Fix lifetime encoding to match the new langref wording #686
Conversation
aqjune
changed the title
|
So how does this handle the acausal nature of |
|
Hmm, I think we should raise unsupported instruction if the program contains a 'bad-looking' lifetimes. A lifetime call is good-looking if it exactly determines which alloca must be dead in the beginning. |
|
|
||
| Value& getSize() const { return *size; } | ||
| Value* getMul() const { return mul; } | ||
| bool initDead() const { return initially_dead; } | ||
| void markAsInitiallyDead() { initially_dead = true; } |
There was a problem hiding this comment.
Previously llvm2alive.cpp's lifetime.star/end encoding was splitted into two functions: hasLifetimeStart (which checks AllocaInst's uses) and visitIntrinsicInst (which adds StartLifetime instance)
I removed hasLifetimeStart, and made visitIntrinsicInst to update this initially-dead flag if lifetime.start uses the alloca.
This markAsInitiallyDead function is added to implement it.
| Pointer p(*this, bid, expr::mkUInt(0, bits_for_offset)); | ||
| expr blksz = p.blockSize(); | ||
| memset(p.release(), IntType("i8", 8).getDummyValue(false), | ||
| move(blksz), bits_byte / 8, {}, false); |
There was a problem hiding this comment.
I guess this encoding can be optimized, but fillPoison() is called only when there are multiple lifetime.start/ends, or they are taking non-stack-allocated objects, etc. These cases are rare.
| hasLifetimeStart(i, visited))); | ||
| auto alloc = make_unique<Alloc>(*ty, value_name(i), *size, mul, | ||
| pref_alignment(i, i.getAllocatedType())); | ||
| allocs.emplace(&i, make_pair(alloc.get(), /*has lifetime.start?*/ false)); |
There was a problem hiding this comment.
This flag will be turned on by getLifetimeKind if it finds lifetime.start with a pointer to this alloca.
| @@ -12,3 +23,4 @@ define void @f() { | |||
| unreachable | |||
| } | |||
|
|
|||
| ; ERROR: Source is more defined than target | |||
There was a problem hiding this comment.
This is a non-functional change; just merges lifetime-noub.src.ll/.tgt.ll into lifetime-noub.srctgt.ll
aqjune
changed the title
|
Now is the time to run LLVM unit tests.. It takes about 3 hrs, I'll leave the result here when it's done. |
|
No regression from LLVM unit tests |
Fixes #433
Conservatively, this patch still raises unsupported feature error if lifetime's pointer has alloca/non-alloca objects mixed.