
963 
964   return TargetLoweringBase::getPreferredVectorAction(VT);
965 }
966 
shouldSimplifyDemandedVectorElts(SDValue Op,const TargetLoweringOpt & TLO) const967 bool WebAssemblyTargetLowering::shouldSimplifyDemandedVectorElts(
968     SDValue Op, const TargetLoweringOpt &TLO) const {
969   // ISel process runs DAGCombiner after legalization; this step is called
970   // SelectionDAG optimization phase. This post-legalization combining process
971   // runs DAGCombiner on each node, and if there was a change to be made,
972   // re-runs legalization again on it and its user nodes to make sure
973   // everythiing is in a legalized state.
974   //
975   // The legalization calls lowering routines, and we do our custom lowering for
976   // build_vectors (LowerBUILD_VECTOR), which converts undef vector elements
977   // into zeros. But there is a set of routines in DAGCombiner that turns unused
978   // (= not demanded) nodes into undef, among which SimplifyDemandedVectorElts
979   // turns unused vector elements into undefs. But this routine does not work
980   // with our custom LowerBUILD_VECTOR, which turns undefs into zeros. This
981   // combination can result in a infinite loop, in which undefs are converted to
982   // zeros in legalization and back to undefs in combining.
983   //
984   // So after DAG is legalized, we prevent SimplifyDemandedVectorElts from
985   // running for build_vectors.
986   if (Op.getOpcode() == ISD::BUILD_VECTOR && TLO.LegalOps && TLO.LegalTys)
987     return false;
988   return true;
989 }
990 
991 //===----------------------------------------------------------------------===//
992 // WebAssembly Lowering private implementation.
993 //===----------------------------------------------------------------------===//
994 
995 //===----------------------------------------------------------------------===//
996 // Lowering Code
997 //===----------------------------------------------------------------------===//
998 
fail(const SDLoc & DL,SelectionDAG & DAG,const char * Msg)999 static void fail(const SDLoc &DL, SelectionDAG &DAG, const char *Msg) {
1000   MachineFunction &MF = DAG.getMachineFunction();
1001   DAG.getContext()->diagnose(
1002       DiagnosticInfoUnsupported(MF.getFunction(), Msg, DL.getDebugLoc()));
1003 }
1004 
1005 // Test whether the given calling convention is supported.
callingConvSupported(CallingConv::ID CallConv)1006 static bool callingConvSupported(CallingConv::ID CallConv) {
1007   // We currently support the language-independent target-independent
1008   // conventions. We don't yet have a way to annotate calls with properties like
1009   // "cold", and we don't have any call-clobbered registers, so these are mostly
1010   // all handled the same.
1011   return CallConv == CallingConv::C || CallConv == CallingConv::Fast ||
1012          CallConv == CallingConv::Cold ||
1013          CallConv == CallingConv::PreserveMost ||
1014          CallConv == CallingConv::PreserveAll ||
1015          CallConv == CallingConv::CXX_FAST_TLS ||
1016          CallConv == CallingConv::WASM_EmscriptenInvoke ||
1017          CallConv == CallingConv::Swift;
1018 }
1019 
1020 SDValue
LowerCall(CallLoweringInfo & CLI,SmallVectorImpl<SDValue> & InVals) const1021 WebAssemblyTargetLowering::LowerCall(CallLoweringInfo &CLI,
1022                                      SmallVectorImpl<SDValue> &InVals) const {
1023   SelectionDAG &DAG = CLI.DAG;
1024   SDLoc DL = CLI.DL;
1025   SDValue Chain = CLI.Chain;
1026   SDValue Callee = CLI.Callee;
1027   MachineFunction &MF = DAG.getMachineFunction();
1028   auto Layout = MF.getDataLayout();
1029 
1030   CallingConv::ID CallConv = CLI.CallConv;
1031   if (!callingConvSupported(CallConv))
1032     fail(DL, DAG,
1033          "WebAssembly doesn't support language-specific or target-specific "
1034          "calling conventions yet");
1035   if (CLI.IsPatchPoint)
1036     fail(DL, DAG, "WebAssembly doesn't support patch point yet");
1037 
1038   if (CLI.IsTailCall) {
1039     auto NoTail = [&](const char *Msg) {
1040       if (CLI.CB && CLI.CB->isMustTailCall())
1041         fail(DL, DAG, Msg);
1042       CLI.IsTailCall = false;
1043     };
1044 
1045     if (!Subtarget->hasTailCall())
1046       NoTail("WebAssembly 'tail-call' feature not enabled");
1047 
1048     // Varargs calls cannot be tail calls because the buffer is on the stack
1049     if (CLI.IsVarArg)
1050       NoTail("WebAssembly does not support varargs tail calls");
1051 
1052     // Do not tail call unless caller and callee return types match
1053     const Function &F = MF.getFunction();
1054     const TargetMachine &TM = getTargetMachine();
1055     Type *RetTy = F.getReturnType();
1056     SmallVector<MVT, 4> CallerRetTys;
1057     SmallVector<MVT, 4> CalleeRetTys;
1058     computeLegalValueVTs(F, TM, RetTy, CallerRetTys);
1059     computeLegalValueVTs(F, TM, CLI.RetTy, CalleeRetTys);
1060     bool TypesMatch = CallerRetTys.size() == CalleeRetTys.size() &&
1061                       std::equal(CallerRetTys.begin(), CallerRetTys.end(),
1062                                  CalleeRetTys.begin());
1063     if (!TypesMatch)
1064       NoTail("WebAssembly tail call requires caller and callee return types to "
1065              "match");
1066 
1067     // If pointers to local stack values are passed, we cannot tail call
1068     if (CLI.CB) {
1069       for (auto &Arg : CLI.CB->args()) {
1070         Value *Val = Arg.get();
1071         // Trace the value back through pointer operations
1072         while (true) {
1073           Value *Src = Val->stripPointerCastsAndAliases();
1074           if (auto *GEP = dyn_cast<GetElementPtrInst>(Src))
1075             Src = GEP->getPointerOperand();
1076           if (Val == Src)
1077             break;
1078           Val = Src;
1079         }
1080         if (isa<AllocaInst>(Val)) {
1081           NoTail(
1082               "WebAssembly does not support tail calling with stack arguments");
1083           break;
1084         }
1085       }
1086     }
1087   }
1088 
1089   SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1090   SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1091   SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1092 
1093   // The generic code may have added an sret argument. If we're lowering an
1094   // invoke function, the ABI requires that the function pointer be the first
1095   // argument, so we may have to swap the arguments.
1096   if (CallConv == CallingConv::WASM_EmscriptenInvoke && Outs.size() >= 2 &&
1097       Outs[0].Flags.isSRet()) {
1098     std::swap(Outs[0], Outs[1]);
1099     std::swap(OutVals[0], OutVals[1]);
1100   }
1101 
1102   bool HasSwiftSelfArg = false;
1103   bool HasSwiftErrorArg = false;
1104   unsigned NumFixedArgs = 0;
1105   for (unsigned I = 0; I < Outs.size(); ++I) {
1106     const ISD::OutputArg &Out = Outs[I];
1107     SDValue &OutVal = OutVals[I];
1108     HasSwiftSelfArg |= Out.Flags.isSwiftSelf();
1109     HasSwiftErrorArg |= Out.Flags.isSwiftError();
1110     if (Out.Flags.isNest())
1111       fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
1112     if (Out.Flags.isInAlloca())
1113       fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
1114     if (Out.Flags.isInConsecutiveRegs())
1115       fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
1116     if (Out.Flags.isInConsecutiveRegsLast())
1117       fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
1118     if (Out.Flags.isByVal() && Out.Flags.getByValSize() != 0) {
1119       auto &MFI = MF.getFrameInfo();
1120       int FI = MFI.CreateStackObject(Out.Flags.getByValSize(),
1121                                      Out.Flags.getNonZeroByValAlign(),
1122                                      /*isSS=*/false);
1123       SDValue SizeNode =
1124           DAG.getConstant(Out.Flags.getByValSize(), DL, MVT::i32);
1125       SDValue FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
1126       Chain = DAG.getMemcpy(Chain, DL, FINode, OutVal, SizeNode,
1127                             Out.Flags.getNonZeroByValAlign(),
1128                             /*isVolatile*/ false, /*AlwaysInline=*/false,
1129                             /*CI=*/nullptr, std::nullopt, MachinePointerInfo(),
1130                             MachinePointerInfo());
1131       OutVal = FINode;
1132     }
1133     // Count the number of fixed args *after* legalization.
1134     NumFixedArgs += Out.IsFixed;
1135   }
1136 
1137   bool IsVarArg = CLI.IsVarArg;
1138   auto PtrVT = getPointerTy(Layout);
1139 
1140   // For swiftcc, emit additional swiftself and swifterror arguments
1141   // if there aren't. These additional arguments are also added for callee
1142   // signature They are necessary to match callee and caller signature for
1143   // indirect call.
1144   if (CallConv == CallingConv::Swift) {
1145     if (!HasSwiftSelfArg) {
1146       NumFixedArgs++;
1147       ISD::OutputArg Arg;
1148       Arg.Flags.setSwiftSelf();
1149       CLI.Outs.push_back(Arg);
1150       SDValue ArgVal = DAG.getUNDEF(PtrVT);
1151       CLI.OutVals.push_back(ArgVal);
1152     }
1153     if (!HasSwiftErrorArg) {
1154       NumFixedArgs++;
1155       ISD::OutputArg Arg;
1156       Arg.Flags.setSwiftError();
1157       CLI.Outs.push_back(Arg);
1158       SDValue ArgVal = DAG.getUNDEF(PtrVT);
1159       CLI.OutVals.push_back(ArgVal);
1160     }
1161   }
1162 
1163   // Analyze operands of the call, assigning locations to each operand.
1164   SmallVector<CCValAssign, 16> ArgLocs;
1165   CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());
1166 
1167   if (IsVarArg) {
1168     // Outgoing non-fixed arguments are placed in a buffer. First
1169     // compute their offsets and the total amount of buffer space needed.
1170     for (unsigned I = NumFixedArgs; I < Outs.size(); ++I) {
1171       const ISD::OutputArg &Out = Outs[I];
1172       SDValue &Arg = OutVals[I];
1173       EVT VT = Arg.getValueType();
1174       assert(VT != MVT::iPTR && "Legalized args should be concrete");
1175       Type *Ty = VT.getTypeForEVT(*DAG.getContext());
1176       Align Alignment =
1177           std::max(Out.Flags.getNonZeroOrigAlign(), Layout.getABITypeAlign(Ty));
1178       unsigned Offset =
1179           CCInfo.AllocateStack(Layout.getTypeAllocSize(Ty), Alignment);
1180       CCInfo.addLoc(CCValAssign::getMem(ArgLocs.size(), VT.getSimpleVT(),
1181                                         Offset, VT.getSimpleVT(),
1182                                         CCValAssign::Full));
1183     }
1184   }
1185 
1186   unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
1187 
1188   SDValue FINode;
1189   if (IsVarArg && NumBytes) {
1190     // For non-fixed arguments, next emit stores to store the argument values
1191     // to the stack buffer at the offsets computed above.
1192     int FI = MF.getFrameInfo().CreateStackObject(NumBytes,
1193                                                  Layout.getStackAlignment(),
1194                                                  /*isSS=*/false);
1195     unsigned ValNo = 0;
1196     SmallVector<SDValue, 8> Chains;
1197     for (SDValue Arg : drop_begin(OutVals, NumFixedArgs)) {
1198       assert(ArgLocs[ValNo].getValNo() == ValNo &&
1199              "ArgLocs should remain in order and only hold varargs args");
1200       unsigned Offset = ArgLocs[ValNo++].getLocMemOffset();
1201       FINode = DAG.getFrameIndex(FI, getPointerTy(Layout));
1202       SDValue Add = DAG.getNode(ISD::ADD, DL, PtrVT, FINode,
1203                                 DAG.getConstant(Offset, DL, PtrVT));
1204       Chains.push_back(
1205           DAG.getStore(Chain, DL, Arg, Add,
1206                        MachinePointerInfo::getFixedStack(MF, FI, Offset)));
1207     }
1208     if (!Chains.empty())
1209       Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
1210   } else if (IsVarArg) {
1211     FINode = DAG.getIntPtrConstant(0, DL);
1212   }
1213 
1214   if (Callee->getOpcode() == ISD::GlobalAddress) {
1215     // If the callee is a GlobalAddress node (quite common, every direct call
1216     // is) turn it into a TargetGlobalAddress node so that LowerGlobalAddress
1217     // doesn't at MO_GOT which is not needed for direct calls.
1218     GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Callee);
1219     Callee = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
1220                                         getPointerTy(DAG.getDataLayout()),
1221                                         GA->getOffset());
1222     Callee = DAG.getNode(WebAssemblyISD::Wrapper, DL,
1223                          getPointerTy(DAG.getDataLayout()), Callee);
1224   }
1225 
1226   // Compute the operands for the CALLn node.
1227   SmallVector<SDValue, 16> Ops;
1228   Ops.push_back(Chain);
1229   Ops.push_back(Callee);
1230 
1231   // Add all fixed arguments. Note that for non-varargs calls, NumFixedArgs
1232   // isn't reliable.
1233   Ops.append(OutVals.begin(),
1234              IsVarArg ? OutVals.begin() + NumFixedArgs : OutVals.end());
1235   // Add a pointer to the vararg buffer.
1236   if (IsVarArg)
1237     Ops.push_back(FINode);
1238 
1239   SmallVector<EVT, 8> InTys;
1240   for (const auto &In : Ins) {
1241     assert(!In.Flags.isByVal() && "byval is not valid for return values");
1242     assert(!In.Flags.isNest() && "nest is not valid for return values");
1243     if (In.Flags.isInAlloca())
1244       fail(DL, DAG, "WebAssembly hasn't implemented inalloca return values");
1245     if (In.Flags.isInConsecutiveRegs())
1246       fail(DL, DAG, "WebAssembly hasn't implemented cons regs return values");
1247     if (In.Flags.isInConsecutiveRegsLast())
1248       fail(DL, DAG,
1249            "WebAssembly hasn't implemented cons regs last return values");
1250     // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in
1251     // registers.
1252     InTys.push_back(In.VT);
1253   }
1254 
1255   // Lastly, if this is a call to a funcref we need to add an instruction
1256   // table.set to the chain and transform the call.
1257   if (CLI.CB && WebAssembly::isWebAssemblyFuncrefType(
1258                     CLI.CB->getCalledOperand()->getType())) {
1259     // In the absence of function references proposal where a funcref call is
1260     // lowered to call_ref, using reference types we generate a table.set to set
1261     // the funcref to a special table used solely for this purpose, followed by
1262     // a call_indirect. Here we just generate the table set, and return the
1263     // SDValue of the table.set so that LowerCall can finalize the lowering by
1264     // generating the call_indirect.
1265     SDValue Chain = Ops[0];
1266 
1267     MCSymbolWasm *Table = WebAssembly::getOrCreateFuncrefCallTableSymbol(
1268         MF.getContext(), Subtarget);
1269     SDValue Sym = DAG.getMCSymbol(Table, PtrVT);
1270     SDValue TableSlot = DAG.getConstant(0, DL, MVT::i32);
1271     SDValue TableSetOps[] = {Chain, Sym, TableSlot, Callee};
1272     SDValue TableSet = DAG.getMemIntrinsicNode(
1273         WebAssemblyISD::TABLE_SET, DL, DAG.getVTList(MVT::Other), TableSetOps,
1274         MVT::funcref,
1275         // Machine Mem Operand args
1276         MachinePointerInfo(
1277             WebAssembly::WasmAddressSpace::WASM_ADDRESS_SPACE_FUNCREF),
1278         CLI.CB->getCalledOperand()->getPointerAlignment(DAG.getDataLayout()),
1279         MachineMemOperand::MOStore);
1280 
1281     Ops[0] = TableSet; // The new chain is the TableSet itself
1282   }
1283 
1284   if (CLI.IsTailCall) {
1285     // ret_calls do not return values to the current frame
1286     SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
1287     return DAG.getNode(WebAssemblyISD::RET_CALL, DL, NodeTys, Ops);
1288   }
1289 
1290   InTys.push_back(MVT::Other);
1291   SDVTList InTyList = DAG.getVTList(InTys);
1292   SDValue Res = DAG.getNode(WebAssemblyISD::CALL, DL, InTyList, Ops);
1293 
1294   for (size_t I = 0; I < Ins.size(); ++I)
1295     InVals.push_back(Res.getValue(I));
1296 
1297   // Return the chain
1298   return Res.getValue(Ins.size());
1299 }
1300 
CanLowerReturn(CallingConv::ID,MachineFunction &,bool,const SmallVectorImpl<ISD::OutputArg> & Outs,LLVMContext &) const1301 bool WebAssemblyTargetLowering::CanLowerReturn(
1302     CallingConv::ID /*CallConv*/, MachineFunction & /*MF*/, bool /*IsVarArg*/,
1303     const SmallVectorImpl<ISD::OutputArg> &Outs,
1304     LLVMContext & /*Context*/) const {
1305   // WebAssembly can only handle returning tuples with multivalue enabled
1306   return WebAssembly::canLowerReturn(Outs.size(), Subtarget);
1307 }
1308 
LowerReturn(SDValue Chain,CallingConv::ID CallConv,bool,const SmallVectorImpl<ISD::OutputArg> & Outs,const SmallVectorImpl<SDValue> & OutVals,const SDLoc & DL,SelectionDAG & DAG) const1309 SDValue WebAssemblyTargetLowering::LowerReturn(
1310     SDValue Chain, CallingConv::ID CallConv, bool /*IsVarArg*/,
1311     const SmallVectorImpl<ISD::OutputArg> &Outs,
1312     const SmallVectorImpl<SDValue> &OutVals, const SDLoc &DL,
1313     SelectionDAG &DAG) const {
1314   assert(WebAssembly::canLowerReturn(Outs.size(), Subtarget) &&
1315          "MVP WebAssembly can only return up to one value");
1316   if (!callingConvSupported(CallConv))
1317     fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
1318 
1319   SmallVector<SDValue, 4> RetOps(1, Chain);
1320   RetOps.append(OutVals.begin(), OutVals.end());
1321   Chain = DAG.getNode(WebAssemblyISD::RETURN, DL, MVT::Other, RetOps);
1322 
1323   // Record the number and types of the return values.
1324   for (const ISD::OutputArg &Out : Outs) {
1325     assert(!Out.Flags.isByVal() && "byval is not valid for return values");
1326     assert(!Out.Flags.isNest() && "nest is not valid for return values");
1327     assert(Out.IsFixed && "non-fixed return value is not valid");
1328     if (Out.Flags.isInAlloca())
1329       fail(DL, DAG, "WebAssembly hasn't implemented inalloca results");
1330     if (Out.Flags.isInConsecutiveRegs())
1331       fail(DL, DAG, "WebAssembly hasn't implemented cons regs results");
1332     if (Out.Flags.isInConsecutiveRegsLast())
1333       fail(DL, DAG, "WebAssembly hasn't implemented cons regs last results");
1334   }
1335 
1336   return Chain;
1337 }
1338 
LowerFormalArguments(SDValue Chain,CallingConv::ID CallConv,bool IsVarArg,const SmallVectorImpl<ISD::InputArg> & Ins,const SDLoc & DL,SelectionDAG & DAG,SmallVectorImpl<SDValue> & InVals) const1339 SDValue WebAssemblyTargetLowering::LowerFormalArguments(
1340     SDValue Chain, CallingConv::ID CallConv, bool IsVarArg,
1341     const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
1342     SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
1343   if (!callingConvSupported(CallConv))
1344     fail(DL, DAG, "WebAssembly doesn't support non-C calling conventions");
1345 
1346   MachineFunction &MF = DAG.getMachineFunction();
1347   auto *MFI = MF.getInfo<WebAssemblyFunctionInfo>();
1348 
1349   // Set up the incoming ARGUMENTS value, which serves to represent the liveness
1350   // of the incoming values before they're represented by virtual registers.
1351   MF.getRegInfo().addLiveIn(WebAssembly::ARGUMENTS);
1352 
1353   bool HasSwiftErrorArg = false;
1354   bool HasSwiftSelfArg = false;
1355   for (const ISD::InputArg &In : Ins) {
1356     HasSwiftSelfArg |= In.Flags.isSwiftSelf();
1357     HasSwiftErrorArg |= In.Flags.isSwiftError();
1358     if (In.Flags.isInAlloca())
1359       fail(DL, DAG, "WebAssembly hasn't implemented inalloca arguments");
1360     if (In.Flags.isNest())
1361       fail(DL, DAG, "WebAssembly hasn't implemented nest arguments");
1362     if (In.Flags.isInConsecutiveRegs())
1363       fail(DL, DAG, "WebAssembly hasn't implemented cons regs arguments");
1364     if (In.Flags.isInConsecutiveRegsLast())
1365       fail(DL, DAG, "WebAssembly hasn't implemented cons regs last arguments");
1366     // Ignore In.getNonZeroOrigAlign() because all our arguments are passed in
1367     // registers.
1368     InVals.push_back(In.Used ? DAG.getNode(WebAssemblyISD::ARGUMENT, DL, In.VT,
1369                                            DAG.getTargetConstant(InVals.size(),
1370                                                                  DL, MVT::i32))
1371                              : DAG.getUNDEF(In.VT));
1372 
1373     // Record the number and types of arguments.
1374     MFI->addParam(In.VT);
1375   }
1376 
1377   // For swiftcc, emit additional swiftself and swifterror arguments
1378   // if there aren't. These additional arguments are also added for callee
1379   // signature They are necessary to match callee and caller signature for
1380   // indirect call.
1381   auto PtrVT = getPointerTy(MF.getDataLayout());
1382   if (CallConv == CallingConv::Swift) {
1383     if (!HasSwiftSelfArg) {
1384       MFI->addParam(PtrVT);
1385     }
1386     if (!HasSwiftErrorArg) {
1387       MFI->addParam(PtrVT);
1388     }
1389   }
1390   // Varargs are copied into a buffer allocated by the caller, and a pointer to
1391   // the buffer is passed as an argument.
1392   if (IsVarArg) {
1393     MVT PtrVT = getPointerTy(MF.getDataLayout());
1394     Register VarargVreg =
1395         MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrVT));
1396     MFI->setVarargBufferVreg(VarargVreg);
1397     Chain = DAG.getCopyToReg(
1398         Chain, DL, VarargVreg,
1399         DAG.getNode(WebAssemblyISD::ARGUMENT, DL, PtrVT,
1400                     DAG.getTargetConstant(Ins.size(), DL, MVT::i32)));
1401     MFI->addParam(PtrVT);
1402   }
1403 
1404   // Record the number and types of arguments and results.
1405   SmallVector<MVT, 4> Params;
1406   SmallVector<MVT, 4> Results;
1407   computeSignatureVTs(MF.getFunction().getFunctionType(), &MF.getFunction(),
1408                       MF.getFunction(), DAG.getTarget(), Params, Results);
1409   for (MVT VT : Results)
1410     MFI->addResult(VT);
1411   // TODO: Use signatures in WebAssemblyMachineFunctionInfo too and unify
1412   // the param logic here with ComputeSignatureVTs
1413   assert(MFI->getParams().size() == Params.size() &&
1414          std::equal(MFI->getParams().begin(), MFI->getParams().end(),
1415                     Params.begin()));
1416 
1417   return Chain;
1418 }
1419 
ReplaceNodeResults(SDNode * N,SmallVectorImpl<SDValue> & Results,SelectionDAG & DAG) const1420 void WebAssemblyTargetLowering::ReplaceNodeResults(
1421     SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
1422   switch (N->getOpcode()) {
1423   case ISD::SIGN_EXTEND_INREG:
1424     // Do not add any results, signifying that N should not be custom lowered
1425     // after all. This happens because simd128 turns on custom lowering for
1426     // SIGN_EXTEND_INREG, but for non-vector sign extends the result might be an
1427     // illegal type.
1428     break;
1429   case ISD::SIGN_EXTEND_VECTOR_INREG:
1430   case ISD::ZERO_EXTEND_VECTOR_INREG:
1431     // Do not add any results, signifying that N should not be custom lowered.
1432     // EXTEND_VECTOR_INREG is implemented for some vectors, but not all.
1433     break;
1434   default:
1435     llvm_unreachable(
1436         "ReplaceNodeResults not implemented for this op for WebAssembly!");
1437   }
1438 }
1439 
1440 //===----------------------------------------------------------------------===//
1441 //  Custom lowering hooks.
1442 //===----------------------------------------------------------------------===//
1443 
LowerOperation(SDValue Op,SelectionDAG & DAG) const1444 SDValue WebAssemblyTargetLowering::LowerOperation(SDValue Op,
1445                                                   SelectionDAG &DAG) const {
1446   SDLoc DL(Op);
1447   switch (Op.getOpcode()) {
1448   default:
1449     llvm_unreachable("unimplemented operation lowering");
1450     return SDValue();
1451   case ISD::FrameIndex:
1452     return LowerFrameIndex(Op, DAG);
1453   case ISD::GlobalAddress:
1454     return LowerGlobalAddress(Op, DAG);
1455   case ISD::GlobalTLSAddress:
1456     return LowerGlobalTLSAddress(Op, DAG);
1457   case ISD::ExternalSymbol:
1458     return LowerExternalSymbol(Op, DAG);
1459   case ISD::JumpTable:
1460     return LowerJumpTable(Op, DAG);
1461   case ISD::BR_JT:
1462     return LowerBR_JT(Op, DAG);
1463   case ISD::VASTART:
1464     return LowerVASTART(Op, DAG);
1465   case ISD::BlockAddress:
1466   case ISD::BRIND:
1467     fail(DL, DAG, "WebAssembly hasn't implemented computed gotos");
1468     return SDValue();
1469   case ISD::RETURNADDR:
1470     return LowerRETURNADDR(Op, DAG);
1471   case ISD::FRAMEADDR:
1472     return LowerFRAMEADDR(Op, DAG);
1473   case ISD::CopyToReg:
1474     return LowerCopyToReg(Op, DAG);
1475   case ISD::EXTRACT_VECTOR_ELT:
1476   case ISD::INSERT_VECTOR_ELT:
1477     return LowerAccessVectorElement(Op, DAG);
1478   case ISD::INTRINSIC_VOID:
1479   case ISD::INTRINSIC_WO_CHAIN:
1480   case ISD::INTRINSIC_W_CHAIN:
1481     return LowerIntrinsic(Op, DAG);
1482   case ISD::SIGN_EXTEND_INREG:
1483     return LowerSIGN_EXTEND_INREG(Op, DAG);
1484   case ISD::ZERO_EXTEND_VECTOR_INREG:
1485   case ISD::SIGN_EXTEND_VECTOR_INREG:
1486     return LowerEXTEND_VECTOR_INREG(Op, DAG);
1487   case ISD::BUILD_VECTOR:
1488     return LowerBUILD_VECTOR(Op, DAG);
1489   case ISD::VECTOR_SHUFFLE:
1490     return LowerVECTOR_SHUFFLE(Op, DAG);
1491   case ISD::SETCC:
1492     return LowerSETCC(Op, DAG);
1493   case ISD::SHL:
1494   case ISD::SRA:
1495   case ISD::SRL:
1496     return LowerShift(Op, DAG);
1497   case ISD::FP_TO_SINT_SAT:
1498   case ISD::FP_TO_UINT_SAT:
1499     return LowerFP_TO_INT_SAT(Op, DAG);
1500   case ISD::LOAD:
1501     return LowerLoad(Op, DAG);
1502   case ISD::STORE:
1503     return LowerStore(Op, DAG);
1504   case ISD::CTPOP:
1505   case ISD::CTLZ:
1506   case ISD::CTTZ:
1507     return DAG.UnrollVectorOp(Op.getNode());
1508   case ISD::CLEAR_CACHE:
1509     report_fatal_error("llvm.clear_cache is not supported on wasm");
1510   }
1511 }
1512 
IsWebAssemblyGlobal(SDValue Op)1513 static bool IsWebAssemblyGlobal(SDValue Op) {
1514   if (const GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Op))
1515     return WebAssembly::isWasmVarAddressSpace(GA->getAddressSpace());
1516 
1517   return false;
1518 }
1519 
IsWebAssemblyLocal(SDValue Op,SelectionDAG & DAG)1520 static std::optional<unsigned> IsWebAssemblyLocal(SDValue Op,
1521                                                   SelectionDAG &DAG) {
1522   const FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(Op);
1523   if (!FI)
1524     return std::nullopt;
1525 
1526   auto &MF = DAG.getMachineFunction();
1527   return WebAssemblyFrameLowering::getLocalForStackObject(MF, FI->getIndex());
1528 }
1529 
LowerStore(SDValue Op,SelectionDAG & DAG) const1530 SDValue WebAssemblyTargetLowering::LowerStore(SDValue Op,
1531                                               SelectionDAG &DAG) const {
1532   SDLoc DL(Op);
1533   StoreSDNode *SN = cast<StoreSDNode>(Op.getNode());
1534   const SDValue &Value = SN->getValue();
1535   const SDValue &Base = SN->getBasePtr();
1536   const SDValue &Offset = SN->getOffset();
1537 
1538   if (IsWebAssemblyGlobal(Base)) {
1539     if (!Offset->isUndef())
1540       report_fatal_error("unexpected offset when storing to webassembly global",
1541                          false);
1542 
1543     SDVTList Tys = DAG.getVTList(MVT::Other);
1544     SDValue Ops[] = {SN->getChain(), Value, Base};
1545     return DAG.getMemIntrinsicNode(WebAssemblyISD::GLOBAL_SET, DL, Tys, Ops,
1546                                    SN->getMemoryVT(), SN->getMemOperand());
1547   }
1548 
1549   if (std::optional<unsigned> Local = IsWebAssemblyLocal(Base, DAG)) {
1550     if (!Offset->isUndef())
1551       report_fatal_error("unexpected offset when storing to webassembly local",
1552                          false);
1553 
1554     SDValue Idx = DAG.getTargetConstant(*Local, Base, MVT::i32);
1555     SDVTList Tys = DAG.getVTList(MVT::Other); // The chain.
1556     SDValue Ops[] = {SN->getChain(), Idx, Value};
1557     return DAG.getNode(WebAssemblyISD::LOCAL_SET, DL, Tys, Ops);
1558   }
1559 
1560   if (WebAssembly::isWasmVarAddressSpace(SN->getAddressSpace()))
1561     report_fatal_error(
1562         "Encountered an unlowerable store to the wasm_var address space",
1563         false);
1564 
1565   return Op;
1566 }
1567 
LowerLoad(SDValue Op,SelectionDAG & DAG) const1568 SDValue WebAssemblyTargetLowering::LowerLoad(SDValue Op,
1569                                              SelectionDAG &DAG) const {
1570   SDLoc DL(Op);
1571   LoadSDNode *LN = cast<LoadSDNode>(Op.getNode());
1572   const SDValue &Base = LN->getBasePtr();
1573   const SDValue &Offset = LN->getOffset();
1574 
1575   if (IsWebAssemblyGlobal(Base)) {
1576     if (!Offset->isUndef())
1577       report_fatal_error(
1578           "unexpected offset when loading from webassembly global", false);
1579 
1580     SDVTList Tys = DAG.getVTList(LN->getValueType(0), MVT::Other);
1581     SDValue Ops[] = {LN->getChain(), Base};
1582     return DAG.getMemIntrinsicNode(WebAssemblyISD::GLOBAL_GET, DL, Tys, Ops,
1583                                    LN->getMemoryVT(), LN->getMemOperand());
1584   }
1585 
1586   if (std::optional<unsigned> Local = IsWebAssemblyLocal(Base, DAG)) {
1587     if (!Offset->isUndef())
1588       report_fatal_error(
1589           "unexpected offset when loading from webassembly local", false);
1590 
1591     SDValue Idx = DAG.getTargetConstant(*Local, Base, MVT::i32);
1592     EVT LocalVT = LN->getValueType(0);
1593     SDValue LocalGet = DAG.getNode(WebAssemblyISD::LOCAL_GET, DL, LocalVT,
1594                                    {LN->getChain(), Idx});
1595     SDValue Result = DAG.getMergeValues({LocalGet, LN->getChain()}, DL);
1596     assert(Result->getNumValues() == 2 && "Loads must carry a chain!");
1597     return Result;
1598   }
1599 
1600   if (WebAssembly::isWasmVarAddressSpace(LN->getAddressSpace()))
1601     report_fatal_error(
1602         "Encountered an unlowerable load from the wasm_var address space",
1603         false);
1604 
1605   return Op;
1606 }
1607 
LowerCopyToReg(SDValue Op,SelectionDAG & DAG) const1608 SDValue WebAssemblyTargetLowering::LowerCopyToReg(SDValue Op,
1609                                                   SelectionDAG &DAG) const {
1610   SDValue Src = Op.getOperand(2);
1611   if (isa<FrameIndexSDNode>(Src.getNode())) {
1612     // CopyToReg nodes don't support FrameIndex operands. Other targets select
1613     // the FI to some LEA-like instruction, but since we don't have that, we
1614     // need to insert some kind of instruction that can take an FI operand and
1615     // produces a value usable by CopyToReg (i.e. in a vreg). So insert a dummy
1616     // local.copy between Op and its FI operand.
1617     SDValue Chain = Op.getOperand(0);
1618     SDLoc DL(Op);
1619     Register Reg = cast<RegisterSDNode>(Op.getOperand(1))->getReg();
1620     EVT VT = Src.getValueType();
1621     SDValue Copy(DAG.getMachineNode(VT == MVT::i32 ? WebAssembly::COPY_I32
1622                                                    : WebAssembly::COPY_I64,
1623                                     DL, VT, Src),
1624                  0);
1625     return Op.getNode()->getNumValues() == 1
1626                ? DAG.getCopyToReg(Chain, DL, Reg, Copy)
1627                : DAG.getCopyToReg(Chain, DL, Reg, Copy,
1628                                   Op.getNumOperands() == 4 ? Op.getOperand(3)
1629                                                            : SDValue());
1630   }
1631   return SDValue();
1632 }
1633 
LowerFrameIndex(SDValue Op,SelectionDAG & DAG) const1634 SDValue WebAssemblyTargetLowering::LowerFrameIndex(SDValue Op,
1635                                                    SelectionDAG &DAG) const {
1636   int FI = cast<FrameIndexSDNode>(Op)->getIndex();
1637   return DAG.getTargetFrameIndex(FI, Op.getValueType());
1638 }
1639 
LowerRETURNADDR(SDValue Op,SelectionDAG & DAG) const1640 SDValue WebAssemblyTargetLowering::LowerRETURNADDR(SDValue Op,
1641                                                    SelectionDAG &DAG) const {
1642   SDLoc DL(Op);
1643 
1644   if (!Subtarget->getTargetTriple().isOSEmscripten()) {
1645     fail(DL, DAG,
1646          "Non-Emscripten WebAssembly hasn't implemented "
1647          "__builtin_return_address");
1648     return SDValue();
1649   }
1650 
1651   if (verifyReturnAddressArgumentIsConstant(Op, DAG))
1652     return SDValue();
1653 
1654   unsigned Depth = Op.getConstantOperandVal(0);
1655   MakeLibCallOptions CallOptions;
1656   return makeLibCall(DAG, RTLIB::RETURN_ADDRESS, Op.getValueType(),
1657                      {DAG.getConstant(Depth, DL, MVT::i32)}, CallOptions, DL)
1658       .first;
1659 }
1660 
LowerFRAMEADDR(SDValue Op,SelectionDAG & DAG) const1661 SDValue WebAssemblyTargetLowering::LowerFRAMEADDR(SDValue Op,
1662                                                   SelectionDAG &DAG) const {
1663   // Non-zero depths are not supported by WebAssembly currently. Use the
1664   // legalizer's default expansion, which is to return 0 (what this function is
1665   // documented to do).
1666   if (Op.getConstantOperandVal(0) > 0)
1667     return SDValue();
1668 
1669   DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true);
1670   EVT VT = Op.getValueType();
1671   Register FP =
1672       Subtarget->getRegisterInfo()->getFrameRegister(DAG.getMachineFunction());
1673   return DAG.getCopyFromReg(DAG.getEntryNode(), SDLoc(Op), FP, VT);
1674 }
1675 
1676 SDValue
LowerGlobalTLSAddress(SDValue Op,SelectionDAG & DAG) const1677 WebAssemblyTargetLowering::LowerGlobalTLSAddress(SDValue Op,
1678                                                  SelectionDAG &DAG) const {
1679   SDLoc DL(Op);
1680   const auto *GA = cast<GlobalAddressSDNode>(Op);
1681 
1682   MachineFunction &MF = DAG.getMachineFunction();
1683   if (!MF.getSubtarget<WebAssemblySubtarget>().hasBulkMemory())
1684     report_fatal_error("cannot use thread-local storage without bulk memory",
1685                        false);
1686 
1687   const GlobalValue *GV = GA->getGlobal();
1688 
1689   // Currently only Emscripten supports dynamic linking with threads. Therefore,
1690   // on other targets, if we have thread-local storage, only the local-exec
1691   // model is possible.
1692   auto model = Subtarget->getTargetTriple().isOSEmscripten()
1693                    ? GV->getThreadLocalMode()
1694                    : GlobalValue::LocalExecTLSModel;
1695 
1696   // Unsupported TLS modes
1697   assert(model != GlobalValue::NotThreadLocal);
1698   assert(model != GlobalValue::InitialExecTLSModel);
1699 
1700   if (model == GlobalValue::LocalExecTLSModel ||
1701       model == GlobalValue::LocalDynamicTLSModel ||
1702       (model == GlobalValue::GeneralDynamicTLSModel &&
1703        getTargetMachine().shouldAssumeDSOLocal(GV))) {
1704     // For DSO-local TLS variables we use offset from __tls_base
1705 
1706     MVT PtrVT = getPointerTy(DAG.getDataLayout());
1707     auto GlobalGet = PtrVT == MVT::i64 ? WebAssembly::GLOBAL_GET_I64
1708                                        : WebAssembly::GLOBAL_GET_I32;
1709     const char *BaseName = MF.createExternalSymbolName("__tls_base");
1710 
1711     SDValue BaseAddr(
1712         DAG.getMachineNode(GlobalGet, DL, PtrVT,
1713                            DAG.getTargetExternalSymbol(BaseName, PtrVT)),
1714         0);
1715 
1716     SDValue TLSOffset = DAG.getTargetGlobalAddress(
1717         GV, DL, PtrVT, GA->getOffset(), WebAssemblyII::MO_TLS_BASE_REL);
1718     SDValue SymOffset =
1719         DAG.getNode(WebAssemblyISD::WrapperREL, DL, PtrVT, TLSOffset);
1720 
1721     return DAG.getNode(ISD::ADD, DL, PtrVT, BaseAddr, SymOffset);
1722   }
1723 
1724   assert(model == GlobalValue::GeneralDynamicTLSModel);
1725 
1726   EVT VT = Op.getValueType();
1727   return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1728                      DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT,
1729                                                 GA->getOffset(),
1730                                                 WebAssemblyII::MO_GOT_TLS));
1731 }
1732 
LowerGlobalAddress(SDValue Op,SelectionDAG & DAG) const1733 SDValue WebAssemblyTargetLowering::LowerGlobalAddress(SDValue Op,
1734                                                       SelectionDAG &DAG) const {
1735   SDLoc DL(Op);
1736   const auto *GA = cast<GlobalAddressSDNode>(Op);
1737   EVT VT = Op.getValueType();
1738   assert(GA->getTargetFlags() == 0 &&
1739          "Unexpected target flags on generic GlobalAddressSDNode");
1740   if (!WebAssembly::isValidAddressSpace(GA->getAddressSpace()))
1741     fail(DL, DAG, "Invalid address space for WebAssembly target");
1742 
1743   unsigned OperandFlags = 0;
1744   const GlobalValue *GV = GA->getGlobal();
1745   // Since WebAssembly tables cannot yet be shared accross modules, we don't
1746   // need special treatment for tables in PIC mode.
1747   if (isPositionIndependent() &&
1748       !WebAssembly::isWebAssemblyTableType(GV->getValueType())) {
1749     if (getTargetMachine().shouldAssumeDSOLocal(GV)) {
1750       MachineFunction &MF = DAG.getMachineFunction();
1751       MVT PtrVT = getPointerTy(MF.getDataLayout());
1752       const char *BaseName;
1753       if (GV->getValueType()->isFunctionTy()) {
1754         BaseName = MF.createExternalSymbolName("__table_base");
1755         OperandFlags = WebAssemblyII::MO_TABLE_BASE_REL;
1756       } else {
1757         BaseName = MF.createExternalSymbolName("__memory_base");
1758         OperandFlags = WebAssemblyII::MO_MEMORY_BASE_REL;
1759       }
1760       SDValue BaseAddr =
1761           DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
1762                       DAG.getTargetExternalSymbol(BaseName, PtrVT));
1763 
1764       SDValue SymAddr = DAG.getNode(
1765           WebAssemblyISD::WrapperREL, DL, VT,
1766           DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT, GA->getOffset(),
1767                                      OperandFlags));
1768 
1769       return DAG.getNode(ISD::ADD, DL, VT, BaseAddr, SymAddr);
1770     }
1771     OperandFlags = WebAssemblyII::MO_GOT;
1772   }
1773 
1774   return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1775                      DAG.getTargetGlobalAddress(GA->getGlobal(), DL, VT,
1776                                                 GA->getOffset(), OperandFlags));
1777 }
1778 
1779 SDValue
LowerExternalSymbol(SDValue Op,SelectionDAG & DAG) const1780 WebAssemblyTargetLowering::LowerExternalSymbol(SDValue Op,
1781                                                SelectionDAG &DAG) const {
1782   SDLoc DL(Op);
1783   const auto *ES = cast<ExternalSymbolSDNode>(Op);
1784   EVT VT = Op.getValueType();
1785   assert(ES->getTargetFlags() == 0 &&
1786          "Unexpected target flags on generic ExternalSymbolSDNode");
1787   return DAG.getNode(WebAssemblyISD::Wrapper, DL, VT,
1788                      DAG.getTargetExternalSymbol(ES->getSymbol(), VT));
1789 }
1790 
LowerJumpTable(SDValue Op,SelectionDAG & DAG) const1791 SDValue WebAssemblyTargetLowering::LowerJumpTable(SDValue Op,
1792                                                   SelectionDAG &DAG) const {
1793   // There's no need for a Wrapper node because we always incorporate a jump
1794   // table operand into a BR_TABLE instruction, rather than ever
1795   // materializing it in a register.
1796   const JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
1797   return DAG.getTargetJumpTable(JT->getIndex(), Op.getValueType(),
1798                                 JT->getTargetFlags());
1799 }
1800 
LowerBR_JT(SDValue Op,SelectionDAG & DAG) const1801 SDValue WebAssemblyTargetLowering::LowerBR_JT(SDValue Op,
1802                                               SelectionDAG &DAG) const {
1803   SDLoc DL(Op);
1804   SDValue Chain = Op.getOperand(0);
1805   const auto *JT = cast<JumpTableSDNode>(Op.getOperand(1));
1806   SDValue Index = Op.getOperand(2);
1807   assert(JT->getTargetFlags() == 0 && "WebAssembly doesn't set target flags");
1808 
1809   SmallVector<SDValue, 8> Ops;
1810   Ops.push_back(Chain);
1811   Ops.push_back(Index);
1812 
1813   MachineJumpTableInfo *MJTI = DAG.getMachineFunction().getJumpTableInfo();
1814   const auto &MBBs = MJTI->getJumpTables()[JT->getIndex()].MBBs;
1815 
1816   // Add an operand for each case.
1817   for (auto *MBB : MBBs)
1818     Ops.push_back(DAG.getBasicBlock(MBB));
1819 
1820   // Add the first MBB as a dummy default target for now. This will be replaced
1821   // with the proper default target (and the preceding range check eliminated)
1822   // if possible by WebAssemblyFixBrTableDefaults.
1823   Ops.push_back(DAG.getBasicBlock(*MBBs.begin()));
1824   return DAG.getNode(WebAssemblyISD::BR_TABLE, DL, MVT::Other, Ops);
1825 }
1826 
LowerVASTART(SDValue Op,SelectionDAG & DAG) const1827 SDValue WebAssemblyTargetLowering::LowerVASTART(SDValue Op,
1828                                                 SelectionDAG &DAG) const {
1829   SDLoc DL(Op);
1830   EVT PtrVT = getPointerTy(DAG.getMachineFunction().getDataLayout());
1831 
1832   auto *MFI = DAG.getMachineFunction().getInfo<WebAssemblyFunctionInfo>();
1833   const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
1834 
1835   SDValue ArgN = DAG.getCopyFromReg(DAG.getEntryNode(), DL,
1836                                     MFI->getVarargBufferVreg(), PtrVT);
1837   return DAG.getStore(Op.getOperand(0), DL, ArgN, Op.getOperand(1),
1838                       MachinePointerInfo(SV));
1839 }
1840 
LowerIntrinsic(SDValue Op,SelectionDAG & DAG) const1841 SDValue WebAssemblyTargetLowering::LowerIntrinsic(SDValue Op,
1842                                                   SelectionDAG &DAG) const {
1843   MachineFunction &MF = DAG.getMachineFunction();
1844   unsigned IntNo;
1845   switch (Op.getOpcode()) {
1846   case ISD::INTRINSIC_VOID:
1847   case ISD::INTRINSIC_W_CHAIN:
1848     IntNo = Op.getConstantOperandVal(1);
1849     break;
1850   case ISD::INTRINSIC_WO_CHAIN:
1851     IntNo = Op.getConstantOperandVal(0);
1852     break;
1853   default:
1854     llvm_unreachable("Invalid intrinsic");
1855   }
1856   SDLoc DL(Op);
1857 
1858   switch (IntNo) {
1859   default:
1860     return SDValue(); // Don't custom lower most intrinsics.
1861 
1862   case Intrinsic::wasm_lsda: {
1863     auto PtrVT = getPointerTy(MF.getDataLayout());
1864     const char *SymName = MF.createExternalSymbolName(
1865         "GCC_except_table" + std::to_string(MF.getFunctionNumber()));
1866     if (isPositionIndependent()) {
1867       SDValue Node = DAG.getTargetExternalSymbol(
1868           SymName, PtrVT, WebAssemblyII::MO_MEMORY_BASE_REL);
1869       const char *BaseName = MF.createExternalSymbolName("__memory_base");
1870       SDValue BaseAddr =
1871           DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT,
1872                       DAG.getTargetExternalSymbol(BaseName, PtrVT));
1873       SDValue SymAddr =
1874           DAG.getNode(WebAssemblyISD::WrapperREL, DL, PtrVT, Node);
1875       return DAG.getNode(ISD::ADD, DL, PtrVT, BaseAddr, SymAddr);
1876     }
1877     SDValue Node = DAG.getTargetExternalSymbol(SymName, PtrVT);
1878     return DAG.getNode(WebAssemblyISD::Wrapper, DL, PtrVT, Node);
1879   }
1880 
1881   case Intrinsic::wasm_shuffle: {
1882     // Drop in-chain and replace undefs, but otherwise pass through unchanged
1883     SDValue Ops[18];
1884     size_t OpIdx = 0;
1885     Ops[OpIdx++] = Op.getOperand(1);
1886     Ops[OpIdx++] = Op.getOperand(2);
1887     while (OpIdx < 18) {
1888       const SDValue &MaskIdx = Op.getOperand(OpIdx + 1);
1889       if (MaskIdx.isUndef() || MaskIdx.getNode()->getAsZExtVal() >= 32) {
1890         bool isTarget = MaskIdx.getNode()->getOpcode() == ISD::TargetConstant;
1891         Ops[OpIdx++] = DAG.getConstant(0, DL, MVT::i32, isTarget);
1892       } else {
1893         Ops[OpIdx++] = MaskIdx;
1894       }
1895     }
1896     return DAG.getNode(WebAssemblyISD::SHUFFLE, DL, Op.getValueType(), Ops);
1897   }
1898   }
1899 }
1900 
1901 SDValue
LowerSIGN_EXTEND_INREG(SDValue Op,SelectionDAG & DAG) const1902 WebAssemblyTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1903                                                   SelectionDAG &DAG) const {
1904   SDLoc DL(Op);
1905   // If sign extension operations are disabled, allow sext_inreg only if operand
1906   // is a vector extract of an i8 or i16 lane. SIMD does not depend on sign
1907   // extension operations, but allowing sext_inreg in this context lets us have
1908   // simple patterns to select extract_lane_s instructions. Expanding sext_inreg
1909   // everywhere would be simpler in this file, but would necessitate large and
1910   // brittle patterns to undo the expansion and select extract_lane_s
1911   // instructions.
1912   assert(!Subtarget->hasSignExt() && Subtarget->hasSIMD128());
1913   if (Op.getOperand(0).getOpcode() != ISD::EXTRACT_VECTOR_ELT)
1914     return SDValue();
1915 
1916   const SDValue &Extract = Op.getOperand(0);
1917   MVT VecT = Extract.getOperand(0).getSimpleValueType();
1918   if (VecT.getVectorElementType().getSizeInBits() > 32)
1919     return SDValue();
1920   MVT ExtractedLaneT =
1921       cast<VTSDNode>(Op.getOperand(1).getNode())->getVT().getSimpleVT();
1922   MVT ExtractedVecT =
1923       MVT::getVectorVT(ExtractedLaneT, 128 / ExtractedLaneT.getSizeInBits());
1924   if (ExtractedVecT == VecT)
1925     return Op;
1926 
1927   // Bitcast vector to appropriate type to ensure ISel pattern coverage
1928   const SDNode *Index = Extract.getOperand(1).getNode();
1929   if (!isa<ConstantSDNode>(Index))
1930     return SDValue();
1931   unsigned IndexVal = Index->getAsZExtVal();
1932   unsigned Scale =
1933       ExtractedVecT.getVectorNumElements() / VecT.getVectorNumElements();
1934   assert(Scale > 1);
1935   SDValue NewIndex =
1936       DAG.getConstant(IndexVal * Scale, DL, Index->getValueType(0));
1937   SDValue NewExtract = DAG.getNode(
1938       ISD::EXTRACT_VECTOR_ELT, DL, Extract.getValueType(),
1939       DAG.getBitcast(ExtractedVecT, Extract.getOperand(0)), NewIndex);
1940   return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, Op.getValueType(), NewExtract,
1941                      Op.getOperand(1));
1942 }
1943 
1944 SDValue
LowerEXTEND_VECTOR_INREG(SDValue Op,SelectionDAG & DAG) const1945 WebAssemblyTargetLowering::LowerEXTEND_VECTOR_INREG(SDValue Op,
1946                                                     SelectionDAG &DAG) const {
1947   SDLoc DL(Op);
1948   EVT VT = Op.getValueType();
1949   SDValue Src = Op.getOperand(0);
1950   EVT SrcVT = Src.getValueType();
1951 
1952   if (SrcVT.getVectorElementType() == MVT::i1 ||
1953       SrcVT.getVectorElementType() == MVT::i64)
1954     return SDValue();
1955 
1956   assert(VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits() == 0 &&
1957          "Unexpected extension factor.");
1958   unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits();
1959 
1960   if (Scale != 2 && Scale != 4 && Scale != 8)
1961     return SDValue();
1962 
1963   unsigned Ext;
1964   switch (Op.getOpcode()) {
1965   case ISD::ZERO_EXTEND_VECTOR_INREG:
1966     Ext = WebAssemblyISD::EXTEND_LOW_U;
1967     break;
1968   case ISD::SIGN_EXTEND_VECTOR_INREG:
1969     Ext = WebAssemblyISD::EXTEND_LOW_S;
1970     break;
1971   }
1972 
1973   SDValue Ret = Src;
1974   while (Scale != 1) {
1975     Ret = DAG.getNode(Ext, DL,
1976                       Ret.getValueType()
1977                           .widenIntegerVectorElementType(*DAG.getContext())
1978                           .getHalfNumVectorElementsVT(*DAG.getContext()),
1979                       Ret);
1980     Scale /= 2;
1981   }
1982   assert(Ret.getValueType() == VT);
1983   return Ret;
1984 }
1985 
LowerConvertLow(SDValue Op,SelectionDAG & DAG)1986 static SDValue LowerConvertLow(SDValue Op, SelectionDAG &DAG) {
1987   SDLoc DL(Op);
1988   if (Op.getValueType() != MVT::v2f64)
1989     return SDValue();
1990 
1991   auto GetConvertedLane = [](SDValue Op, unsigned &Opcode, SDValue &SrcVec,
1992                              unsigned &Index) -> bool {
1993     switch (Op.getOpcode()) {
1994     case ISD::SINT_TO_FP:
1995       Opcode = WebAssemblyISD::CONVERT_LOW_S;
1996       break;
1997     case ISD::UINT_TO_FP:
1998       Opcode = WebAssemblyISD::CONVERT_LOW_U;
1999       break;
2000     case ISD::FP_EXTEND:
2001       Opcode = WebAssemblyISD::PROMOTE_LOW;
2002       break;
2003     default:
2004       return false;
2005     }
2006 
2007     auto ExtractVector = Op.getOperand(0);
2008     if (ExtractVector.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2009       return false;
2010 
2011     if (!isa<ConstantSDNode>(ExtractVector.getOperand(1).getNode()))
2012       return false;
2013 
2014     SrcVec = ExtractVector.getOperand(0);
2015     Index = ExtractVector.getConstantOperandVal(1);
2016     return true;
2017   };
2018 
2019   unsigned LHSOpcode, RHSOpcode, LHSIndex, RHSIndex;
2020   SDValue LHSSrcVec, RHSSrcVec;
2021   if (!GetConvertedLane(Op.getOperand(0), LHSOpcode, LHSSrcVec, LHSIndex) ||
2022       !GetConvertedLane(Op.getOperand(1), RHSOpcode, RHSSrcVec, RHSIndex))
2023     return SDValue();
2024 
2025   if (LHSOpcode != RHSOpcode)
2026     return SDValue();
2027 
2028   MVT ExpectedSrcVT;
2029   switch (LHSOpcode) {
2030   case WebAssemblyISD::CONVERT_LOW_S:
2031   case WebAssemblyISD::CONVERT_LOW_U:
2032     ExpectedSrcVT = MVT::v4i32;
2033     break;
2034   case WebAssemblyISD::PROMOTE_LOW:
2035     ExpectedSrcVT = MVT::v4f32;
2036     break;
2037   }
2038   if (LHSSrcVec.getValueType() != ExpectedSrcVT)
2039     return SDValue();
2040 
2041   auto Src = LHSSrcVec;
2042   if (LHSIndex != 0 || RHSIndex != 1 || LHSSrcVec != RHSSrcVec) {
2043     // Shuffle the source vector so that the converted lanes are the low lanes.
2044     Src = DAG.getVectorShuffle(
2045         ExpectedSrcVT, DL, LHSSrcVec, RHSSrcVec,
2046         {static_cast<int>(LHSIndex), static_cast<int>(RHSIndex) + 4, -1, -1});
2047   }
2048   return DAG.getNode(LHSOpcode, DL, MVT::v2f64, Src);
2049 }
2050 
LowerBUILD_VECTOR(SDValue Op,SelectionDAG & DAG) const2051 SDValue WebAssemblyTargetLowering::LowerBUILD_VECTOR(SDValue Op,
2052                                                      SelectionDAG &DAG) const {
2053   if (auto ConvertLow = LowerConvertLow(Op, DAG))
2054     return ConvertLow;
2055 
2056   SDLoc DL(Op);
2057   const EVT VecT = Op.getValueType();
2058   const EVT LaneT = Op.getOperand(0).getValueType();
2059   const size_t Lanes = Op.getNumOperands();
2060   bool CanSwizzle = VecT == MVT::v16i8;
2061 
2062   // BUILD_VECTORs are lowered to the instruction that initializes the highest
2063   // possible number of lanes at once followed by a sequence of replace_lane
2064   // instructions to individually initialize any remaining lanes.
2065 
2066   // TODO: Tune this. For example, lanewise swizzling is very expensive, so
2067   // swizzled lanes should be given greater weight.
2068 
2069   // TODO: Investigate looping rather than always extracting/replacing specific
2070   // lanes to fill gaps.
2071 
2072   auto IsConstant = [](const SDValue &V) {
2073     return V.getOpcode() == ISD::Constant || V.getOpcode() == ISD::ConstantFP;
2074   };
2075 
2076   // Returns the source vector and index vector pair if they exist. Checks for:
2077   //   (extract_vector_elt
2078   //     $src,
2079   //     (sign_extend_inreg (extract_vector_elt $indices, $i))
2080   //   )
2081   auto GetSwizzleSrcs = [](size_t I, const SDValue &Lane) {
2082     auto Bail = std::make_pair(SDValue(), SDValue());
2083     if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2084       return Bail;
2085     const SDValue &SwizzleSrc = Lane->getOperand(0);
2086     const SDValue &IndexExt = Lane->getOperand(1);
2087     if (IndexExt->getOpcode() != ISD::SIGN_EXTEND_INREG)
2088       return Bail;
2089     const SDValue &Index = IndexExt->getOperand(0);
2090     if (Index->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2091       return Bail;
2092     const SDValue &SwizzleIndices = Index->getOperand(0);
2093     if (SwizzleSrc.getValueType() != MVT::v16i8 ||
2094         SwizzleIndices.getValueType() != MVT::v16i8 ||
2095         Index->getOperand(1)->getOpcode() != ISD::Constant ||
2096         Index->getConstantOperandVal(1) != I)
2097       return Bail;
2098     return std::make_pair(SwizzleSrc, SwizzleIndices);
2099   };
2100 
2101   // If the lane is extracted from another vector at a constant index, return
2102   // that vector. The source vector must not have more lanes than the dest
2103   // because the shufflevector indices are in terms of the destination lanes and
2104   // would not be able to address the smaller individual source lanes.
2105   auto GetShuffleSrc = [&](const SDValue &Lane) {
2106     if (Lane->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
2107       return SDValue();
2108     if (!isa<ConstantSDNode>(Lane->getOperand(1).getNode()))
2109       return SDValue();
2110     if (Lane->getOperand(0).getValueType().getVectorNumElements() >
2111         VecT.getVectorNumElements())
2112       return SDValue();
2113     return Lane->getOperand(0);
2114   };
2115 
2116   using ValueEntry = std::pair<SDValue, size_t>;
2117   SmallVector<ValueEntry, 16> SplatValueCounts;
2118 
2119   using SwizzleEntry = std::pair<std::pair<SDValue, SDValue>, size_t>;
2120   SmallVector<SwizzleEntry, 16> SwizzleCounts;
2121 
2122   using ShuffleEntry = std::pair<SDValue, size_t>;
2123   SmallVector<ShuffleEntry, 16> ShuffleCounts;
2124 
2125   auto AddCount = [](auto &Counts, const auto &Val) {
2126     auto CountIt =
2127         llvm::find_if(Counts, [&Val](auto E) { return E.first == Val; });
2128     if (CountIt == Counts.end()) {
2129       Counts.emplace_back(Val, 1);
2130     } else {
2131       CountIt->second++;
2132     }
2133   };
2134 
2135   auto GetMostCommon = [](auto &Counts) {
2136     auto CommonIt =
2137         std::max_element(Counts.begin(), Counts.end(), llvm::less_second());
2138     assert(CommonIt != Counts.end() && "Unexpected all-undef build_vector");
2139     return *CommonIt;
2140   };
2141 
2142   size_t NumConstantLanes = 0;
2143 
2144   // Count eligible lanes for each type of vector creation op
2145   for (size_t I = 0; I < Lanes; ++I) {
2146     const SDValue &Lane = Op->getOperand(I);
2147     if (Lane.isUndef())
2148       continue;
2149 
2150     AddCount(SplatValueCounts, Lane);
2151 
2152     if (IsConstant(Lane))
2153       NumConstantLanes++;
2154     if (auto ShuffleSrc = GetShuffleSrc(Lane))
2155       AddCount(ShuffleCounts, ShuffleSrc);
2156     if (CanSwizzle) {
2157       auto SwizzleSrcs = GetSwizzleSrcs(I, Lane);
2158       if (SwizzleSrcs.first)
2159         AddCount(SwizzleCounts, SwizzleSrcs);
2160     }
2161   }
2162 
2163   SDValue SplatValue;
2164   size_t NumSplatLanes;
2165   std::tie(SplatValue, NumSplatLanes) = GetMostCommon(SplatValueCounts);
2166 
2167   SDValue SwizzleSrc;
2168   SDValue SwizzleIndices;
2169   size_t NumSwizzleLanes = 0;
2170   if (SwizzleCounts.size())
2171     std::forward_as_tuple(std::tie(SwizzleSrc, SwizzleIndices),
2172                           NumSwizzleLanes) = GetMostCommon(SwizzleCounts);
2173 
2174   // Shuffles can draw from up to two vectors, so find the two most common
2175   // sources.
2176   SDValue ShuffleSrc1, ShuffleSrc2;
2177   size_t NumShuffleLanes = 0;
2178   if (ShuffleCounts.size()) {
2179     std::tie(ShuffleSrc1, NumShuffleLanes) = GetMostCommon(ShuffleCounts);
2180     llvm::erase_if(ShuffleCounts,
2181                    [&](const auto &Pair) { return Pair.first == ShuffleSrc1; });
2182   }
2183   if (ShuffleCounts.size()) {
2184     size_t AdditionalShuffleLanes;
2185     std::tie(ShuffleSrc2, AdditionalShuffleLanes) =
2186         GetMostCommon(ShuffleCounts);
2187     NumShuffleLanes += AdditionalShuffleLanes;
2188   }
2189 
2190   // Predicate returning true if the lane is properly initialized by the
2191   // original instruction
2192   std::function<bool(size_t, const SDValue &)> IsLaneConstructed;
2193   SDValue Result;
2194   // Prefer swizzles over shuffles over vector consts over splats
2195   if (NumSwizzleLanes >= NumShuffleLanes &&
2196       NumSwizzleLanes >= NumConstantLanes && NumSwizzleLanes >= NumSplatLanes) {
2197     Result = DAG.getNode(WebAssemblyISD::SWIZZLE, DL, VecT, SwizzleSrc,
2198                          SwizzleIndices);
2199     auto Swizzled = std::make_pair(SwizzleSrc, SwizzleIndices);
2200     IsLaneConstructed = [&, Swizzled](size_t I, const SDValue &Lane) {
2201       return Swizzled == GetSwizzleSrcs(I, Lane);
2202     };
2203   } else if (NumShuffleLanes >= NumConstantLanes &&
2204              NumShuffleLanes >= NumSplatLanes) {
2205     size_t DestLaneSize = VecT.getVectorElementType().getFixedSizeInBits() / 8;
2206     size_t DestLaneCount = VecT.getVectorNumElements();
2207     size_t Scale1 = 1;
2208     size_t Scale2 = 1;
2209     SDValue Src1 = ShuffleSrc1;
2210     SDValue Src2 = ShuffleSrc2 ? ShuffleSrc2 : DAG.getUNDEF(VecT);
2211     if (Src1.getValueType() != VecT) {
2212       size_t LaneSize =
2213           Src1.getValueType().getVectorElementType().getFixedSizeInBits() / 8;
2214       assert(LaneSize > DestLaneSize);
2215       Scale1 = LaneSize / DestLaneSize;
2216       Src1 = DAG.getBitcast(VecT, Src1);
2217     }
2218     if (Src2.getValueType() != VecT) {
2219       size_t LaneSize =
2220           Src2.getValueType().getVectorElementType().getFixedSizeInBits() / 8;
2221       assert(LaneSize > DestLaneSize);
2222       Scale2 = LaneSize / DestLaneSize;
2223       Src2 = DAG.getBitcast(VecT, Src2);
2224     }
2225 
2226     int Mask[16];
2227     assert(DestLaneCount <= 16);
2228     for (size_t I = 0; I < DestLaneCount; ++I) {
2229       const SDValue &Lane = Op->getOperand(I);
2230       SDValue Src = GetShuffleSrc(Lane);
2231       if (Src == ShuffleSrc1) {
2232         Mask[I] = Lane->getConstantOperandVal(1) * Scale1;
2233       } else if (Src && Src == ShuffleSrc2) {
2234         Mask[I] = DestLaneCount + Lane->getConstantOperandVal(1) * Scale2;
2235       } else {
2236         Mask[I] = -1;
2237       }
2238     }
2239     ArrayRef<int> MaskRef(Mask, DestLaneCount);
2240     Result = DAG.getVectorShuffle(VecT, DL, Src1, Src2, MaskRef);
2241     IsLaneConstructed = [&](size_t, const SDValue &Lane) {
2242       auto Src = GetShuffleSrc(Lane);
2243       return Src == ShuffleSrc1 || (Src && Src == ShuffleSrc2);
2244     };
2245   } else if (NumConstantLanes >= NumSplatLanes) {
2246     SmallVector<SDValue, 16> ConstLanes;
2247     for (const SDValue &Lane : Op->op_values()) {
2248       if (IsConstant(Lane)) {
2249         // Values may need to be fixed so that they will sign extend to be
2250         // within the expected range during ISel. Check whether the value is in
2251         // bounds based on the lane bit width and if it is out of bounds, lop
2252         // off the extra bits and subtract 2^n to reflect giving the high bit
2253         // value -2^(n-1) rather than +2^(n-1). Skip the i64 case because it
2254         // cannot possibly be out of range.
2255         auto *Const = dyn_cast<ConstantSDNode>(Lane.getNode());
2256         int64_t Val = Const ? Const->getSExtValue() : 0;
2257         uint64_t LaneBits = 128 / Lanes;
2258         assert((LaneBits == 64 || Val >= -(1ll << (LaneBits - 1))) &&
2259                "Unexpected out of bounds negative value");
2260         if (Const && LaneBits != 64 && Val > (1ll << (LaneBits - 1)) - 1) {
2261           uint64_t Mask = (1ll << LaneBits) - 1;
2262           auto NewVal = (((uint64_t)Val & Mask) - (1ll << LaneBits)) & Mask;
2263           ConstLanes.push_back(DAG.getConstant(NewVal, SDLoc(Lane), LaneT));
2264         } else {
2265           ConstLanes.push_back(Lane);
2266         }
2267       } else if (LaneT.isFloatingPoint()) {
2268         ConstLanes.push_back(DAG.getConstantFP(0, DL, LaneT));
2269       } else {
2270         ConstLanes.push_back(DAG.getConstant(0, DL, LaneT));
2271       }
2272     }
2273     Result = DAG.getBuildVector(VecT, DL, ConstLanes);
2274     IsLaneConstructed = [&IsConstant](size_t _, const SDValue &Lane) {
2275       return IsConstant(Lane);
2276     };
2277   } else {
2278     // Use a splat (which might be selected as a load splat)
2279     Result = DAG.getSplatBuildVector(VecT, DL, SplatValue);
2280     IsLaneConstructed = [&SplatValue](size_t _, const SDValue &Lane) {
2281       return Lane == SplatValue;
2282     };
2283   }
2284 
2285   assert(Result);
2286   assert(IsLaneConstructed);
2287 
2288   // Add replace_lane instructions for any unhandled values
2289   for (size_t I = 0; I < Lanes; ++I) {
2290     const SDValue &Lane = Op->getOperand(I);
2291     if (!Lane.isUndef() && !IsLaneConstructed(I, Lane))
2292       Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecT, Result, Lane,
2293                            DAG.getConstant(I, DL, MVT::i32));
2294   }
2295 
2296   return Result;
2297 }
2298 
2299 SDValue
LowerVECTOR_SHUFFLE(SDValue Op,SelectionDAG & DAG) const2300 WebAssemblyTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
2301                                                SelectionDAG &DAG) const {
2302   SDLoc DL(Op);
2303   ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op.getNode())->getMask();
2304   MVT VecType = Op.getOperand(0).getSimpleValueType();
2305   assert(VecType.is128BitVector() && "Unexpected shuffle vector type");
2306   size_t LaneBytes = VecType.getVectorElementType().getSizeInBits() / 8;
2307 
2308   // Space for two vector args and sixteen mask indices
2309   SDValue Ops[18];
2310   size_t OpIdx = 0;
2311   Ops[OpIdx++] = Op.getOperand(0);
2312   Ops[OpIdx++] = Op.getOperand(1);
2313 
2314   // Expand mask indices to byte indices and materialize them as operands
2315   for (int M : Mask) {
2316     for (size_t J = 0; J < LaneBytes; ++J) {
2317       // Lower undefs (represented by -1 in mask) to {0..J}, which use a
2318       // whole lane of vector input, to allow further reduction at VM. E.g.
2319       // match an 8x16 byte shuffle to an equivalent cheaper 32x4 shuffle.
2320       uint64_t ByteIndex = M == -1 ? J : (uint64_t)M * LaneBytes + J;
2321       Ops[OpIdx++] = DAG.getConstant(ByteIndex, DL, MVT::i32);
2322     }
2323   }
2324 
2325   return DAG.getNode(WebAssemblyISD::SHUFFLE, DL, Op.getValueType(), Ops);
2326 }
2327 
LowerSETCC(SDValue Op,SelectionDAG & DAG) const2328 SDValue WebAssemblyTargetLowering::LowerSETCC(SDValue Op,
2329                                               SelectionDAG &DAG) const {
2330   SDLoc DL(Op);
2331   // The legalizer does not know how to expand the unsupported comparison modes
2332   // of i64x2 vectors, so we manually unroll them here.
2333   assert(Op->getOperand(0)->getSimpleValueType(0) == MVT::v2i64);
2334   SmallVector<SDValue, 2> LHS, RHS;
2335   DAG.ExtractVectorElements(Op->getOperand(0), LHS);
2336   DAG.ExtractVectorElements(Op->getOperand(1), RHS);
2337   const SDValue &CC = Op->getOperand(2);
2338   auto MakeLane = [&](unsigned I) {
2339     return DAG.getNode(ISD::SELECT_CC, DL, MVT::i64, LHS[I], RHS[I],
2340                        DAG.getConstant(uint64_t(-1), DL, MVT::i64),
2341                        DAG.getConstant(uint64_t(0), DL, MVT::i64), CC);
2342   };
2343   return DAG.getBuildVector(Op->getValueType(0), DL,
2344                             {MakeLane(0), MakeLane(1)});
2345 }
2346 
2347 SDValue
LowerAccessVectorElement(SDValue Op,SelectionDAG & DAG) const2348 WebAssemblyTargetLowering::LowerAccessVectorElement(SDValue Op,
2349                                                     SelectionDAG &DAG) const {
2350   // Allow constant lane indices, expand variable lane indices
2351   SDNode *IdxNode = Op.getOperand(Op.getNumOperands() - 1).getNode();
2352   if (isa<ConstantSDNode>(IdxNode)) {
2353     // Ensure the index type is i32 to match the tablegen patterns
2354     uint64_t Idx = IdxNode->getAsZExtVal();
2355     SmallVector<SDValue, 3> Ops(Op.getNode()->ops());
2356     Ops[Op.getNumOperands() - 1] =
2357         DAG.getConstant(Idx, SDLoc(IdxNode), MVT::i32);
2358     return DAG.getNode(Op.getOpcode(), SDLoc(Op), Op.getValueType(), Ops);
2359   }
2360   // Perform default expansion
2361   return SDValue();
2362 }
2363 
unrollVectorShift(SDValue Op,SelectionDAG & DAG)2364 static SDValue unrollVectorShift(SDValue Op, SelectionDAG &DAG) {
2365   EVT LaneT = Op.getSimpleValueType().getVectorElementType();
2366   // 32-bit and 64-bit unrolled shifts will have proper semantics
2367   if (LaneT.bitsGE(MVT::i32))
2368     return DAG.UnrollVectorOp(Op.getNode());
2369   // Otherwise mask the shift value to get proper semantics from 32-bit shift
2370   SDLoc DL(Op);
2371   size_t NumLanes = Op.getSimpleValueType().getVectorNumElements();
2372   SDValue Mask = DAG.getConstant(LaneT.getSizeInBits() - 1, DL, MVT::i32);
2373   unsigned ShiftOpcode = Op.getOpcode();
2374   SmallVector<SDValue, 16> ShiftedElements;
2375   DAG.ExtractVectorElements(Op.getOperand(0), ShiftedElements, 0, 0, MVT::i32);
2376   SmallVector<SDValue, 16> ShiftElements;
2377   DAG.ExtractVectorElements(Op.getOperand(1), ShiftElements, 0, 0, MVT::i32);
2378   SmallVector<SDValue, 16> UnrolledOps;
2379   for (size_t i = 0; i < NumLanes; ++i) {
2380     SDValue MaskedShiftValue =
2381         DAG.getNode(ISD::AND, DL, MVT::i32, ShiftElements[i], Mask);
2382     SDValue ShiftedValue = ShiftedElements[i];
2383     if (ShiftOpcode == ISD::SRA)
2384       ShiftedValue = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32,
2385                                  ShiftedValue, DAG.getValueType(LaneT));
2386     UnrolledOps.push_back(
2387         DAG.getNode(ShiftOpcode, DL, MVT::i32, ShiftedValue, MaskedShiftValue));
2388   }
2389   return DAG.getBuildVector(Op.getValueType(), DL, UnrolledOps);
2390 }
2391 
LowerShift(SDValue Op,SelectionDAG & DAG) const2392 SDValue WebAssemblyTargetLowering::LowerShift(SDValue Op,
2393                                               SelectionDAG &DAG) const {
2394   SDLoc DL(Op);
2395 
2396   // Only manually lower vector shifts
2397   assert(Op.getSimpleValueType().isVector());
2398 
2399   uint64_t LaneBits = Op.getValueType().getScalarSizeInBits();
2400   auto ShiftVal = Op.getOperand(1);
2401 
2402   // Try to skip bitmask operation since it is implied inside shift instruction
2403   auto SkipImpliedMask = [](SDValue MaskOp, uint64_t MaskBits) {
2404     if (MaskOp.getOpcode() != ISD::AND)
2405       return MaskOp;
2406     SDValue LHS = MaskOp.getOperand(0);
2407     SDValue RHS = MaskOp.getOperand(1);
2408     if (MaskOp.getValueType().isVector()) {
2409       APInt MaskVal;
2410       if (!ISD::isConstantSplatVector(RHS.getNode(), MaskVal))
2411         std::swap(LHS, RHS);
2412 
2413       if (ISD::isConstantSplatVector(RHS.getNode(), MaskVal) &&
2414           MaskVal == MaskBits)
2415         MaskOp = LHS;
2416     } else {
2417       if (!isa<ConstantSDNode>(RHS.getNode()))
2418         std::swap(LHS, RHS);
2419 
2420       auto ConstantRHS = dyn_cast<ConstantSDNode>(RHS.getNode());
2421       if (ConstantRHS && ConstantRHS->getAPIntValue() == MaskBits)
2422         MaskOp = LHS;
2423     }
2424 
2425     return MaskOp;
2426   };
2427 
2428   // Skip vector and operation
2429   ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1);
2430   ShiftVal = DAG.getSplatValue(ShiftVal);
2431   if (!ShiftVal)
2432     return unrollVectorShift(Op, DAG);
2433 
2434   // Skip scalar and operation
2435   ShiftVal = SkipImpliedMask(ShiftVal, LaneBits - 1);
2436   // Use anyext because none of the high bits can affect the shift
2437   ShiftVal = DAG.getAnyExtOrTrunc(ShiftVal, DL, MVT::i32);
2438 
2439   unsigned Opcode;
2440   switch (Op.getOpcode()) {
2441   case ISD::SHL:
2442     Opcode = WebAssemblyISD::VEC_SHL;
2443     break;
2444   case ISD::SRA:
2445     Opcode = WebAssemblyISD::VEC_SHR_S;
2446     break;
2447   case ISD::SRL:
2448     Opcode = WebAssemblyISD::VEC_SHR_U;
2449     break;
2450   default:
2451     llvm_unreachable("unexpected opcode");
2452   }
2453 
2454   return DAG.getNode(Opcode, DL, Op.getValueType(), Op.getOperand(0), ShiftVal);
2455 }
2456 
LowerFP_TO_INT_SAT(SDValue Op,SelectionDAG & DAG) const2457 SDValue WebAssemblyTargetLowering::LowerFP_TO_INT_SAT(SDValue Op,
2458                                                       SelectionDAG &DAG) const {
2459   SDLoc DL(Op);
2460   EVT ResT = Op.getValueType();
2461   EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
2462 
2463   if ((ResT == MVT::i32 || ResT == MVT::i64) &&
2464       (SatVT == MVT::i32 || SatVT == MVT::i64))
2465     return Op;
2466 
2467   if (ResT == MVT::v4i32 && SatVT == MVT::i32)
2468     return Op;
2469 
2470   return SDValue();
2471 }
2472 
2473 //===----------------------------------------------------------------------===//
2474 //   Custom DAG combine hooks
2475 //===----------------------------------------------------------------------===//
2476 static SDValue
performVECTOR_SHUFFLECombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2477 performVECTOR_SHUFFLECombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
2478   auto &DAG = DCI.DAG;
2479   auto Shuffle = cast<ShuffleVectorSDNode>(N);
2480 
2481   // Hoist vector bitcasts that don't change the number of lanes out of unary
2482   // shuffles, where they are less likely to get in the way of other combines.
2483   // (shuffle (vNxT1 (bitcast (vNxT0 x))), undef, mask) ->
2484   //  (vNxT1 (bitcast (vNxT0 (shuffle x, undef, mask))))
2485   SDValue Bitcast = N->getOperand(0);
2486   if (Bitcast.getOpcode() != ISD::BITCAST)
2487     return SDValue();
2488   if (!N->getOperand(1).isUndef())
2489     return SDValue();
2490   SDValue CastOp = Bitcast.getOperand(0);
2491   EVT SrcType = CastOp.getValueType();
2492   EVT DstType = Bitcast.getValueType();
2493   if (!SrcType.is128BitVector() ||
2494       SrcType.getVectorNumElements() != DstType.getVectorNumElements())
2495     return SDValue();
2496   SDValue NewShuffle = DAG.getVectorShuffle(
2497       SrcType, SDLoc(N), CastOp, DAG.getUNDEF(SrcType), Shuffle->getMask());
2498   return DAG.getBitcast(DstType, NewShuffle);
2499 }
2500 
2501 /// Convert ({u,s}itofp vec) --> ({u,s}itofp ({s,z}ext vec)) so it doesn't get
2502 /// split up into scalar instructions during legalization, and the vector
2503 /// extending instructions are selected in performVectorExtendCombine below.
2504 static SDValue
performVectorExtendToFPCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2505 performVectorExtendToFPCombine(SDNode *N,
2506                                TargetLowering::DAGCombinerInfo &DCI) {
2507   auto &DAG = DCI.DAG;
2508   assert(N->getOpcode() == ISD::UINT_TO_FP ||
2509          N->getOpcode() == ISD::SINT_TO_FP);
2510 
2511   EVT InVT = N->getOperand(0)->getValueType(0);
2512   EVT ResVT = N->getValueType(0);
2513   MVT ExtVT;
2514   if (ResVT == MVT::v4f32 && (InVT == MVT::v4i16 || InVT == MVT::v4i8))
2515     ExtVT = MVT::v4i32;
2516   else if (ResVT == MVT::v2f64 && (InVT == MVT::v2i16 || InVT == MVT::v2i8))
2517     ExtVT = MVT::v2i32;
2518   else
2519     return SDValue();
2520 
2521   unsigned Op =
2522       N->getOpcode() == ISD::UINT_TO_FP ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
2523   SDValue Conv = DAG.getNode(Op, SDLoc(N), ExtVT, N->getOperand(0));
2524   return DAG.getNode(N->getOpcode(), SDLoc(N), ResVT, Conv);
2525 }
2526 
2527 static SDValue
performVectorExtendCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2528 performVectorExtendCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
2529   auto &DAG = DCI.DAG;
2530   assert(N->getOpcode() == ISD::SIGN_EXTEND ||
2531          N->getOpcode() == ISD::ZERO_EXTEND);
2532 
2533   // Combine ({s,z}ext (extract_subvector src, i)) into a widening operation if
2534   // possible before the extract_subvector can be expanded.
2535   auto Extract = N->getOperand(0);
2536   if (Extract.getOpcode() != ISD::EXTRACT_SUBVECTOR)
2537     return SDValue();
2538   auto Source = Extract.getOperand(0);
2539   auto *IndexNode = dyn_cast<ConstantSDNode>(Extract.getOperand(1));
2540   if (IndexNode == nullptr)
2541     return SDValue();
2542   auto Index = IndexNode->getZExtValue();
2543 
2544   // Only v8i8, v4i16, and v2i32 extracts can be widened, and only if the
2545   // extracted subvector is the low or high half of its source.
2546   EVT ResVT = N->getValueType(0);
2547   if (ResVT == MVT::v8i16) {
2548     if (Extract.getValueType() != MVT::v8i8 ||
2549         Source.getValueType() != MVT::v16i8 || (Index != 0 && Index != 8))
2550       return SDValue();
2551   } else if (ResVT == MVT::v4i32) {
2552     if (Extract.getValueType() != MVT::v4i16 ||
2553         Source.getValueType() != MVT::v8i16 || (Index != 0 && Index != 4))
2554       return SDValue();
2555   } else if (ResVT == MVT::v2i64) {
2556     if (Extract.getValueType() != MVT::v2i32 ||
2557         Source.getValueType() != MVT::v4i32 || (Index != 0 && Index != 2))
2558       return SDValue();
2559   } else {
2560     return SDValue();
2561   }
2562 
2563   bool IsSext = N->getOpcode() == ISD::SIGN_EXTEND;
2564   bool IsLow = Index == 0;
2565 
2566   unsigned Op = IsSext ? (IsLow ? WebAssemblyISD::EXTEND_LOW_S
2567                                 : WebAssemblyISD::EXTEND_HIGH_S)
2568                        : (IsLow ? WebAssemblyISD::EXTEND_LOW_U
2569                                 : WebAssemblyISD::EXTEND_HIGH_U);
2570 
2571   return DAG.getNode(Op, SDLoc(N), ResVT, Source);
2572 }
2573 
2574 static SDValue
performVectorTruncZeroCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2575 performVectorTruncZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
2576   auto &DAG = DCI.DAG;
2577 
2578   auto GetWasmConversionOp = [](unsigned Op) {
2579     switch (Op) {
2580     case ISD::FP_TO_SINT_SAT:
2581       return WebAssemblyISD::TRUNC_SAT_ZERO_S;
2582     case ISD::FP_TO_UINT_SAT:
2583       return WebAssemblyISD::TRUNC_SAT_ZERO_U;
2584     case ISD::FP_ROUND:
2585       return WebAssemblyISD::DEMOTE_ZERO;
2586     }
2587     llvm_unreachable("unexpected op");
2588   };
2589 
2590   auto IsZeroSplat = [](SDValue SplatVal) {
2591     auto *Splat = dyn_cast<BuildVectorSDNode>(SplatVal.getNode());
2592     APInt SplatValue, SplatUndef;
2593     unsigned SplatBitSize;
2594     bool HasAnyUndefs;
2595     // Endianness doesn't matter in this context because we are looking for
2596     // an all-zero value.
2597     return Splat &&
2598            Splat->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
2599                                   HasAnyUndefs) &&
2600            SplatValue == 0;
2601   };
2602 
2603   if (N->getOpcode() == ISD::CONCAT_VECTORS) {
2604     // Combine this:
2605     //
2606     //   (concat_vectors (v2i32 (fp_to_{s,u}int_sat $x, 32)), (v2i32 (splat 0)))
2607     //
2608     // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x).
2609     //
2610     // Or this:
2611     //
2612     //   (concat_vectors (v2f32 (fp_round (v2f64 $x))), (v2f32 (splat 0)))
2613     //
2614     // into (f32x4.demote_zero_f64x2 $x).
2615     EVT ResVT;
2616     EVT ExpectedConversionType;
2617     auto Conversion = N->getOperand(0);
2618     auto ConversionOp = Conversion.getOpcode();
2619     switch (ConversionOp) {
2620     case ISD::FP_TO_SINT_SAT:
2621     case ISD::FP_TO_UINT_SAT:
2622       ResVT = MVT::v4i32;
2623       ExpectedConversionType = MVT::v2i32;
2624       break;
2625     case ISD::FP_ROUND:
2626       ResVT = MVT::v4f32;
2627       ExpectedConversionType = MVT::v2f32;
2628       break;
2629     default:
2630       return SDValue();
2631     }
2632 
2633     if (N->getValueType(0) != ResVT)
2634       return SDValue();
2635 
2636     if (Conversion.getValueType() != ExpectedConversionType)
2637       return SDValue();
2638 
2639     auto Source = Conversion.getOperand(0);
2640     if (Source.getValueType() != MVT::v2f64)
2641       return SDValue();
2642 
2643     if (!IsZeroSplat(N->getOperand(1)) ||
2644         N->getOperand(1).getValueType() != ExpectedConversionType)
2645       return SDValue();
2646 
2647     unsigned Op = GetWasmConversionOp(ConversionOp);
2648     return DAG.getNode(Op, SDLoc(N), ResVT, Source);
2649   }
2650 
2651   // Combine this:
2652   //
2653   //   (fp_to_{s,u}int_sat (concat_vectors $x, (v2f64 (splat 0))), 32)
2654   //
2655   // into (i32x4.trunc_sat_f64x2_zero_{s,u} $x).
2656   //
2657   // Or this:
2658   //
2659   //   (v4f32 (fp_round (concat_vectors $x, (v2f64 (splat 0)))))
2660   //
2661   // into (f32x4.demote_zero_f64x2 $x).
2662   EVT ResVT;
2663   auto ConversionOp = N->getOpcode();
2664   switch (ConversionOp) {
2665   case ISD::FP_TO_SINT_SAT:
2666   case ISD::FP_TO_UINT_SAT:
2667     ResVT = MVT::v4i32;
2668     break;
2669   case ISD::FP_ROUND:
2670     ResVT = MVT::v4f32;
2671     break;
2672   default:
2673     llvm_unreachable("unexpected op");
2674   }
2675 
2676   if (N->getValueType(0) != ResVT)
2677     return SDValue();
2678 
2679   auto Concat = N->getOperand(0);
2680   if (Concat.getValueType() != MVT::v4f64)
2681     return SDValue();
2682 
2683   auto Source = Concat.getOperand(0);
2684   if (Source.getValueType() != MVT::v2f64)
2685     return SDValue();
2686 
2687   if (!IsZeroSplat(Concat.getOperand(1)) ||
2688       Concat.getOperand(1).getValueType() != MVT::v2f64)
2689     return SDValue();
2690 
2691   unsigned Op = GetWasmConversionOp(ConversionOp);
2692   return DAG.getNode(Op, SDLoc(N), ResVT, Source);
2693 }
2694 
2695 // Helper to extract VectorWidth bits from Vec, starting from IdxVal.
extractSubVector(SDValue Vec,unsigned IdxVal,SelectionDAG & DAG,const SDLoc & DL,unsigned VectorWidth)2696 static SDValue extractSubVector(SDValue Vec, unsigned IdxVal, SelectionDAG &DAG,
2697                                 const SDLoc &DL, unsigned VectorWidth) {
2698   EVT VT = Vec.getValueType();
2699   EVT ElVT = VT.getVectorElementType();
2700   unsigned Factor = VT.getSizeInBits() / VectorWidth;
2701   EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
2702                                   VT.getVectorNumElements() / Factor);
2703 
2704   // Extract the relevant VectorWidth bits.  Generate an EXTRACT_SUBVECTOR
2705   unsigned ElemsPerChunk = VectorWidth / ElVT.getSizeInBits();
2706   assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
2707 
2708   // This is the index of the first element of the VectorWidth-bit chunk
2709   // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
2710   IdxVal &= ~(ElemsPerChunk - 1);
2711 
2712   // If the input is a buildvector just emit a smaller one.
2713   if (Vec.getOpcode() == ISD::BUILD_VECTOR)
2714     return DAG.getBuildVector(ResultVT, DL,
2715                               Vec->ops().slice(IdxVal, ElemsPerChunk));
2716 
2717   SDValue VecIdx = DAG.getIntPtrConstant(IdxVal, DL);
2718   return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ResultVT, Vec, VecIdx);
2719 }
2720 
2721 // Helper to recursively truncate vector elements in half with NARROW_U. DstVT
2722 // is the expected destination value type after recursion. In is the initial
2723 // input. Note that the input should have enough leading zero bits to prevent
2724 // NARROW_U from saturating results.
truncateVectorWithNARROW(EVT DstVT,SDValue In,const SDLoc & DL,SelectionDAG & DAG)2725 static SDValue truncateVectorWithNARROW(EVT DstVT, SDValue In, const SDLoc &DL,
2726                                         SelectionDAG &DAG) {
2727   EVT SrcVT = In.getValueType();
2728 
2729   // No truncation required, we might get here due to recursive calls.
2730   if (SrcVT == DstVT)
2731     return In;
2732 
2733   unsigned SrcSizeInBits = SrcVT.getSizeInBits();
2734   unsigned NumElems = SrcVT.getVectorNumElements();
2735   if (!isPowerOf2_32(NumElems))
2736     return SDValue();
2737   assert(DstVT.getVectorNumElements() == NumElems && "Illegal truncation");
2738   assert(SrcSizeInBits > DstVT.getSizeInBits() && "Illegal truncation");
2739 
2740   LLVMContext &Ctx = *DAG.getContext();
2741   EVT PackedSVT = EVT::getIntegerVT(Ctx, SrcVT.getScalarSizeInBits() / 2);
2742 
2743   // Narrow to the largest type possible:
2744   // vXi64/vXi32 -> i16x8.narrow_i32x4_u and vXi16 -> i8x16.narrow_i16x8_u.
2745   EVT InVT = MVT::i16, OutVT = MVT::i8;
2746   if (SrcVT.getScalarSizeInBits() > 16) {
2747     InVT = MVT::i32;
2748     OutVT = MVT::i16;
2749   }
2750   unsigned SubSizeInBits = SrcSizeInBits / 2;
2751   InVT = EVT::getVectorVT(Ctx, InVT, SubSizeInBits / InVT.getSizeInBits());
2752   OutVT = EVT::getVectorVT(Ctx, OutVT, SubSizeInBits / OutVT.getSizeInBits());
2753 
2754   // Split lower/upper subvectors.
2755   SDValue Lo = extractSubVector(In, 0, DAG, DL, SubSizeInBits);
2756   SDValue Hi = extractSubVector(In, NumElems / 2, DAG, DL, SubSizeInBits);
2757 
2758   // 256bit -> 128bit truncate - Narrow lower/upper 128-bit subvectors.
2759   if (SrcVT.is256BitVector() && DstVT.is128BitVector()) {
2760     Lo = DAG.getBitcast(InVT, Lo);
2761     Hi = DAG.getBitcast(InVT, Hi);
2762     SDValue Res = DAG.getNode(WebAssemblyISD::NARROW_U, DL, OutVT, Lo, Hi);
2763     return DAG.getBitcast(DstVT, Res);
2764   }
2765 
2766   // Recursively narrow lower/upper subvectors, concat result and narrow again.
2767   EVT PackedVT = EVT::getVectorVT(Ctx, PackedSVT, NumElems / 2);
2768   Lo = truncateVectorWithNARROW(PackedVT, Lo, DL, DAG);
2769   Hi = truncateVectorWithNARROW(PackedVT, Hi, DL, DAG);
2770 
2771   PackedVT = EVT::getVectorVT(Ctx, PackedSVT, NumElems);
2772   SDValue Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, PackedVT, Lo, Hi);
2773   return truncateVectorWithNARROW(DstVT, Res, DL, DAG);
2774 }
2775 
performTruncateCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2776 static SDValue performTruncateCombine(SDNode *N,
2777                                       TargetLowering::DAGCombinerInfo &DCI) {
2778   auto &DAG = DCI.DAG;
2779 
2780   SDValue In = N->getOperand(0);
2781   EVT InVT = In.getValueType();
2782   if (!InVT.isSimple())
2783     return SDValue();
2784 
2785   EVT OutVT = N->getValueType(0);
2786   if (!OutVT.isVector())
2787     return SDValue();
2788 
2789   EVT OutSVT = OutVT.getVectorElementType();
2790   EVT InSVT = InVT.getVectorElementType();
2791   // Currently only cover truncate to v16i8 or v8i16.
2792   if (!((InSVT == MVT::i16 || InSVT == MVT::i32 || InSVT == MVT::i64) &&
2793         (OutSVT == MVT::i8 || OutSVT == MVT::i16) && OutVT.is128BitVector()))
2794     return SDValue();
2795 
2796   SDLoc DL(N);
2797   APInt Mask = APInt::getLowBitsSet(InVT.getScalarSizeInBits(),
2798                                     OutVT.getScalarSizeInBits());
2799   In = DAG.getNode(ISD::AND, DL, InVT, In, DAG.getConstant(Mask, DL, InVT));
2800   return truncateVectorWithNARROW(OutVT, In, DL, DAG);
2801 }
2802 
performBitcastCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2803 static SDValue performBitcastCombine(SDNode *N,
2804                                      TargetLowering::DAGCombinerInfo &DCI) {
2805   auto &DAG = DCI.DAG;
2806   SDLoc DL(N);
2807   SDValue Src = N->getOperand(0);
2808   EVT VT = N->getValueType(0);
2809   EVT SrcVT = Src.getValueType();
2810 
2811   // bitcast <N x i1> to iN
2812   //   ==> bitmask
2813   if (DCI.isBeforeLegalize() && VT.isScalarInteger() &&
2814       SrcVT.isFixedLengthVector() && SrcVT.getScalarType() == MVT::i1) {
2815     unsigned NumElts = SrcVT.getVectorNumElements();
2816     if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2817       return SDValue();
2818     EVT Width = MVT::getIntegerVT(128 / NumElts);
2819     return DAG.getZExtOrTrunc(
2820         DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
2821                     {DAG.getConstant(Intrinsic::wasm_bitmask, DL, MVT::i32),
2822                      DAG.getSExtOrTrunc(N->getOperand(0), DL,
2823                                         SrcVT.changeVectorElementType(Width))}),
2824         DL, VT);
2825   }
2826 
2827   return SDValue();
2828 }
2829 
performSETCCCombine(SDNode * N,TargetLowering::DAGCombinerInfo & DCI)2830 static SDValue performSETCCCombine(SDNode *N,
2831                                    TargetLowering::DAGCombinerInfo &DCI) {
2832   auto &DAG = DCI.DAG;
2833 
2834   SDValue LHS = N->getOperand(0);
2835   SDValue RHS = N->getOperand(1);
2836   ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(2))->get();
2837   SDLoc DL(N);
2838   EVT VT = N->getValueType(0);
2839 
2840   // setcc (iN (bitcast (vNi1 X))), 0, ne
2841   //   ==> any_true (vNi1 X)
2842   // setcc (iN (bitcast (vNi1 X))), 0, eq
2843   //   ==> xor (any_true (vNi1 X)), -1
2844   // setcc (iN (bitcast (vNi1 X))), -1, eq
2845   //   ==> all_true (vNi1 X)
2846   // setcc (iN (bitcast (vNi1 X))), -1, ne
2847   //   ==> xor (all_true (vNi1 X)), -1
2848   if (DCI.isBeforeLegalize() && VT.isScalarInteger() &&
2849       (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
2850       (isNullConstant(RHS) || isAllOnesConstant(RHS)) &&
2851       LHS->getOpcode() == ISD::BITCAST) {
2852     EVT FromVT = LHS->getOperand(0).getValueType();
2853     if (FromVT.isFixedLengthVector() &&
2854         FromVT.getVectorElementType() == MVT::i1) {
2855       int Intrin = isNullConstant(RHS) ? Intrinsic::wasm_anytrue
2856                                        : Intrinsic::wasm_alltrue;
2857       unsigned NumElts = FromVT.getVectorNumElements();
2858       if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2859         return SDValue();
2860       EVT Width = MVT::getIntegerVT(128 / NumElts);
2861       SDValue Ret = DAG.getZExtOrTrunc(
2862           DAG.getNode(
2863               ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
2864               {DAG.getConstant(Intrin, DL, MVT::i32),
2865                DAG.getSExtOrTrunc(LHS->getOperand(0), DL,
2866                                   FromVT.changeVectorElementType(Width))}),
2867           DL, MVT::i1);
2868       if ((isNullConstant(RHS) && (Cond == ISD::SETEQ)) ||
2869           (isAllOnesConstant(RHS) && (Cond == ISD::SETNE))) {
2870         Ret = DAG.getNOT(DL, Ret, MVT::i1);
2871       }
2872       return DAG.getZExtOrTrunc(Ret, DL, VT);
2873     }
2874   }
2875 
2876   return SDValue();
2877 }
2878 
2879 SDValue
PerformDAGCombine(SDNode * N,DAGCombinerInfo & DCI) const2880 WebAssemblyTargetLowering::PerformDAGCombine(SDNode *N,
2881                                              DAGCombinerInfo &DCI) const {
2882   switch (N->getOpcode()) {
2883   default:
2884     return SDValue();
2885   case ISD::BITCAST:
2886     return performBitcastCombine(N, DCI);
2887   case ISD::SETCC:
2888     return performSETCCCombine(N, DCI);
2889   case ISD::VECTOR_SHUFFLE:
2890     return performVECTOR_SHUFFLECombine(N, DCI);
2891   case ISD::SIGN_EXTEND:
2892   case ISD::ZERO_EXTEND:
2893     return performVectorExtendCombine(N, DCI);
2894   case ISD::UINT_TO_FP:
2895   case ISD::SINT_TO_FP:
2896     return performVectorExtendToFPCombine(N, DCI);
2897   case ISD::FP_TO_SINT_SAT:
2898   case ISD::FP_TO_UINT_SAT:
2899   case ISD::FP_ROUND:
2900   case ISD::CONCAT_VECTORS:
2901     return performVectorTruncZeroCombine(N, DCI);
2902   case ISD::TRUNCATE:
2903     return performTruncateCombine(N, DCI);
2904   }
2905 }
2906