Skip to content

Commit

Permalink
Handle IntrinsicInst uses in allocation optimization pass
Browse files Browse the repository at this point in the history 
This is needed on LLVM 5.0.
  • Loading branch information
@yuyichao
yuyichao committed Sep 17, 2017
1 parent 3b45cdc commit 625bd9a
Show file tree
Hide file tree
Showing 2 changed files with 100 additions and 23 deletions.
62 changes: 59 additions & 3 deletions src/llvm-alloc-opt.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -150,6 +150,7 @@ struct AllocOpt : public FunctionPass {
bool checkInst(Instruction *I, CheckInstStack &stack, std::set<Instruction*> &uses,
bool &ignore_tag);
void replaceUsesWith(Instruction *orig_i, Instruction *new_i, ReplaceUsesStack &stack);
void replaceIntrinsicUseWith(IntrinsicInst *call, Instruction *orig_i, Instruction *new_i);
bool isSafepoint(Instruction *inst);
void getAnalysisUsage(AnalysisUsage &AU) const override
{
Expand Down Expand Up @@ -381,9 +382,10 @@ bool AllocOpt::checkInst(Instruction *I, CheckInstStack &stack, std::set<Instruc
if (isa<LoadInst>(inst))
return true;
if (auto call = dyn_cast<CallInst>(inst)) {
// TODO: on LLVM 5.0 we may need to handle certain llvm intrinsics
// including `memcpy`, `memset` etc. We might also need to handle
// `memcmp` by coverting to our own intrinsic and lower it after the gc root pass.
// TODO handle `memcmp`
// None of the intrinsics should care if the memory is stack or heap allocated.
if (isa<IntrinsicInst>(call))
return true;
if (ptr_from_objref && ptr_from_objref == call->getCalledFunction())
return true;
auto opno = use->getOperandNo();
Expand Down Expand Up @@ -443,6 +445,56 @@ bool AllocOpt::checkInst(Instruction *I, CheckInstStack &stack, std::set<Instruc
}
}

void AllocOpt::replaceIntrinsicUseWith(IntrinsicInst *call, Instruction *orig_i,
Instruction *new_i)
{
Intrinsic::ID ID = call->getIntrinsicID();
assert(ID);
auto nargs = call->getNumArgOperands();
SmallVector<Value*, 8> args(nargs);
SmallVector<Type*, 8> argTys(nargs);
for (unsigned i = 0; i < nargs; i++) {
auto arg = call->getArgOperand(i);
args[i] = arg == orig_i ? new_i : arg;
argTys[i] = args[i]->getType();
}

// Accumulate an array of overloaded types for the given intrinsic
SmallVector<Type*, 4> overloadTys;
{
SmallVector<Intrinsic::IITDescriptor, 8> Table;
getIntrinsicInfoTableEntries(ID, Table);
ArrayRef<Intrinsic::IITDescriptor> TableRef = Table;
auto oldfType = call->getFunctionType();
bool res = Intrinsic::matchIntrinsicType(oldfType->getReturnType(), TableRef, overloadTys);
assert(!res);
for (auto Ty : argTys) {
res = Intrinsic::matchIntrinsicType(Ty, TableRef, overloadTys);
assert(!res);
}
res = Intrinsic::matchIntrinsicVarArg(oldfType->isVarArg(), TableRef);
assert(!res);
(void)res;
}
auto newF = Intrinsic::getDeclaration(call->getModule(), ID, overloadTys);
newF->setCallingConv(call->getCallingConv());
auto newCall = CallInst::Create(newF, args, "", call);
newCall->setTailCallKind(call->getTailCallKind());
auto old_attrs = call->getAttributes();
#if JL_LLVM_VERSION >= 50000
newCall->setAttributes(AttributeList::get(*ctx, old_attrs.getFnAttributes(),
old_attrs.getRetAttributes(), {}));
#else
AttributeSet attr;
attr = attr.addAttributes(*ctx, AttributeSet::ReturnIndex, old_attrs.getRetAttributes())
.addAttributes(*ctx, AttributeSet::FunctionIndex, old_attrs.getFnAttributes());
newCall->setAttributes(attr);
#endif
newCall->setDebugLoc(call->getDebugLoc());
call->replaceAllUsesWith(newCall);
call->eraseFromParent();
}

// This function needs to handle all cases `AllocOpt::checkInst` can handle.
// This function should not erase any safepoint so that the lifetime marker can find and cache
// all the original safepoints.
Expand Down Expand Up @@ -496,6 +548,10 @@ void AllocOpt::replaceUsesWith(Instruction *orig_inst, Instruction *new_inst,
call->eraseFromParent();
return;
}
if (auto intrinsic = dyn_cast<IntrinsicInst>(call)) {
replaceIntrinsicUseWith(intrinsic, orig_i, new_i);
return;
}
// remove from operand bundle
Type *new_t = new_i->getType();
user->replaceUsesOfWith(orig_i, ConstantPointerNull::get(cast<PointerType>(new_t)));
Expand Down
61 changes: 41 additions & 20 deletions test/llvmpasses/alloc-opt.jl
View file
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,7 @@ define %jl_value_t addrspace(10)* @return_obj() {
# CHECK: alloca i64, align 8
# CHECK-NOT: @julia.gc_alloc_obj
# CHECK-NOT: @jl_gc_pool_alloc
# CHECK: call void @llvm.lifetime.start(i64 8, i8*
# CHECK: call void @llvm.lifetime.start{{.*}}(i64 8, i8*
# CHECK-NEXT: %v64 = bitcast %jl_value_t* %v to i64*
# CHECK-NOT: @tag
# CHECK-NOT: @llvm.lifetime.end
Expand All @@ -49,7 +49,7 @@ define i64 @return_load(i64 %i) {
# CHECK: call %jl_value_t*** @jl_get_ptls_states()
# CHECK-NOT: @julia.gc_alloc_obj
# CHECK-NOT: @jl_gc_pool_alloc
# CHECK: call void @llvm.lifetime.start(i64 16, i8*
# CHECK: call void @llvm.lifetime.start{{.*}}(i64 16, i8*
# CHECK: store %jl_value_t addrspace(10)* @tag, %jl_value_t addrspace(10)** {{.*}}, !tbaa !0
# CHECK-NOT: @llvm.lifetime.end
println("""
Expand Down Expand Up @@ -88,20 +88,20 @@ define void @ccall_obj(i8* %fptr) {
# CHECK: call %jl_value_t*** @jl_get_ptls_states()
# CHECK-NOT: @julia.gc_alloc_obj
# CHECK-NOT: @jl_gc_pool_alloc
# CHECK: call void @llvm.lifetime.start(i64 8, i8*
# CHECK-NEXT: %f = bitcast i8* %fptr to void (%jl_value_t*)*
# CHECK: call void @llvm.lifetime.start{{.*}}(i64 8, i8*
# CHECK: %f = bitcast i8* %fptr to void (i64)*
# Currently the GC frame lowering pass strips away all operand bundles
# CHECK-NEXT: call void %f(%jl_value_t* %v)
# CHECK-NEXT: call void %f(i64
# CHECK-NEXT: ret void
println("""
define void @ccall_ptr(i8* %fptr) {
%ptls = call %jl_value_t*** @jl_get_ptls_states()
%ptls_i8 = bitcast %jl_value_t*** %ptls to i8*
%v = call noalias %jl_value_t addrspace(10)* @julia.gc_alloc_obj(i8* %ptls_i8, $isz 8, %jl_value_t addrspace(10)* @tag)
%va = addrspacecast %jl_value_t addrspace(10)* %v to %jl_value_t addrspace(11)*
%ptr = call %jl_value_t* @julia.pointer_from_objref(%jl_value_t addrspace(11)* %va)
%f = bitcast i8* %fptr to void (%jl_value_t*)*
call void %f(%jl_value_t* %ptr) [ "jl_roots"(%jl_value_t addrspace(10)* %v), "unknown_bundle"(%jl_value_t* %ptr) ]
%ptr = call i64 @julia.pointer_from_objref(%jl_value_t addrspace(11)* %va)
%f = bitcast i8* %fptr to void (i64)*
call void %f(i64 %ptr) [ "jl_roots"(%jl_value_t addrspace(10)* %v), "unknown_bundle"(i64 %ptr) ]
ret void
}
""")
Expand All @@ -118,9 +118,9 @@ define void @ccall_unknown_bundle(i8* %fptr) {
%ptls_i8 = bitcast %jl_value_t*** %ptls to i8*
%v = call noalias %jl_value_t addrspace(10)* @julia.gc_alloc_obj(i8* %ptls_i8, $isz 8, %jl_value_t addrspace(10)* @tag)
%va = addrspacecast %jl_value_t addrspace(10)* %v to %jl_value_t addrspace(11)*
%ptr = call %jl_value_t* @julia.pointer_from_objref(%jl_value_t addrspace(11)* %va)
%f = bitcast i8* %fptr to void (%jl_value_t*)*
call void %f(%jl_value_t* %ptr) [ "jl_not_jl_roots"(%jl_value_t addrspace(10)* %v) ]
%ptr = call i64 @julia.pointer_from_objref(%jl_value_t addrspace(11)* %va)
%f = bitcast i8* %fptr to void (i64)*
call void %f(i64 %ptr) [ "jl_not_jl_roots"(%jl_value_t addrspace(10)* %v) ]
ret void
}
""")
Expand All @@ -130,18 +130,18 @@ define void @ccall_unknown_bundle(i8* %fptr) {
# CHECK: alloca i64
# CHECK: call %jl_value_t*** @jl_get_ptls_states()
# CHECK: L1:
# CHECK-NEXT: call void @llvm.lifetime.start(i64 8,
# CHECK-NEXT: %f = bitcast i8* %fptr to void (%jl_value_t*)*
# CHECK-NEXT: call void %f(%jl_value_t* %v)
# CHECK-NEXT: call void @llvm.lifetime.start{{.*}}(i64 8,
# CHECK: %f = bitcast i8* %fptr to void (i64)*
# CHECK-NEXT: call void %f(i64
# CHECK-NEXT: br i1 %b2, label %L2, label %L3

# CHECK: L2:
# CHECK-NEXT: %f2 = bitcast i8* %fptr to void (%jl_value_t*)*
# CHECK-NEXT: call void @llvm.lifetime.end(i64 8,
# CHECK-NEXT: call void @llvm.lifetime.end{{.*}}(i64 8,
# CHECK-NEXT: call void %f2(%jl_value_t* null)

# CHECK: L3:
# CHECK-NEXT: call void @llvm.lifetime.end(i64 8,
# CHECK-NEXT: call void @llvm.lifetime.end{{.*}}(i64 8,
println("""
define void @lifetime_branches(i8* %fptr, i1 %b, i1 %b2) {
%ptls = call %jl_value_t*** @jl_get_ptls_states()
Expand All @@ -151,9 +151,9 @@ define void @lifetime_branches(i8* %fptr, i1 %b, i1 %b2) {
L1:
%v = call noalias %jl_value_t addrspace(10)* @julia.gc_alloc_obj(i8* %ptls_i8, $isz 8, %jl_value_t addrspace(10)* @tag)
%va = addrspacecast %jl_value_t addrspace(10)* %v to %jl_value_t addrspace(11)*
%ptr = call %jl_value_t* @julia.pointer_from_objref(%jl_value_t addrspace(11)* %va)
%f = bitcast i8* %fptr to void (%jl_value_t*)*
call void %f(%jl_value_t* %ptr) [ "jl_roots"(%jl_value_t addrspace(10)* %v) ]
%ptr = call i64 @julia.pointer_from_objref(%jl_value_t addrspace(11)* %va)
%f = bitcast i8* %fptr to void (i64)*
call void %f(i64 %ptr) [ "jl_roots"(%jl_value_t addrspace(10)* %v) ]
br i1 %b2, label %L2, label %L3
L2:
Expand Down Expand Up @@ -185,13 +185,34 @@ define void @object_field(%jl_value_t addrspace(10)* %field) {
""")
# CHECK-LABEL: }

# CHECK-LABEL: @memcpy_opt
# CHECK: alloca i128, align 16
# CHECK: call %jl_value_t*** @jl_get_ptls_states()
# CHECK-NOT: @julia.gc_alloc_obj
# CHECK-NOT: @jl_gc_pool_alloc
# CHECK: call void @llvm.memcpy.p0i8.p0i8.i64
println("""
define void @memcpy_opt(i8* %v22) {
top:
%v6 = call %jl_value_t*** @jl_get_ptls_states()
%v18 = bitcast %jl_value_t*** %v6 to i8*
%v19 = call noalias %jl_value_t addrspace(10)* @julia.gc_alloc_obj(i8* %v18, $isz 16, %jl_value_t addrspace(10)* @tag)
%v20 = bitcast %jl_value_t addrspace(10)* %v19 to i8 addrspace(10)*
%v21 = addrspacecast i8 addrspace(10)* %v20 to i8 addrspace(11)*
call void @llvm.memcpy.p11i8.p0i8.i64(i8 addrspace(11)* %v21, i8* %v22, i64 16, i32 8, i1 false)
ret void
}
""")
# CHECK-LABEL: }

# CHECK: declare noalias %jl_value_t addrspace(10)* @jl_gc_pool_alloc(i8*,
# CHECK: declare noalias %jl_value_t addrspace(10)* @jl_gc_big_alloc(i8*,
println("""
declare void @external_function()
declare %jl_value_t*** @jl_get_ptls_states()
declare noalias %jl_value_t addrspace(10)* @julia.gc_alloc_obj(i8*, $isz, %jl_value_t addrspace(10)*)
declare %jl_value_t* @julia.pointer_from_objref(%jl_value_t addrspace(11)*)
declare i64 @julia.pointer_from_objref(%jl_value_t addrspace(11)*)
declare void @llvm.memcpy.p11i8.p0i8.i64(i8 addrspace(11)* nocapture writeonly, i8* nocapture readonly, i64, i32, i1)
!0 = !{!1, !1, i64 0}
!1 = !{!"jtbaa_tag", !2, i64 0}
Expand Down

0 comments on commit 625bd9a

Please sign in to comment.