[LLVM 21] fix calls to llvm.lifetime.start
#4979
Conversation
thewilsonator
force-pushed
the
lifetime
branch
from
aacda96 to
812d365
Compare
|
Did you check whether those lifetime calls are there for LLVM 20? (in the exact same functions?) Would help to get a standalone smaller testcase btw. Something like this: https://d.godbolt.org/z/d6f9763KE |
|
I think I found the extremely weird answer in the clang source (see #4990) - the calls still feature the size argument, although the declaration has a single pointer parameter: https://github.com/llvm/llvm-project/blob/220bac16a417e97bf97fdcb34855e28b2e6dfdf7/clang/lib/CodeGen/CGDecl.cpp#L1363 |
|
that is weird! |
thewilsonator
force-pushed
the
lifetime
branch
from
812d365 to
9ebd597
Compare
|
reduced to struct CodepointSet
{
CodepointSet opBinary(string op, U)(U )
{
return this;
}
this(this) { }
}
CodepointSet memoizeExpr(string expr)()
{
if (__ctfe)
return mixin(expr);
CodepointSet slot;
return slot;
}
void wordCharacter() {
memoizeExpr!"unicode.A | unicode.M";
}
struct unicode
{
static CodepointSet opDispatch(string name)()
{
CodepointSet set;
return set;
}
}it generates a broken call ; Function Attrs: uwtable(sync)
define weak_odr void @_D7package__T11memoizeExprVAyaa21_756e69636f64652e41207c20756e69636f64652e4dZQCmFNaNbNiNfZSQDl__T13InversionListTSQEi8GcPolicyZQBe(ptr noalias sret(%"package.InversionList!(GcPolicy).InversionList") align 1 %.sret_arg) #0 {
%slot = alloca %"package.InversionList!(GcPolicy).InversionList", align 1
%__copytmp3 = alloca %"package.InversionList!(GcPolicy).InversionList", align 1
call void @llvm.lifetime.start.p0(ptr captures(none) %slot) #2
call void @llvm.memset.p0.i64(ptr align 1 %slot, i8 0, i64 1, i1 false)
call void @llvm.lifetime.start.p0(ptr captures(none) %.sret_arg) #2 ; <<<<<< HERE
call void @llvm.memcpy.p0.p0.i64(ptr align 1 %.sret_arg, ptr align 1 %slot, i64 1, i1 false)
call void @_D7package__T13InversionListTSQBc8GcPolicyZQBe15__fieldPostblitMFNaNbNiNeZv(ptr nonnull %.sret_arg) #0
ret void
} |
|
53d7c1d is not the correct fix, and fails when built with recent LLVM. and removing the early return from if (lltype->isVoidTy()) {
irLocal->value = getNullPtr();
return;
}does not fix it either. I think this is a DMD bug where if (__ctfe)
return mixin(expr);
alias T = typeof(mixin(expr));
T slot;
return slot; // This NRVO is broken
}the |
thewilsonator
force-pushed
the
lifetime
branch
from
9ebd597 to
69b3f0e
Compare
|
Hopefully this is now fixed. |
thewilsonator
force-pushed
the
lifetime
branch
from
69b3f0e to
5424e87
Compare
thewilsonator
force-pushed
the
lifetime
branch
from
5424e87 to
0e742c1
Compare
3 participants
This mostly fixes compilation for LLVM21, except for some breakage with regards to moving
llvm.lifetime.startto taking only andAllocaInst(or aPoison).and
(the above ` | grep "%0")
which both produce a call to
@llvm.lifetime.start.p0(ptr captures(none) %0)and%0in this case is anArgumentnot anAllocaInst.I'm trying to figure out where those calls are coming from. Is it the ABI messing things up somehow? I don't know what is special about those functions.
there is only the one use of it and