blob: 7ab8e361a8becb29cd6d31642e3e8802abb0037e [file] [log] [blame]
Derek Schuffccdceda2016-08-18 15:27:25 +00001//=== WebAssemblyLowerEmscriptenEHSjLj.cpp - Lower exceptions for Emscripten =//
Derek Schufff41f67d2016-08-01 21:34:04 +00002//
Chandler Carruth2946cd72019-01-19 08:50:56 +00003// 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
Derek Schufff41f67d2016-08-01 21:34:04 +00006//
7//===----------------------------------------------------------------------===//
8///
9/// \file
Adrian Prantl5f8f34e42018-05-01 15:54:18 +000010/// This file lowers exception-related instructions and setjmp/longjmp
Heejin Ahnc0f18172016-09-01 21:05:15 +000011/// function calls in order to use Emscripten's JavaScript try and catch
12/// mechanism.
Derek Schufff41f67d2016-08-01 21:34:04 +000013///
Heejin Ahnc0f18172016-09-01 21:05:15 +000014/// To handle exceptions and setjmp/longjmps, this scheme relies on JavaScript's
15/// try and catch syntax and relevant exception-related libraries implemented
16/// in JavaScript glue code that will be produced by Emscripten. This is similar
17/// to the current Emscripten asm.js exception handling in fastcomp. For
18/// fastcomp's EH / SjLj scheme, see these files in fastcomp LLVM branch:
Derek Schufff41f67d2016-08-01 21:34:04 +000019/// (Location: https://github.com/kripken/emscripten-fastcomp)
20/// lib/Target/JSBackend/NaCl/LowerEmExceptionsPass.cpp
Heejin Ahnc0f18172016-09-01 21:05:15 +000021/// lib/Target/JSBackend/NaCl/LowerEmSetjmp.cpp
Derek Schufff41f67d2016-08-01 21:34:04 +000022/// lib/Target/JSBackend/JSBackend.cpp
23/// lib/Target/JSBackend/CallHandlers.h
24///
Heejin Ahnc0f18172016-09-01 21:05:15 +000025/// * Exception handling
26/// This pass lowers invokes and landingpads into library functions in JS glue
27/// code. Invokes are lowered into function wrappers called invoke wrappers that
28/// exist in JS side, which wraps the original function call with JS try-catch.
29/// If an exception occurred, cxa_throw() function in JS side sets some
30/// variables (see below) so we can check whether an exception occurred from
31/// wasm code and handle it appropriately.
Derek Schufff41f67d2016-08-01 21:34:04 +000032///
Heejin Ahnc0f18172016-09-01 21:05:15 +000033/// * Setjmp-longjmp handling
34/// This pass lowers setjmp to a reasonably-performant approach for emscripten.
35/// The idea is that each block with a setjmp is broken up into two parts: the
36/// part containing setjmp and the part right after the setjmp. The latter part
37/// is either reached from the setjmp, or later from a longjmp. To handle the
38/// longjmp, all calls that might longjmp are also called using invoke wrappers
39/// and thus JS / try-catch. JS longjmp() function also sets some variables so
40/// we can check / whether a longjmp occurred from wasm code. Each block with a
41/// function call that might longjmp is also split up after the longjmp call.
42/// After the longjmp call, we check whether a longjmp occurred, and if it did,
43/// which setjmp it corresponds to, and jump to the right post-setjmp block.
44/// We assume setjmp-longjmp handling always run after EH handling, which means
45/// we don't expect any exception-related instructions when SjLj runs.
46/// FIXME Currently this scheme does not support indirect call of setjmp,
47/// because of the limitation of the scheme itself. fastcomp does not support it
48/// either.
49///
50/// In detail, this pass does following things:
51///
Sam Clegg4791a662018-11-20 19:25:07 +000052/// 1) Assumes the existence of global variables: __THREW__, __threwValue
53/// __THREW__ and __threwValue will be set in invoke wrappers
Heejin Ahnc0f18172016-09-01 21:05:15 +000054/// in JS glue code. For what invoke wrappers are, refer to 3). These
55/// variables are used for both exceptions and setjmp/longjmps.
56/// __THREW__ indicates whether an exception or a longjmp occurred or not. 0
57/// means nothing occurred, 1 means an exception occurred, and other numbers
Heejin Ahn99bd16b2016-09-10 02:33:47 +000058/// mean a longjmp occurred. In the case of longjmp, __threwValue variable
Heejin Ahnc0f18172016-09-01 21:05:15 +000059/// indicates the corresponding setjmp buffer the longjmp corresponds to.
Heejin Ahnc0f18172016-09-01 21:05:15 +000060///
61/// * Exception handling
Derek Schufff41f67d2016-08-01 21:34:04 +000062///
Sam Clegg4791a662018-11-20 19:25:07 +000063/// 2) We assume the existence of setThrew and setTempRet0/getTempRet0 functions
64/// at link time.
Sam Cleggb2486f12018-10-02 22:12:15 +000065/// The global variables in 1) will exist in wasm address space,
66/// but their values should be set in JS code, so these functions
Derek Schufff41f67d2016-08-01 21:34:04 +000067/// as interfaces to JS glue code. These functions are equivalent to the
68/// following JS functions, which actually exist in asm.js version of JS
69/// library.
70///
71/// function setThrew(threw, value) {
72/// if (__THREW__ == 0) {
73/// __THREW__ = threw;
Derek Schuffccdceda2016-08-18 15:27:25 +000074/// __threwValue = value;
Derek Schufff41f67d2016-08-01 21:34:04 +000075/// }
76/// }
Sam Clegg4791a662018-11-20 19:25:07 +000077//
78/// setTempRet0 is called from __cxa_find_matching_catch() in JS glue code.
Derek Schufff41f67d2016-08-01 21:34:04 +000079///
Sam Clegg4791a662018-11-20 19:25:07 +000080/// In exception handling, getTempRet0 indicates the type of an exception
81/// caught, and in setjmp/longjmp, it means the second argument to longjmp
82/// function.
Derek Schufff41f67d2016-08-01 21:34:04 +000083///
84/// 3) Lower
85/// invoke @func(arg1, arg2) to label %invoke.cont unwind label %lpad
86/// into
87/// __THREW__ = 0;
Heejin Ahn99bd16b2016-09-10 02:33:47 +000088/// call @__invoke_SIG(func, arg1, arg2)
Derek Schufff41f67d2016-08-01 21:34:04 +000089/// %__THREW__.val = __THREW__;
90/// __THREW__ = 0;
Heejin Ahn99bd16b2016-09-10 02:33:47 +000091/// if (%__THREW__.val == 1)
92/// goto %lpad
93/// else
94/// goto %invoke.cont
Derek Schufff41f67d2016-08-01 21:34:04 +000095/// SIG is a mangled string generated based on the LLVM IR-level function
96/// signature. After LLVM IR types are lowered to the target wasm types,
97/// the names for these wrappers will change based on wasm types as well,
98/// as in invoke_vi (function takes an int and returns void). The bodies of
99/// these wrappers will be generated in JS glue code, and inside those
100/// wrappers we use JS try-catch to generate actual exception effects. It
101/// also calls the original callee function. An example wrapper in JS code
102/// would look like this:
103/// function invoke_vi(index,a1) {
104/// try {
105/// Module["dynCall_vi"](index,a1); // This calls original callee
106/// } catch(e) {
107/// if (typeof e !== 'number' && e !== 'longjmp') throw e;
108/// asm["setThrew"](1, 0); // setThrew is called here
109/// }
110/// }
111/// If an exception is thrown, __THREW__ will be set to true in a wrapper,
112/// so we can jump to the right BB based on this value.
113///
114/// 4) Lower
115/// %val = landingpad catch c1 catch c2 catch c3 ...
116/// ... use %val ...
117/// into
Derek Schuff53b9af02016-08-09 00:29:55 +0000118/// %fmc = call @__cxa_find_matching_catch_N(c1, c2, c3, ...)
Sam Clegg4791a662018-11-20 19:25:07 +0000119/// %val = {%fmc, getTempRet0()}
Derek Schufff41f67d2016-08-01 21:34:04 +0000120/// ... use %val ...
121/// Here N is a number calculated based on the number of clauses.
Sam Clegg4791a662018-11-20 19:25:07 +0000122/// setTempRet0 is called from __cxa_find_matching_catch() in JS glue code.
Derek Schufff41f67d2016-08-01 21:34:04 +0000123///
124/// 5) Lower
125/// resume {%a, %b}
126/// into
Derek Schuff53b9af02016-08-09 00:29:55 +0000127/// call @__resumeException(%a)
128/// where __resumeException() is a function in JS glue code.
Derek Schufff41f67d2016-08-01 21:34:04 +0000129///
Derek Schuff53b9af02016-08-09 00:29:55 +0000130/// 6) Lower
131/// call @llvm.eh.typeid.for(type) (intrinsic)
132/// into
133/// call @llvm_eh_typeid_for(type)
134/// llvm_eh_typeid_for function will be generated in JS glue code.
Derek Schufff41f67d2016-08-01 21:34:04 +0000135///
Heejin Ahnc0f18172016-09-01 21:05:15 +0000136/// * Setjmp / Longjmp handling
137///
Heejin Ahn0c68a872018-11-08 22:56:26 +0000138/// In case calls to longjmp() exists
139///
140/// 1) Lower
141/// longjmp(buf, value)
142/// into
143/// emscripten_longjmp_jmpbuf(buf, value)
144/// emscripten_longjmp_jmpbuf will be lowered to emscripten_longjmp later.
145///
146/// In case calls to setjmp() exists
147///
148/// 2) In the function entry that calls setjmp, initialize setjmpTable and
Heejin Ahnc0f18172016-09-01 21:05:15 +0000149/// sejmpTableSize as follows:
150/// setjmpTableSize = 4;
151/// setjmpTable = (int *) malloc(40);
152/// setjmpTable[0] = 0;
153/// setjmpTable and setjmpTableSize are used in saveSetjmp() function in JS
154/// code.
155///
Heejin Ahn0c68a872018-11-08 22:56:26 +0000156/// 3) Lower
Heejin Ahnc0f18172016-09-01 21:05:15 +0000157/// setjmp(buf)
158/// into
159/// setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
Sam Clegg4791a662018-11-20 19:25:07 +0000160/// setjmpTableSize = getTempRet0();
Heejin Ahnc0f18172016-09-01 21:05:15 +0000161/// For each dynamic setjmp call, setjmpTable stores its ID (a number which
162/// is incrementally assigned from 0) and its label (a unique number that
163/// represents each callsite of setjmp). When we need more entries in
164/// setjmpTable, it is reallocated in saveSetjmp() in JS code and it will
165/// return the new table address, and assign the new table size in
Sam Clegg4791a662018-11-20 19:25:07 +0000166/// setTempRet0(). saveSetjmp also stores the setjmp's ID into the buffer
167/// buf. A BB with setjmp is split into two after setjmp call in order to
168/// make the post-setjmp BB the possible destination of longjmp BB.
Heejin Ahnc0f18172016-09-01 21:05:15 +0000169///
Heejin Ahnc0f18172016-09-01 21:05:15 +0000170///
Heejin Ahn0c68a872018-11-08 22:56:26 +0000171/// 4) Lower every call that might longjmp into
Heejin Ahnc0f18172016-09-01 21:05:15 +0000172/// __THREW__ = 0;
Heejin Ahn99bd16b2016-09-10 02:33:47 +0000173/// call @__invoke_SIG(func, arg1, arg2)
Heejin Ahnc0f18172016-09-01 21:05:15 +0000174/// %__THREW__.val = __THREW__;
175/// __THREW__ = 0;
Heejin Ahn99bd16b2016-09-10 02:33:47 +0000176/// if (%__THREW__.val != 0 & __threwValue != 0) {
Heejin Ahnc0f18172016-09-01 21:05:15 +0000177/// %label = testSetjmp(mem[%__THREW__.val], setjmpTable,
178/// setjmpTableSize);
179/// if (%label == 0)
Heejin Ahn99bd16b2016-09-10 02:33:47 +0000180/// emscripten_longjmp(%__THREW__.val, __threwValue);
Sam Clegg4791a662018-11-20 19:25:07 +0000181/// setTempRet0(__threwValue);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000182/// } else {
183/// %label = -1;
184/// }
Sam Clegg4791a662018-11-20 19:25:07 +0000185/// longjmp_result = getTempRet0();
Heejin Ahnc0f18172016-09-01 21:05:15 +0000186/// switch label {
Heejin Ahn99bd16b2016-09-10 02:33:47 +0000187/// label 1: goto post-setjmp BB 1
188/// label 2: goto post-setjmp BB 2
Heejin Ahnc0f18172016-09-01 21:05:15 +0000189/// ...
Heejin Ahn99bd16b2016-09-10 02:33:47 +0000190/// default: goto splitted next BB
Heejin Ahnc0f18172016-09-01 21:05:15 +0000191/// }
Heejin Ahn0c68a872018-11-08 22:56:26 +0000192/// testSetjmp examines setjmpTable to see if there is a matching setjmp
193/// call. After calling an invoke wrapper, if a longjmp occurred, __THREW__
194/// will be the address of matching jmp_buf buffer and __threwValue be the
195/// second argument to longjmp. mem[__THREW__.val] is a setjmp ID that is
196/// stored in saveSetjmp. testSetjmp returns a setjmp label, a unique ID to
197/// each setjmp callsite. Label 0 means this longjmp buffer does not
198/// correspond to one of the setjmp callsites in this function, so in this
199/// case we just chain the longjmp to the caller. (Here we call
200/// emscripten_longjmp, which is different from emscripten_longjmp_jmpbuf.
201/// emscripten_longjmp_jmpbuf takes jmp_buf as its first argument, while
202/// emscripten_longjmp takes an int. Both of them will eventually be lowered
203/// to emscripten_longjmp in s2wasm, but here we need two signatures - we
204/// can't translate an int value to a jmp_buf.)
205/// Label -1 means no longjmp occurred. Otherwise we jump to the right
206/// post-setjmp BB based on the label.
Heejin Ahnc0f18172016-09-01 21:05:15 +0000207///
Derek Schufff41f67d2016-08-01 21:34:04 +0000208///===----------------------------------------------------------------------===//
209
210#include "WebAssembly.h"
Derek Schuff53b9af02016-08-09 00:29:55 +0000211#include "llvm/IR/CallSite.h"
Heejin Ahnc0f18172016-09-01 21:05:15 +0000212#include "llvm/IR/Dominators.h"
Derek Schufff41f67d2016-08-01 21:34:04 +0000213#include "llvm/IR/IRBuilder.h"
Heejin Ahnc0f18172016-09-01 21:05:15 +0000214#include "llvm/Transforms/Utils/BasicBlockUtils.h"
215#include "llvm/Transforms/Utils/SSAUpdater.h"
Derek Schufff41f67d2016-08-01 21:34:04 +0000216
217using namespace llvm;
218
Derek Schuffccdceda2016-08-18 15:27:25 +0000219#define DEBUG_TYPE "wasm-lower-em-ehsjlj"
Derek Schufff41f67d2016-08-01 21:34:04 +0000220
Derek Schuff66641322016-08-09 22:37:00 +0000221static cl::list<std::string>
Derek Schuffccdceda2016-08-18 15:27:25 +0000222 EHWhitelist("emscripten-cxx-exceptions-whitelist",
223 cl::desc("The list of function names in which Emscripten-style "
224 "exception handling is enabled (see emscripten "
225 "EMSCRIPTEN_CATCHING_WHITELIST options)"),
226 cl::CommaSeparated);
Derek Schuff66641322016-08-09 22:37:00 +0000227
Derek Schufff41f67d2016-08-01 21:34:04 +0000228namespace {
Derek Schuffccdceda2016-08-18 15:27:25 +0000229class WebAssemblyLowerEmscriptenEHSjLj final : public ModulePass {
Derek Schuffccdceda2016-08-18 15:27:25 +0000230 static const char *ResumeFName;
231 static const char *EHTypeIDFName;
Heejin Ahnc0f18172016-09-01 21:05:15 +0000232 static const char *EmLongjmpFName;
233 static const char *EmLongjmpJmpbufFName;
234 static const char *SaveSetjmpFName;
235 static const char *TestSetjmpFName;
Derek Schuffccdceda2016-08-18 15:27:25 +0000236 static const char *FindMatchingCatchPrefix;
237 static const char *InvokePrefix;
238
Heejin Ahnc0f18172016-09-01 21:05:15 +0000239 bool EnableEH; // Enable exception handling
240 bool EnableSjLj; // Enable setjmp/longjmp handling
Derek Schuffccdceda2016-08-18 15:27:25 +0000241
Heejin Ahn18c56a02019-02-04 19:13:39 +0000242 GlobalVariable *ThrewGV = nullptr;
243 GlobalVariable *ThrewValueGV = nullptr;
244 Function *GetTempRet0Func = nullptr;
245 Function *SetTempRet0Func = nullptr;
246 Function *ResumeF = nullptr;
247 Function *EHTypeIDF = nullptr;
248 Function *EmLongjmpF = nullptr;
249 Function *EmLongjmpJmpbufF = nullptr;
250 Function *SaveSetjmpF = nullptr;
251 Function *TestSetjmpF = nullptr;
Heejin Ahnc0f18172016-09-01 21:05:15 +0000252
Derek Schuffccdceda2016-08-18 15:27:25 +0000253 // __cxa_find_matching_catch_N functions.
254 // Indexed by the number of clauses in an original landingpad instruction.
255 DenseMap<int, Function *> FindMatchingCatches;
256 // Map of <function signature string, invoke_ wrappers>
257 StringMap<Function *> InvokeWrappers;
258 // Set of whitelisted function names for exception handling
259 std::set<std::string> EHWhitelistSet;
260
Mehdi Amini117296c2016-10-01 02:56:57 +0000261 StringRef getPassName() const override {
Derek Schufff41f67d2016-08-01 21:34:04 +0000262 return "WebAssembly Lower Emscripten Exceptions";
263 }
264
Derek Schuffccdceda2016-08-18 15:27:25 +0000265 bool runEHOnFunction(Function &F);
266 bool runSjLjOnFunction(Function &F);
Derek Schufff41f67d2016-08-01 21:34:04 +0000267 Function *getFindMatchingCatch(Module &M, unsigned NumClauses);
268
Heejin Ahnc0f18172016-09-01 21:05:15 +0000269 template <typename CallOrInvoke> Value *wrapInvoke(CallOrInvoke *CI);
270 void wrapTestSetjmp(BasicBlock *BB, Instruction *InsertPt, Value *Threw,
271 Value *SetjmpTable, Value *SetjmpTableSize, Value *&Label,
272 Value *&LongjmpResult, BasicBlock *&EndBB);
273 template <typename CallOrInvoke> Function *getInvokeWrapper(CallOrInvoke *CI);
274
Derek Schuffccdceda2016-08-18 15:27:25 +0000275 bool areAllExceptionsAllowed() const { return EHWhitelistSet.empty(); }
Heejin Ahnc0f18172016-09-01 21:05:15 +0000276 bool canLongjmp(Module &M, const Value *Callee) const;
277
Heejin Ahnc0f18172016-09-01 21:05:15 +0000278 void rebuildSSA(Function &F);
Derek Schufff41f67d2016-08-01 21:34:04 +0000279
280public:
281 static char ID;
282
Heejin Ahnc0f18172016-09-01 21:05:15 +0000283 WebAssemblyLowerEmscriptenEHSjLj(bool EnableEH = true, bool EnableSjLj = true)
Heejin Ahn18c56a02019-02-04 19:13:39 +0000284 : ModulePass(ID), EnableEH(EnableEH), EnableSjLj(EnableSjLj) {
Derek Schuffccdceda2016-08-18 15:27:25 +0000285 EHWhitelistSet.insert(EHWhitelist.begin(), EHWhitelist.end());
Derek Schuff66641322016-08-09 22:37:00 +0000286 }
Derek Schufff41f67d2016-08-01 21:34:04 +0000287 bool runOnModule(Module &M) override;
Heejin Ahnc0f18172016-09-01 21:05:15 +0000288
289 void getAnalysisUsage(AnalysisUsage &AU) const override {
290 AU.addRequired<DominatorTreeWrapperPass>();
291 }
Derek Schufff41f67d2016-08-01 21:34:04 +0000292};
293} // End anonymous namespace
294
Derek Schuffccdceda2016-08-18 15:27:25 +0000295const char *WebAssemblyLowerEmscriptenEHSjLj::ResumeFName = "__resumeException";
296const char *WebAssemblyLowerEmscriptenEHSjLj::EHTypeIDFName =
297 "llvm_eh_typeid_for";
Heejin Ahnc0f18172016-09-01 21:05:15 +0000298const char *WebAssemblyLowerEmscriptenEHSjLj::EmLongjmpFName =
299 "emscripten_longjmp";
300const char *WebAssemblyLowerEmscriptenEHSjLj::EmLongjmpJmpbufFName =
301 "emscripten_longjmp_jmpbuf";
302const char *WebAssemblyLowerEmscriptenEHSjLj::SaveSetjmpFName = "saveSetjmp";
303const char *WebAssemblyLowerEmscriptenEHSjLj::TestSetjmpFName = "testSetjmp";
Derek Schuffccdceda2016-08-18 15:27:25 +0000304const char *WebAssemblyLowerEmscriptenEHSjLj::FindMatchingCatchPrefix =
305 "__cxa_find_matching_catch_";
306const char *WebAssemblyLowerEmscriptenEHSjLj::InvokePrefix = "__invoke_";
Derek Schufff41f67d2016-08-01 21:34:04 +0000307
Derek Schuffccdceda2016-08-18 15:27:25 +0000308char WebAssemblyLowerEmscriptenEHSjLj::ID = 0;
309INITIALIZE_PASS(WebAssemblyLowerEmscriptenEHSjLj, DEBUG_TYPE,
310 "WebAssembly Lower Emscripten Exceptions / Setjmp / Longjmp",
311 false, false)
312
Heejin Ahnc0f18172016-09-01 21:05:15 +0000313ModulePass *llvm::createWebAssemblyLowerEmscriptenEHSjLj(bool EnableEH,
314 bool EnableSjLj) {
315 return new WebAssemblyLowerEmscriptenEHSjLj(EnableEH, EnableSjLj);
Derek Schufff41f67d2016-08-01 21:34:04 +0000316}
317
318static bool canThrow(const Value *V) {
319 if (const auto *F = dyn_cast<const Function>(V)) {
320 // Intrinsics cannot throw
321 if (F->isIntrinsic())
322 return false;
323 StringRef Name = F->getName();
324 // leave setjmp and longjmp (mostly) alone, we process them properly later
325 if (Name == "setjmp" || Name == "longjmp")
326 return false;
Heejin Ahnc0f18172016-09-01 21:05:15 +0000327 return !F->doesNotThrow();
Derek Schufff41f67d2016-08-01 21:34:04 +0000328 }
Heejin Ahnb6cd5122016-08-24 22:53:00 +0000329 // not a function, so an indirect call - can throw, we can't tell
330 return true;
Derek Schufff41f67d2016-08-01 21:34:04 +0000331}
332
Sam Cleggb2486f12018-10-02 22:12:15 +0000333// Get a global variable with the given name. If it doesn't exist declare it,
334// which will generate an import and asssumes that it will exist at link time.
335static GlobalVariable *getGlobalVariableI32(Module &M, IRBuilder<> &IRB,
336 const char *Name) {
Sam Clegg28b3e992018-07-17 16:40:03 +0000337
Sam Clegg12011fa2019-04-04 01:43:21 +0000338 auto* GV = dyn_cast<GlobalVariable>(M.getOrInsertGlobal(Name, IRB.getInt32Ty()));
339 if (!GV)
340 report_fatal_error(Twine("unable to create global: ") + Name);
341
342 return GV;
Derek Schufff41f67d2016-08-01 21:34:04 +0000343}
344
345// Simple function name mangler.
346// This function simply takes LLVM's string representation of parameter types
Derek Schuff53b9af02016-08-09 00:29:55 +0000347// and concatenate them with '_'. There are non-alphanumeric characters but llc
348// is ok with it, and we need to postprocess these names after the lowering
349// phase anyway.
Derek Schufff41f67d2016-08-01 21:34:04 +0000350static std::string getSignature(FunctionType *FTy) {
351 std::string Sig;
352 raw_string_ostream OS(Sig);
353 OS << *FTy->getReturnType();
354 for (Type *ParamTy : FTy->params())
355 OS << "_" << *ParamTy;
356 if (FTy->isVarArg())
357 OS << "_...";
358 Sig = OS.str();
David Majnemerc7004902016-08-12 04:32:37 +0000359 Sig.erase(remove_if(Sig, isspace), Sig.end());
Derek Schuff53b9af02016-08-09 00:29:55 +0000360 // When s2wasm parses .s file, a comma means the end of an argument. So a
361 // mangled function name can contain any character but a comma.
362 std::replace(Sig.begin(), Sig.end(), ',', '.');
Derek Schufff41f67d2016-08-01 21:34:04 +0000363 return Sig;
364}
365
Heejin Ahnc0f18172016-09-01 21:05:15 +0000366// Returns __cxa_find_matching_catch_N function, where N = NumClauses + 2.
367// This is because a landingpad instruction contains two more arguments, a
368// personality function and a cleanup bit, and __cxa_find_matching_catch_N
369// functions are named after the number of arguments in the original landingpad
370// instruction.
Derek Schuffccdceda2016-08-18 15:27:25 +0000371Function *
372WebAssemblyLowerEmscriptenEHSjLj::getFindMatchingCatch(Module &M,
373 unsigned NumClauses) {
Derek Schufff41f67d2016-08-01 21:34:04 +0000374 if (FindMatchingCatches.count(NumClauses))
375 return FindMatchingCatches[NumClauses];
376 PointerType *Int8PtrTy = Type::getInt8PtrTy(M.getContext());
377 SmallVector<Type *, 16> Args(NumClauses, Int8PtrTy);
378 FunctionType *FTy = FunctionType::get(Int8PtrTy, Args, false);
Derek Schuffccdceda2016-08-18 15:27:25 +0000379 Function *F =
380 Function::Create(FTy, GlobalValue::ExternalLinkage,
381 FindMatchingCatchPrefix + Twine(NumClauses + 2), &M);
Derek Schufff41f67d2016-08-01 21:34:04 +0000382 FindMatchingCatches[NumClauses] = F;
383 return F;
384}
385
Heejin Ahnc0f18172016-09-01 21:05:15 +0000386// Generate invoke wrapper seqence with preamble and postamble
387// Preamble:
388// __THREW__ = 0;
389// Postamble:
390// %__THREW__.val = __THREW__; __THREW__ = 0;
391// Returns %__THREW__.val, which indicates whether an exception is thrown (or
392// whether longjmp occurred), for future use.
393template <typename CallOrInvoke>
394Value *WebAssemblyLowerEmscriptenEHSjLj::wrapInvoke(CallOrInvoke *CI) {
395 LLVMContext &C = CI->getModule()->getContext();
396
397 // If we are calling a function that is noreturn, we must remove that
398 // attribute. The code we insert here does expect it to return, after we
399 // catch the exception.
400 if (CI->doesNotReturn()) {
401 if (auto *F = dyn_cast<Function>(CI->getCalledValue()))
402 F->removeFnAttr(Attribute::NoReturn);
Reid Klecknerb5180542017-03-21 16:57:19 +0000403 CI->removeAttribute(AttributeList::FunctionIndex, Attribute::NoReturn);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000404 }
405
406 IRBuilder<> IRB(C);
407 IRB.SetInsertPoint(CI);
408
409 // Pre-invoke
410 // __THREW__ = 0;
411 IRB.CreateStore(IRB.getInt32(0), ThrewGV);
412
413 // Invoke function wrapper in JavaScript
414 SmallVector<Value *, 16> Args;
415 // Put the pointer to the callee as first argument, so it can be called
416 // within the invoke wrapper later
417 Args.push_back(CI->getCalledValue());
418 Args.append(CI->arg_begin(), CI->arg_end());
419 CallInst *NewCall = IRB.CreateCall(getInvokeWrapper(CI), Args);
420 NewCall->takeName(CI);
421 NewCall->setCallingConv(CI->getCallingConv());
422 NewCall->setDebugLoc(CI->getDebugLoc());
423
424 // Because we added the pointer to the callee as first argument, all
425 // argument attribute indices have to be incremented by one.
Reid Kleckner7f720332017-04-13 00:58:09 +0000426 SmallVector<AttributeSet, 8> ArgAttributes;
Derek Schuff0db0ca32017-04-12 16:03:00 +0000427 const AttributeList &InvokeAL = CI->getAttributes();
428
Derek Schuff0db0ca32017-04-12 16:03:00 +0000429 // No attributes for the callee pointer.
Reid Kleckner7f720332017-04-13 00:58:09 +0000430 ArgAttributes.push_back(AttributeSet());
Derek Schuff0db0ca32017-04-12 16:03:00 +0000431 // Copy the argument attributes from the original
Heejin Ahn18c56a02019-02-04 19:13:39 +0000432 for (unsigned I = 0, E = CI->getNumArgOperands(); I < E; ++I)
433 ArgAttributes.push_back(InvokeAL.getParamAttributes(I));
Derek Schuff0db0ca32017-04-12 16:03:00 +0000434
Heejin Ahnc0f18172016-09-01 21:05:15 +0000435 // Reconstruct the AttributesList based on the vector we constructed.
Reid Kleckner7f720332017-04-13 00:58:09 +0000436 AttributeList NewCallAL =
437 AttributeList::get(C, InvokeAL.getFnAttributes(),
438 InvokeAL.getRetAttributes(), ArgAttributes);
Derek Schuff0db0ca32017-04-12 16:03:00 +0000439 NewCall->setAttributes(NewCallAL);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000440
441 CI->replaceAllUsesWith(NewCall);
442
443 // Post-invoke
444 // %__THREW__.val = __THREW__; __THREW__ = 0;
James Y Knight14359ef2019-02-01 20:44:24 +0000445 Value *Threw =
446 IRB.CreateLoad(IRB.getInt32Ty(), ThrewGV, ThrewGV->getName() + ".val");
Heejin Ahnc0f18172016-09-01 21:05:15 +0000447 IRB.CreateStore(IRB.getInt32(0), ThrewGV);
448 return Threw;
449}
450
451// Get matching invoke wrapper based on callee signature
452template <typename CallOrInvoke>
453Function *WebAssemblyLowerEmscriptenEHSjLj::getInvokeWrapper(CallOrInvoke *CI) {
454 Module *M = CI->getModule();
Derek Schufff41f67d2016-08-01 21:34:04 +0000455 SmallVector<Type *, 16> ArgTys;
Heejin Ahnc0f18172016-09-01 21:05:15 +0000456 Value *Callee = CI->getCalledValue();
Derek Schufff41f67d2016-08-01 21:34:04 +0000457 FunctionType *CalleeFTy;
458 if (auto *F = dyn_cast<Function>(Callee))
459 CalleeFTy = F->getFunctionType();
460 else {
Heejin Ahnc0f18172016-09-01 21:05:15 +0000461 auto *CalleeTy = cast<PointerType>(Callee->getType())->getElementType();
Derek Schufff41f67d2016-08-01 21:34:04 +0000462 CalleeFTy = dyn_cast<FunctionType>(CalleeTy);
463 }
464
465 std::string Sig = getSignature(CalleeFTy);
466 if (InvokeWrappers.find(Sig) != InvokeWrappers.end())
467 return InvokeWrappers[Sig];
468
469 // Put the pointer to the callee as first argument
470 ArgTys.push_back(PointerType::getUnqual(CalleeFTy));
471 // Add argument types
472 ArgTys.append(CalleeFTy->param_begin(), CalleeFTy->param_end());
473
474 FunctionType *FTy = FunctionType::get(CalleeFTy->getReturnType(), ArgTys,
475 CalleeFTy->isVarArg());
476 Function *F = Function::Create(FTy, GlobalValue::ExternalLinkage,
Heejin Ahnc0f18172016-09-01 21:05:15 +0000477 InvokePrefix + Sig, M);
Derek Schufff41f67d2016-08-01 21:34:04 +0000478 InvokeWrappers[Sig] = F;
479 return F;
480}
481
Heejin Ahnc0f18172016-09-01 21:05:15 +0000482bool WebAssemblyLowerEmscriptenEHSjLj::canLongjmp(Module &M,
483 const Value *Callee) const {
484 if (auto *CalleeF = dyn_cast<Function>(Callee))
485 if (CalleeF->isIntrinsic())
486 return false;
Heejin Ahn23d57102016-08-31 22:40:34 +0000487
Heejin Ahnc0f18172016-09-01 21:05:15 +0000488 // The reason we include malloc/free here is to exclude the malloc/free
489 // calls generated in setjmp prep / cleanup routines.
490 Function *SetjmpF = M.getFunction("setjmp");
491 Function *MallocF = M.getFunction("malloc");
492 Function *FreeF = M.getFunction("free");
493 if (Callee == SetjmpF || Callee == MallocF || Callee == FreeF)
Heejin Ahn10a70862016-09-01 00:44:37 +0000494 return false;
495
Heejin Ahnc0f18172016-09-01 21:05:15 +0000496 // There are functions in JS glue code
497 if (Callee == ResumeF || Callee == EHTypeIDF || Callee == SaveSetjmpF ||
498 Callee == TestSetjmpF)
499 return false;
Heejin Ahn10a70862016-09-01 00:44:37 +0000500
Heejin Ahnc0f18172016-09-01 21:05:15 +0000501 // __cxa_find_matching_catch_N functions cannot longjmp
502 if (Callee->getName().startswith(FindMatchingCatchPrefix))
503 return false;
504
505 // Exception-catching related functions
506 Function *BeginCatchF = M.getFunction("__cxa_begin_catch");
507 Function *EndCatchF = M.getFunction("__cxa_end_catch");
508 Function *AllocExceptionF = M.getFunction("__cxa_allocate_exception");
509 Function *ThrowF = M.getFunction("__cxa_throw");
510 Function *TerminateF = M.getFunction("__clang_call_terminate");
511 if (Callee == BeginCatchF || Callee == EndCatchF ||
Sam Clegg4791a662018-11-20 19:25:07 +0000512 Callee == AllocExceptionF || Callee == ThrowF || Callee == TerminateF ||
513 Callee == GetTempRet0Func || Callee == SetTempRet0Func)
Heejin Ahnc0f18172016-09-01 21:05:15 +0000514 return false;
515
516 // Otherwise we don't know
517 return true;
518}
519
520// Generate testSetjmp function call seqence with preamble and postamble.
521// The code this generates is equivalent to the following JavaScript code:
522// if (%__THREW__.val != 0 & threwValue != 0) {
523// %label = _testSetjmp(mem[%__THREW__.val], setjmpTable, setjmpTableSize);
524// if (%label == 0)
525// emscripten_longjmp(%__THREW__.val, threwValue);
Sam Clegg4791a662018-11-20 19:25:07 +0000526// setTempRet0(threwValue);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000527// } else {
528// %label = -1;
529// }
Sam Clegg4791a662018-11-20 19:25:07 +0000530// %longjmp_result = getTempRet0();
Heejin Ahnc0f18172016-09-01 21:05:15 +0000531//
532// As output parameters. returns %label, %longjmp_result, and the BB the last
533// instruction (%longjmp_result = ...) is in.
534void WebAssemblyLowerEmscriptenEHSjLj::wrapTestSetjmp(
535 BasicBlock *BB, Instruction *InsertPt, Value *Threw, Value *SetjmpTable,
536 Value *SetjmpTableSize, Value *&Label, Value *&LongjmpResult,
537 BasicBlock *&EndBB) {
538 Function *F = BB->getParent();
539 LLVMContext &C = BB->getModule()->getContext();
540 IRBuilder<> IRB(C);
541 IRB.SetInsertPoint(InsertPt);
542
543 // if (%__THREW__.val != 0 & threwValue != 0)
544 IRB.SetInsertPoint(BB);
545 BasicBlock *ThenBB1 = BasicBlock::Create(C, "if.then1", F);
546 BasicBlock *ElseBB1 = BasicBlock::Create(C, "if.else1", F);
547 BasicBlock *EndBB1 = BasicBlock::Create(C, "if.end", F);
548 Value *ThrewCmp = IRB.CreateICmpNE(Threw, IRB.getInt32(0));
James Y Knight14359ef2019-02-01 20:44:24 +0000549 Value *ThrewValue = IRB.CreateLoad(IRB.getInt32Ty(), ThrewValueGV,
550 ThrewValueGV->getName() + ".val");
Heejin Ahnc0f18172016-09-01 21:05:15 +0000551 Value *ThrewValueCmp = IRB.CreateICmpNE(ThrewValue, IRB.getInt32(0));
552 Value *Cmp1 = IRB.CreateAnd(ThrewCmp, ThrewValueCmp, "cmp1");
553 IRB.CreateCondBr(Cmp1, ThenBB1, ElseBB1);
554
555 // %label = _testSetjmp(mem[%__THREW__.val], _setjmpTable, _setjmpTableSize);
556 // if (%label == 0)
557 IRB.SetInsertPoint(ThenBB1);
558 BasicBlock *ThenBB2 = BasicBlock::Create(C, "if.then2", F);
559 BasicBlock *EndBB2 = BasicBlock::Create(C, "if.end2", F);
560 Value *ThrewInt = IRB.CreateIntToPtr(Threw, Type::getInt32PtrTy(C),
561 Threw->getName() + ".i32p");
James Y Knight14359ef2019-02-01 20:44:24 +0000562 Value *LoadedThrew = IRB.CreateLoad(IRB.getInt32Ty(), ThrewInt,
563 ThrewInt->getName() + ".loaded");
Heejin Ahnc0f18172016-09-01 21:05:15 +0000564 Value *ThenLabel = IRB.CreateCall(
565 TestSetjmpF, {LoadedThrew, SetjmpTable, SetjmpTableSize}, "label");
566 Value *Cmp2 = IRB.CreateICmpEQ(ThenLabel, IRB.getInt32(0));
567 IRB.CreateCondBr(Cmp2, ThenBB2, EndBB2);
568
569 // emscripten_longjmp(%__THREW__.val, threwValue);
570 IRB.SetInsertPoint(ThenBB2);
571 IRB.CreateCall(EmLongjmpF, {Threw, ThrewValue});
572 IRB.CreateUnreachable();
573
Sam Clegg4791a662018-11-20 19:25:07 +0000574 // setTempRet0(threwValue);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000575 IRB.SetInsertPoint(EndBB2);
Sam Clegg4791a662018-11-20 19:25:07 +0000576 IRB.CreateCall(SetTempRet0Func, ThrewValue);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000577 IRB.CreateBr(EndBB1);
578
579 IRB.SetInsertPoint(ElseBB1);
580 IRB.CreateBr(EndBB1);
581
Sam Clegg4791a662018-11-20 19:25:07 +0000582 // longjmp_result = getTempRet0();
Heejin Ahnc0f18172016-09-01 21:05:15 +0000583 IRB.SetInsertPoint(EndBB1);
584 PHINode *LabelPHI = IRB.CreatePHI(IRB.getInt32Ty(), 2, "label");
585 LabelPHI->addIncoming(ThenLabel, EndBB2);
586
587 LabelPHI->addIncoming(IRB.getInt32(-1), ElseBB1);
588
589 // Output parameter assignment
590 Label = LabelPHI;
591 EndBB = EndBB1;
Sam Clegg4791a662018-11-20 19:25:07 +0000592 LongjmpResult = IRB.CreateCall(GetTempRet0Func, None, "longjmp_result");
Heejin Ahnc0f18172016-09-01 21:05:15 +0000593}
594
Heejin Ahnc0f18172016-09-01 21:05:15 +0000595void WebAssemblyLowerEmscriptenEHSjLj::rebuildSSA(Function &F) {
596 DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>(F).getDomTree();
597 DT.recalculate(F); // CFG has been changed
598 SSAUpdater SSA;
599 for (BasicBlock &BB : F) {
600 for (Instruction &I : BB) {
601 for (auto UI = I.use_begin(), UE = I.use_end(); UI != UE;) {
602 Use &U = *UI;
603 ++UI;
604 SSA.Initialize(I.getType(), I.getName());
605 SSA.AddAvailableValue(&BB, &I);
Heejin Ahn18c56a02019-02-04 19:13:39 +0000606 auto *User = cast<Instruction>(U.getUser());
Heejin Ahnc0f18172016-09-01 21:05:15 +0000607 if (User->getParent() == &BB)
608 continue;
609
Heejin Ahn18c56a02019-02-04 19:13:39 +0000610 if (auto *UserPN = dyn_cast<PHINode>(User))
Heejin Ahnc0f18172016-09-01 21:05:15 +0000611 if (UserPN->getIncomingBlock(U) == &BB)
612 continue;
613
614 if (DT.dominates(&I, User))
615 continue;
616 SSA.RewriteUseAfterInsertions(U);
617 }
618 }
619 }
620}
621
622bool WebAssemblyLowerEmscriptenEHSjLj::runOnModule(Module &M) {
Heejin Ahn569f0902019-01-09 23:05:21 +0000623 LLVM_DEBUG(dbgs() << "********** Lower Emscripten EH & SjLj **********\n");
624
Heejin Ahnc0f18172016-09-01 21:05:15 +0000625 LLVMContext &C = M.getContext();
626 IRBuilder<> IRB(C);
627
628 Function *SetjmpF = M.getFunction("setjmp");
629 Function *LongjmpF = M.getFunction("longjmp");
630 bool SetjmpUsed = SetjmpF && !SetjmpF->use_empty();
631 bool LongjmpUsed = LongjmpF && !LongjmpF->use_empty();
632 bool DoSjLj = EnableSjLj && (SetjmpUsed || LongjmpUsed);
633
Sam Clegg4791a662018-11-20 19:25:07 +0000634 // Declare (or get) global variables __THREW__, __threwValue, and
635 // getTempRet0/setTempRet0 function which are used in common for both
636 // exception handling and setjmp/longjmp handling
Sam Cleggb2486f12018-10-02 22:12:15 +0000637 ThrewGV = getGlobalVariableI32(M, IRB, "__THREW__");
638 ThrewValueGV = getGlobalVariableI32(M, IRB, "__threwValue");
Sam Clegg4791a662018-11-20 19:25:07 +0000639 GetTempRet0Func =
640 Function::Create(FunctionType::get(IRB.getInt32Ty(), false),
641 GlobalValue::ExternalLinkage, "getTempRet0", &M);
642 SetTempRet0Func = Function::Create(
643 FunctionType::get(IRB.getVoidTy(), IRB.getInt32Ty(), false),
644 GlobalValue::ExternalLinkage, "setTempRet0", &M);
645 GetTempRet0Func->setDoesNotThrow();
646 SetTempRet0Func->setDoesNotThrow();
Heejin Ahnc0f18172016-09-01 21:05:15 +0000647
648 bool Changed = false;
649
650 // Exception handling
651 if (EnableEH) {
652 // Register __resumeException function
653 FunctionType *ResumeFTy =
654 FunctionType::get(IRB.getVoidTy(), IRB.getInt8PtrTy(), false);
655 ResumeF = Function::Create(ResumeFTy, GlobalValue::ExternalLinkage,
656 ResumeFName, &M);
657
658 // Register llvm_eh_typeid_for function
659 FunctionType *EHTypeIDTy =
660 FunctionType::get(IRB.getInt32Ty(), IRB.getInt8PtrTy(), false);
661 EHTypeIDF = Function::Create(EHTypeIDTy, GlobalValue::ExternalLinkage,
662 EHTypeIDFName, &M);
663
664 for (Function &F : M) {
665 if (F.isDeclaration())
666 continue;
667 Changed |= runEHOnFunction(F);
668 }
669 }
670
671 // Setjmp/longjmp handling
672 if (DoSjLj) {
673 Changed = true; // We have setjmp or longjmp somewhere
674
Heejin Ahnc0f18172016-09-01 21:05:15 +0000675 if (LongjmpF) {
676 // Replace all uses of longjmp with emscripten_longjmp_jmpbuf, which is
677 // defined in JS code
678 EmLongjmpJmpbufF = Function::Create(LongjmpF->getFunctionType(),
679 GlobalValue::ExternalLinkage,
680 EmLongjmpJmpbufFName, &M);
681
682 LongjmpF->replaceAllUsesWith(EmLongjmpJmpbufF);
683 }
Heejin Ahnc0f18172016-09-01 21:05:15 +0000684
Heejin Ahn0c68a872018-11-08 22:56:26 +0000685 if (SetjmpF) {
686 // Register saveSetjmp function
687 FunctionType *SetjmpFTy = SetjmpF->getFunctionType();
688 SmallVector<Type *, 4> Params = {SetjmpFTy->getParamType(0),
689 IRB.getInt32Ty(), Type::getInt32PtrTy(C),
690 IRB.getInt32Ty()};
691 FunctionType *FTy =
692 FunctionType::get(Type::getInt32PtrTy(C), Params, false);
693 SaveSetjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage,
694 SaveSetjmpFName, &M);
695
696 // Register testSetjmp function
697 Params = {IRB.getInt32Ty(), Type::getInt32PtrTy(C), IRB.getInt32Ty()};
698 FTy = FunctionType::get(IRB.getInt32Ty(), Params, false);
699 TestSetjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage,
700 TestSetjmpFName, &M);
701
702 FTy = FunctionType::get(IRB.getVoidTy(),
703 {IRB.getInt32Ty(), IRB.getInt32Ty()}, false);
704 EmLongjmpF = Function::Create(FTy, GlobalValue::ExternalLinkage,
705 EmLongjmpFName, &M);
706
707 // Only traverse functions that uses setjmp in order not to insert
708 // unnecessary prep / cleanup code in every function
709 SmallPtrSet<Function *, 8> SetjmpUsers;
710 for (User *U : SetjmpF->users()) {
711 auto *UI = cast<Instruction>(U);
712 SetjmpUsers.insert(UI->getFunction());
713 }
714 for (Function *F : SetjmpUsers)
715 runSjLjOnFunction(*F);
Heejin Ahnc0f18172016-09-01 21:05:15 +0000716 }
Heejin Ahnc0f18172016-09-01 21:05:15 +0000717 }
718
719 if (!Changed) {
720 // Delete unused global variables and functions
Heejin Ahnc0f18172016-09-01 21:05:15 +0000721 if (ResumeF)
722 ResumeF->eraseFromParent();
723 if (EHTypeIDF)
724 EHTypeIDF->eraseFromParent();
725 if (EmLongjmpF)
726 EmLongjmpF->eraseFromParent();
727 if (SaveSetjmpF)
728 SaveSetjmpF->eraseFromParent();
729 if (TestSetjmpF)
730 TestSetjmpF->eraseFromParent();
731 return false;
732 }
733
Derek Schufff41f67d2016-08-01 21:34:04 +0000734 return true;
735}
736
Derek Schuffccdceda2016-08-18 15:27:25 +0000737bool WebAssemblyLowerEmscriptenEHSjLj::runEHOnFunction(Function &F) {
Derek Schufff41f67d2016-08-01 21:34:04 +0000738 Module &M = *F.getParent();
Derek Schuff53b9af02016-08-09 00:29:55 +0000739 LLVMContext &C = F.getContext();
Derek Schuffccdceda2016-08-18 15:27:25 +0000740 IRBuilder<> IRB(C);
Derek Schufff41f67d2016-08-01 21:34:04 +0000741 bool Changed = false;
742 SmallVector<Instruction *, 64> ToErase;
743 SmallPtrSet<LandingPadInst *, 32> LandingPads;
Derek Schuff66641322016-08-09 22:37:00 +0000744 bool AllowExceptions =
Derek Schuffccdceda2016-08-18 15:27:25 +0000745 areAllExceptionsAllowed() || EHWhitelistSet.count(F.getName());
Derek Schufff41f67d2016-08-01 21:34:04 +0000746
747 for (BasicBlock &BB : F) {
748 auto *II = dyn_cast<InvokeInst>(BB.getTerminator());
749 if (!II)
750 continue;
751 Changed = true;
752 LandingPads.insert(II->getLandingPadInst());
Derek Schuffccdceda2016-08-18 15:27:25 +0000753 IRB.SetInsertPoint(II);
Derek Schufff41f67d2016-08-01 21:34:04 +0000754
Derek Schuff53b9af02016-08-09 00:29:55 +0000755 bool NeedInvoke = AllowExceptions && canThrow(II->getCalledValue());
756 if (NeedInvoke) {
Heejin Ahnc0f18172016-09-01 21:05:15 +0000757 // Wrap invoke with invoke wrapper and generate preamble/postamble
758 Value *Threw = wrapInvoke(II);
Derek Schufff41f67d2016-08-01 21:34:04 +0000759 ToErase.push_back(II);
760
Derek Schufff41f67d2016-08-01 21:34:04 +0000761 // Insert a branch based on __THREW__ variable
Heejin Ahnc0f18172016-09-01 21:05:15 +0000762 Value *Cmp = IRB.CreateICmpEQ(Threw, IRB.getInt32(1), "cmp");
763 IRB.CreateCondBr(Cmp, II->getUnwindDest(), II->getNormalDest());
Derek Schufff41f67d2016-08-01 21:34:04 +0000764
765 } else {
766 // This can't throw, and we don't need this invoke, just replace it with a
767 // call+branch
Heejin Ahnc0f18172016-09-01 21:05:15 +0000768 SmallVector<Value *, 16> Args(II->arg_begin(), II->arg_end());
James Y Knight7976eb52019-02-01 20:43:25 +0000769 CallInst *NewCall =
770 IRB.CreateCall(II->getFunctionType(), II->getCalledValue(), Args);
Derek Schufff41f67d2016-08-01 21:34:04 +0000771 NewCall->takeName(II);
772 NewCall->setCallingConv(II->getCallingConv());
Derek Schufff41f67d2016-08-01 21:34:04 +0000773 NewCall->setDebugLoc(II->getDebugLoc());
Derek Schuff53b9af02016-08-09 00:29:55 +0000774 NewCall->setAttributes(II->getAttributes());
Derek Schufff41f67d2016-08-01 21:34:04 +0000775 II->replaceAllUsesWith(NewCall);
776 ToErase.push_back(II);
777
Derek Schuffccdceda2016-08-18 15:27:25 +0000778 IRB.CreateBr(II->getNormalDest());
Derek Schufff41f67d2016-08-01 21:34:04 +0000779
780 // Remove any PHI node entries from the exception destination
781 II->getUnwindDest()->removePredecessor(&BB);
782 }
783 }
784
785 // Process resume instructions
786 for (BasicBlock &BB : F) {
787 // Scan the body of the basic block for resumes
788 for (Instruction &I : BB) {
789 auto *RI = dyn_cast<ResumeInst>(&I);
790 if (!RI)
791 continue;
792
793 // Split the input into legal values
794 Value *Input = RI->getValue();
Derek Schuffccdceda2016-08-18 15:27:25 +0000795 IRB.SetInsertPoint(RI);
796 Value *Low = IRB.CreateExtractValue(Input, 0, "low");
Derek Schuff53b9af02016-08-09 00:29:55 +0000797 // Create a call to __resumeException function
Heejin Ahnc0f18172016-09-01 21:05:15 +0000798 IRB.CreateCall(ResumeF, {Low});
Derek Schufff41f67d2016-08-01 21:34:04 +0000799 // Add a terminator to the block
Derek Schuffccdceda2016-08-18 15:27:25 +0000800 IRB.CreateUnreachable();
Derek Schufff41f67d2016-08-01 21:34:04 +0000801 ToErase.push_back(RI);
802 }
803 }
804
Derek Schuff53b9af02016-08-09 00:29:55 +0000805 // Process llvm.eh.typeid.for intrinsics
806 for (BasicBlock &BB : F) {
807 for (Instruction &I : BB) {
808 auto *CI = dyn_cast<CallInst>(&I);
809 if (!CI)
810 continue;
811 const Function *Callee = CI->getCalledFunction();
812 if (!Callee)
813 continue;
814 if (Callee->getIntrinsicID() != Intrinsic::eh_typeid_for)
815 continue;
816
Derek Schuffccdceda2016-08-18 15:27:25 +0000817 IRB.SetInsertPoint(CI);
Derek Schuff53b9af02016-08-09 00:29:55 +0000818 CallInst *NewCI =
Derek Schuffccdceda2016-08-18 15:27:25 +0000819 IRB.CreateCall(EHTypeIDF, CI->getArgOperand(0), "typeid");
Derek Schuff53b9af02016-08-09 00:29:55 +0000820 CI->replaceAllUsesWith(NewCI);
821 ToErase.push_back(CI);
822 }
823 }
824
Hiroshi Inouec3969642017-06-30 07:17:53 +0000825 // Look for orphan landingpads, can occur in blocks with no predecessors
Derek Schufff41f67d2016-08-01 21:34:04 +0000826 for (BasicBlock &BB : F) {
827 Instruction *I = BB.getFirstNonPHI();
828 if (auto *LPI = dyn_cast<LandingPadInst>(I))
829 LandingPads.insert(LPI);
830 }
831
832 // Handle all the landingpad for this function together, as multiple invokes
833 // may share a single lp
834 for (LandingPadInst *LPI : LandingPads) {
Derek Schuffccdceda2016-08-18 15:27:25 +0000835 IRB.SetInsertPoint(LPI);
Derek Schufff41f67d2016-08-01 21:34:04 +0000836 SmallVector<Value *, 16> FMCArgs;
Heejin Ahn18c56a02019-02-04 19:13:39 +0000837 for (unsigned I = 0, E = LPI->getNumClauses(); I < E; ++I) {
838 Constant *Clause = LPI->getClause(I);
Derek Schufff41f67d2016-08-01 21:34:04 +0000839 // As a temporary workaround for the lack of aggregate varargs support
840 // in the interface between JS and wasm, break out filter operands into
841 // their component elements.
Heejin Ahn18c56a02019-02-04 19:13:39 +0000842 if (LPI->isFilter(I)) {
Heejin Ahnc0f18172016-09-01 21:05:15 +0000843 auto *ATy = cast<ArrayType>(Clause->getType());
Heejin Ahn18c56a02019-02-04 19:13:39 +0000844 for (unsigned J = 0, E = ATy->getNumElements(); J < E; ++J) {
845 Value *EV = IRB.CreateExtractValue(Clause, makeArrayRef(J), "filter");
Derek Schufff41f67d2016-08-01 21:34:04 +0000846 FMCArgs.push_back(EV);
847 }
848 } else
849 FMCArgs.push_back(Clause);
850 }
851
Derek Schuff53b9af02016-08-09 00:29:55 +0000852 // Create a call to __cxa_find_matching_catch_N function
Derek Schufff41f67d2016-08-01 21:34:04 +0000853 Function *FMCF = getFindMatchingCatch(M, FMCArgs.size());
Derek Schuffccdceda2016-08-18 15:27:25 +0000854 CallInst *FMCI = IRB.CreateCall(FMCF, FMCArgs, "fmc");
Derek Schufff41f67d2016-08-01 21:34:04 +0000855 Value *Undef = UndefValue::get(LPI->getType());
Derek Schuffccdceda2016-08-18 15:27:25 +0000856 Value *Pair0 = IRB.CreateInsertValue(Undef, FMCI, 0, "pair0");
Sam Clegg4791a662018-11-20 19:25:07 +0000857 Value *TempRet0 = IRB.CreateCall(GetTempRet0Func, None, "tempret0");
Derek Schuffccdceda2016-08-18 15:27:25 +0000858 Value *Pair1 = IRB.CreateInsertValue(Pair0, TempRet0, 1, "pair1");
Derek Schufff41f67d2016-08-01 21:34:04 +0000859
860 LPI->replaceAllUsesWith(Pair1);
861 ToErase.push_back(LPI);
862 }
863
864 // Erase everything we no longer need in this function
865 for (Instruction *I : ToErase)
866 I->eraseFromParent();
867
868 return Changed;
869}
Derek Schuffccdceda2016-08-18 15:27:25 +0000870
871bool WebAssemblyLowerEmscriptenEHSjLj::runSjLjOnFunction(Function &F) {
Heejin Ahnc0f18172016-09-01 21:05:15 +0000872 Module &M = *F.getParent();
873 LLVMContext &C = F.getContext();
874 IRBuilder<> IRB(C);
875 SmallVector<Instruction *, 64> ToErase;
876 // Vector of %setjmpTable values
877 std::vector<Instruction *> SetjmpTableInsts;
878 // Vector of %setjmpTableSize values
879 std::vector<Instruction *> SetjmpTableSizeInsts;
880
881 // Setjmp preparation
882
883 // This instruction effectively means %setjmpTableSize = 4.
884 // We create this as an instruction intentionally, and we don't want to fold
885 // this instruction to a constant 4, because this value will be used in
886 // SSAUpdater.AddAvailableValue(...) later.
887 BasicBlock &EntryBB = F.getEntryBlock();
888 BinaryOperator *SetjmpTableSize = BinaryOperator::Create(
889 Instruction::Add, IRB.getInt32(4), IRB.getInt32(0), "setjmpTableSize",
890 &*EntryBB.getFirstInsertionPt());
891 // setjmpTable = (int *) malloc(40);
892 Instruction *SetjmpTable = CallInst::CreateMalloc(
893 SetjmpTableSize, IRB.getInt32Ty(), IRB.getInt32Ty(), IRB.getInt32(40),
894 nullptr, nullptr, "setjmpTable");
895 // setjmpTable[0] = 0;
896 IRB.SetInsertPoint(SetjmpTableSize);
897 IRB.CreateStore(IRB.getInt32(0), SetjmpTable);
898 SetjmpTableInsts.push_back(SetjmpTable);
899 SetjmpTableSizeInsts.push_back(SetjmpTableSize);
900
901 // Setjmp transformation
902 std::vector<PHINode *> SetjmpRetPHIs;
903 Function *SetjmpF = M.getFunction("setjmp");
904 for (User *U : SetjmpF->users()) {
905 auto *CI = dyn_cast<CallInst>(U);
906 if (!CI)
907 report_fatal_error("Does not support indirect calls to setjmp");
908
909 BasicBlock *BB = CI->getParent();
910 if (BB->getParent() != &F) // in other function
911 continue;
912
913 // The tail is everything right after the call, and will be reached once
914 // when setjmp is called, and later when longjmp returns to the setjmp
915 BasicBlock *Tail = SplitBlock(BB, CI->getNextNode());
916 // Add a phi to the tail, which will be the output of setjmp, which
917 // indicates if this is the first call or a longjmp back. The phi directly
918 // uses the right value based on where we arrive from
919 IRB.SetInsertPoint(Tail->getFirstNonPHI());
920 PHINode *SetjmpRet = IRB.CreatePHI(IRB.getInt32Ty(), 2, "setjmp.ret");
921
922 // setjmp initial call returns 0
923 SetjmpRet->addIncoming(IRB.getInt32(0), BB);
924 // The proper output is now this, not the setjmp call itself
925 CI->replaceAllUsesWith(SetjmpRet);
926 // longjmp returns to the setjmp will add themselves to this phi
927 SetjmpRetPHIs.push_back(SetjmpRet);
928
929 // Fix call target
930 // Our index in the function is our place in the array + 1 to avoid index
931 // 0, because index 0 means the longjmp is not ours to handle.
932 IRB.SetInsertPoint(CI);
933 Value *Args[] = {CI->getArgOperand(0), IRB.getInt32(SetjmpRetPHIs.size()),
934 SetjmpTable, SetjmpTableSize};
935 Instruction *NewSetjmpTable =
936 IRB.CreateCall(SaveSetjmpF, Args, "setjmpTable");
937 Instruction *NewSetjmpTableSize =
Sam Clegg4791a662018-11-20 19:25:07 +0000938 IRB.CreateCall(GetTempRet0Func, None, "setjmpTableSize");
Heejin Ahnc0f18172016-09-01 21:05:15 +0000939 SetjmpTableInsts.push_back(NewSetjmpTable);
940 SetjmpTableSizeInsts.push_back(NewSetjmpTableSize);
941 ToErase.push_back(CI);
942 }
943
944 // Update each call that can longjmp so it can return to a setjmp where
945 // relevant.
946
947 // Because we are creating new BBs while processing and don't want to make
948 // all these newly created BBs candidates again for longjmp processing, we
949 // first make the vector of candidate BBs.
950 std::vector<BasicBlock *> BBs;
951 for (BasicBlock &BB : F)
952 BBs.push_back(&BB);
953
954 // BBs.size() will change within the loop, so we query it every time
Heejin Ahn18c56a02019-02-04 19:13:39 +0000955 for (unsigned I = 0; I < BBs.size(); I++) {
956 BasicBlock *BB = BBs[I];
Heejin Ahnc0f18172016-09-01 21:05:15 +0000957 for (Instruction &I : *BB) {
958 assert(!isa<InvokeInst>(&I));
959 auto *CI = dyn_cast<CallInst>(&I);
960 if (!CI)
961 continue;
962
963 const Value *Callee = CI->getCalledValue();
964 if (!canLongjmp(M, Callee))
965 continue;
966
967 Value *Threw = nullptr;
968 BasicBlock *Tail;
969 if (Callee->getName().startswith(InvokePrefix)) {
970 // If invoke wrapper has already been generated for this call in
971 // previous EH phase, search for the load instruction
972 // %__THREW__.val = __THREW__;
973 // in postamble after the invoke wrapper call
974 LoadInst *ThrewLI = nullptr;
975 StoreInst *ThrewResetSI = nullptr;
976 for (auto I = std::next(BasicBlock::iterator(CI)), IE = BB->end();
977 I != IE; ++I) {
978 if (auto *LI = dyn_cast<LoadInst>(I))
979 if (auto *GV = dyn_cast<GlobalVariable>(LI->getPointerOperand()))
980 if (GV == ThrewGV) {
981 Threw = ThrewLI = LI;
982 break;
983 }
984 }
985 // Search for the store instruction after the load above
986 // __THREW__ = 0;
987 for (auto I = std::next(BasicBlock::iterator(ThrewLI)), IE = BB->end();
988 I != IE; ++I) {
989 if (auto *SI = dyn_cast<StoreInst>(I))
990 if (auto *GV = dyn_cast<GlobalVariable>(SI->getPointerOperand()))
991 if (GV == ThrewGV && SI->getValueOperand() == IRB.getInt32(0)) {
992 ThrewResetSI = SI;
993 break;
994 }
995 }
996 assert(Threw && ThrewLI && "Cannot find __THREW__ load after invoke");
997 assert(ThrewResetSI && "Cannot find __THREW__ store after invoke");
998 Tail = SplitBlock(BB, ThrewResetSI->getNextNode());
999
1000 } else {
1001 // Wrap call with invoke wrapper and generate preamble/postamble
1002 Threw = wrapInvoke(CI);
1003 ToErase.push_back(CI);
1004 Tail = SplitBlock(BB, CI->getNextNode());
1005 }
1006
1007 // We need to replace the terminator in Tail - SplitBlock makes BB go
1008 // straight to Tail, we need to check if a longjmp occurred, and go to the
1009 // right setjmp-tail if so
1010 ToErase.push_back(BB->getTerminator());
1011
1012 // Generate a function call to testSetjmp function and preamble/postamble
1013 // code to figure out (1) whether longjmp occurred (2) if longjmp
1014 // occurred, which setjmp it corresponds to
1015 Value *Label = nullptr;
1016 Value *LongjmpResult = nullptr;
1017 BasicBlock *EndBB = nullptr;
1018 wrapTestSetjmp(BB, CI, Threw, SetjmpTable, SetjmpTableSize, Label,
1019 LongjmpResult, EndBB);
1020 assert(Label && LongjmpResult && EndBB);
1021
1022 // Create switch instruction
1023 IRB.SetInsertPoint(EndBB);
1024 SwitchInst *SI = IRB.CreateSwitch(Label, Tail, SetjmpRetPHIs.size());
1025 // -1 means no longjmp happened, continue normally (will hit the default
1026 // switch case). 0 means a longjmp that is not ours to handle, needs a
1027 // rethrow. Otherwise the index is the same as the index in P+1 (to avoid
1028 // 0).
Heejin Ahn18c56a02019-02-04 19:13:39 +00001029 for (unsigned I = 0; I < SetjmpRetPHIs.size(); I++) {
1030 SI->addCase(IRB.getInt32(I + 1), SetjmpRetPHIs[I]->getParent());
1031 SetjmpRetPHIs[I]->addIncoming(LongjmpResult, EndBB);
Heejin Ahnc0f18172016-09-01 21:05:15 +00001032 }
1033
1034 // We are splitting the block here, and must continue to find other calls
1035 // in the block - which is now split. so continue to traverse in the Tail
1036 BBs.push_back(Tail);
1037 }
1038 }
1039
1040 // Erase everything we no longer need in this function
1041 for (Instruction *I : ToErase)
1042 I->eraseFromParent();
1043
1044 // Free setjmpTable buffer before each return instruction
1045 for (BasicBlock &BB : F) {
Chandler Carruthedb12a82018-10-15 10:04:59 +00001046 Instruction *TI = BB.getTerminator();
Heejin Ahnc0f18172016-09-01 21:05:15 +00001047 if (isa<ReturnInst>(TI))
1048 CallInst::CreateFree(SetjmpTable, TI);
1049 }
1050
1051 // Every call to saveSetjmp can change setjmpTable and setjmpTableSize
1052 // (when buffer reallocation occurs)
1053 // entry:
1054 // setjmpTableSize = 4;
1055 // setjmpTable = (int *) malloc(40);
1056 // setjmpTable[0] = 0;
1057 // ...
1058 // somebb:
1059 // setjmpTable = saveSetjmp(buf, label, setjmpTable, setjmpTableSize);
Sam Clegg4791a662018-11-20 19:25:07 +00001060 // setjmpTableSize = getTempRet0();
Heejin Ahnc0f18172016-09-01 21:05:15 +00001061 // So we need to make sure the SSA for these variables is valid so that every
1062 // saveSetjmp and testSetjmp calls have the correct arguments.
1063 SSAUpdater SetjmpTableSSA;
1064 SSAUpdater SetjmpTableSizeSSA;
1065 SetjmpTableSSA.Initialize(Type::getInt32PtrTy(C), "setjmpTable");
1066 SetjmpTableSizeSSA.Initialize(Type::getInt32Ty(C), "setjmpTableSize");
1067 for (Instruction *I : SetjmpTableInsts)
1068 SetjmpTableSSA.AddAvailableValue(I->getParent(), I);
1069 for (Instruction *I : SetjmpTableSizeInsts)
1070 SetjmpTableSizeSSA.AddAvailableValue(I->getParent(), I);
1071
1072 for (auto UI = SetjmpTable->use_begin(), UE = SetjmpTable->use_end();
1073 UI != UE;) {
1074 // Grab the use before incrementing the iterator.
1075 Use &U = *UI;
1076 // Increment the iterator before removing the use from the list.
1077 ++UI;
Heejin Ahn18c56a02019-02-04 19:13:39 +00001078 if (auto *I = dyn_cast<Instruction>(U.getUser()))
Heejin Ahnc0f18172016-09-01 21:05:15 +00001079 if (I->getParent() != &EntryBB)
1080 SetjmpTableSSA.RewriteUse(U);
1081 }
1082 for (auto UI = SetjmpTableSize->use_begin(), UE = SetjmpTableSize->use_end();
1083 UI != UE;) {
1084 Use &U = *UI;
1085 ++UI;
Heejin Ahn18c56a02019-02-04 19:13:39 +00001086 if (auto *I = dyn_cast<Instruction>(U.getUser()))
Heejin Ahnc0f18172016-09-01 21:05:15 +00001087 if (I->getParent() != &EntryBB)
1088 SetjmpTableSizeSSA.RewriteUse(U);
1089 }
1090
1091 // Finally, our modifications to the cfg can break dominance of SSA variables.
1092 // For example, in this code,
1093 // if (x()) { .. setjmp() .. }
1094 // if (y()) { .. longjmp() .. }
1095 // We must split the longjmp block, and it can jump into the block splitted
1096 // from setjmp one. But that means that when we split the setjmp block, it's
1097 // first part no longer dominates its second part - there is a theoretically
1098 // possible control flow path where x() is false, then y() is true and we
1099 // reach the second part of the setjmp block, without ever reaching the first
1100 // part. So, we rebuild SSA form here.
1101 rebuildSSA(F);
1102 return true;
Derek Schuffccdceda2016-08-18 15:27:25 +00001103}