Skip to content

8385746: S390: Improve receiver type profiling reliability #31510

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
offamitkumar wants to merge 2 commits into
base: master
Choose a base branch
from
Open

8385746: S390: Improve receiver type profiling reliability #31510

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2aca63b
init
offamitkumar Jun 12, 2026
ee85677
revert mistakenly deleted line
offamitkumar Jun 15, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion src/hotspot/cpu/ppc/macroAssembler_ppc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4450,7 +4450,7 @@ void MacroAssembler::profile_receiver_type(Register recv, Register mdp, int mdp_
addi(offset, offset, receiver_step);
bdnz(L_loop_search_receiver);

// Fast: no receiver, but profile is full
// Fast: no receiver, but profile is not full
if (count != noreg) {
mtctr(count);
} else {
Expand Down
73 changes: 10 additions & 63 deletions src/hotspot/cpu/s390/c1_LIRAssembler_s390.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2413,32 +2413,9 @@ void LIR_Assembler::emit_alloc_array(LIR_OpAllocArray* op) {
}

void LIR_Assembler::type_profile_helper(Register mdo, ciMethodData *md, ciProfileData *data,
Register recv, Register tmp1, Label* update_done) {
uint i;
for (i = 0; i < VirtualCallData::row_limit(); i++) {
Label next_test;
// See if the receiver is receiver[n].
Address receiver_addr(mdo, md->byte_offset_of_slot(data, ReceiverTypeData::receiver_offset(i)));
__ z_cg(recv, receiver_addr);
__ z_brne(next_test);
Address data_addr(mdo, md->byte_offset_of_slot(data, ReceiverTypeData::receiver_count_offset(i)));
__ add2mem_64(data_addr, DataLayout::counter_increment, tmp1);
__ branch_optimized(Assembler::bcondAlways, *update_done);
__ bind(next_test);
}

// Didn't find receiver; find next empty slot and fill it in.
for (i = 0; i < VirtualCallData::row_limit(); i++) {
Label next_test;
Address recv_addr(mdo, md->byte_offset_of_slot(data, ReceiverTypeData::receiver_offset(i)));
__ z_ltg(Z_R0_scratch, recv_addr);
__ z_brne(next_test);
__ z_stg(recv, recv_addr);
__ load_const_optimized(tmp1, DataLayout::counter_increment);
__ z_stg(tmp1, md->byte_offset_of_slot(data, ReceiverTypeData::receiver_count_offset(i)), mdo);
__ branch_optimized(Assembler::bcondAlways, *update_done);
__ bind(next_test);
}
Register recv, Register tmp1) {
int mdp_offset = md->byte_offset_of_slot(data, in_ByteSize(0));
__ profile_receiver_type(recv, mdo, mdp_offset, tmp1);
}

void LIR_Assembler::setup_md_access(ciMethod* method, int bci,
Expand Down Expand Up @@ -2510,13 +2487,9 @@ void LIR_Assembler::emit_typecheck_helper(LIR_OpTypeCheck *op, Label* success, L
__ branch_optimized(Assembler::bcondAlways, *obj_is_null);
__ bind(not_null);

NearLabel update_done;
Register recv = k_RInfo;
__ load_klass(recv, obj);
type_profile_helper(mdo, md, data, recv, Rtmp1, &update_done);
Address counter_addr(mdo, md->byte_offset_of_slot(data, CounterData::count_offset()));
__ add2mem_64(counter_addr, DataLayout::counter_increment, Rtmp1);
__ bind(update_done);
type_profile_helper(mdo, md, data, recv, Rtmp1);
} else {
__ compareU64_and_branch(obj, (intptr_t) 0, Assembler::bcondEqual, *obj_is_null);
}
Expand Down Expand Up @@ -2606,13 +2579,9 @@ void LIR_Assembler::emit_opTypeCheck(LIR_OpTypeCheck* op) {
__ branch_optimized(Assembler::bcondAlways, done);
__ bind(not_null);

NearLabel update_done;
Register recv = k_RInfo;
__ load_klass(recv, value);
type_profile_helper(mdo, md, data, recv, Rtmp1, &update_done);
Address counter_addr(mdo, md->byte_offset_of_slot(data, CounterData::count_offset()));
__ add2mem_64(counter_addr, DataLayout::counter_increment, Rtmp1);
__ bind(update_done);
type_profile_helper(mdo, md, data, recv, Rtmp1);
} else {
__ compareU64_and_branch(value, (intptr_t) 0, Assembler::bcondEqual, done);
}
Expand Down Expand Up @@ -2772,44 +2741,22 @@ void LIR_Assembler::emit_profile_call(LIR_OpProfileCall* op) {
// statically update the MethodData* rather than needing to do
// dynamic tests on the receiver type.

// NOTE: we should probably put a lock around this search to
// avoid collisions by concurrent compilations.
ciVirtualCallData* vc_data = (ciVirtualCallData*) data;
uint i;
for (i = 0; i < VirtualCallData::row_limit(); i++) {
for (uint i = 0; i < VirtualCallData::row_limit(); i++) {
ciKlass* receiver = vc_data->receiver(i);
if (known_klass->equals(receiver)) {
Address data_addr(mdo, md->byte_offset_of_slot(data, VirtualCallData::receiver_count_offset(i)));
__ add2mem_64(data_addr, DataLayout::counter_increment, tmp1);
return;
}
}

// Receiver type not found in profile data. Select an empty slot.

// Note that this is less efficient than it should be because it
// always does a write to the receiver part of the
// VirtualCallData rather than just the first time.
for (i = 0; i < VirtualCallData::row_limit(); i++) {
ciKlass* receiver = vc_data->receiver(i);
if (receiver == nullptr) {
Address recv_addr(mdo, md->byte_offset_of_slot(data, VirtualCallData::receiver_offset(i)));
metadata2reg(known_klass->constant_encoding(), tmp1);
__ z_stg(tmp1, recv_addr);
Address data_addr(mdo, md->byte_offset_of_slot(data, VirtualCallData::receiver_count_offset(i)));
__ add2mem_64(data_addr, DataLayout::counter_increment, tmp1);
return;
}
}
// Receiver type is not found in profile data.
// Fall back to runtime helper to handle the rest at runtime.
metadata2reg(known_klass->constant_encoding(), recv);
} else {
__ load_klass(recv, recv);
NearLabel update_done;
type_profile_helper(mdo, md, data, recv, tmp1, &update_done);
// Receiver did not match any saved receiver and there is no empty row for it.
// Increment total counter to indicate polymorphic case.
__ add2mem_64(counter_addr, DataLayout::counter_increment, tmp1);
__ bind(update_done);
}
type_profile_helper(mdo, md, data, recv, tmp1);
} else {
// static call
__ add2mem_64(counter_addr, DataLayout::counter_increment, tmp1);
Expand Down
2 changes: 1 addition & 1 deletion src/hotspot/cpu/s390/c1_LIRAssembler_s390.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,7 +30,7 @@

// Record the type of the receiver in ReceiverTypeData.
void type_profile_helper(Register mdo, ciMethodData *md, ciProfileData *data,
Register recv, Register tmp1, Label* update_done);
Register recv, Register tmp1);
// Setup pointers to MDO, MDO slot, also compute offset bias to access the slot.
void setup_md_access(ciMethod* method, int bci,
ciMethodData*& md, ciProfileData*& data, int& mdo_offset_bias);
Expand Down
123 changes: 2 additions & 121 deletions src/hotspot/cpu/s390/interp_masm_s390.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1267,133 +1267,14 @@ void InterpreterMacroAssembler::profile_virtual_call(Register receiver,
test_method_data_pointer(mdp, profile_continue);

// Record the receiver type.
record_klass_in_profile(receiver, mdp, reg2);
profile_receiver_type(receiver, mdp, 0, reg2);

// The method data pointer needs to be updated to reflect the new target.
update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));
bind(profile_continue);
}
}

// This routine creates a state machine for updating the multi-row
// type profile at a virtual call site (or other type-sensitive bytecode).
// The machine visits each row (of receiver/count) until the receiver type
// is found, or until it runs out of rows. At the same time, it remembers
// the location of the first empty row. (An empty row records null for its
// receiver, and can be allocated for a newly-observed receiver type.)
// Because there are two degrees of freedom in the state, a simple linear
// search will not work; it must be a decision tree. Hence this helper
// function is recursive, to generate the required tree structured code.
// It's the interpreter, so we are trading off code space for speed.
// See below for example code.
void InterpreterMacroAssembler::record_klass_in_profile_helper(
Register receiver, Register mdp,
Register reg2, int start_row,
Label& done) {
if (TypeProfileWidth == 0) {
increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
return;
}

int last_row = VirtualCallData::row_limit() - 1;
assert(start_row <= last_row, "must be work left to do");
// Test this row for both the receiver and for null.
// Take any of three different outcomes:
// 1. found receiver => increment count and goto done
// 2. found null => keep looking for case 1, maybe allocate this cell
// 3. found something else => keep looking for cases 1 and 2
// Case 3 is handled by a recursive call.
for (int row = start_row; row <= last_row; row++) {
NearLabel next_test;
bool test_for_null_also = (row == start_row);

// See if the receiver is receiver[n].
int recvr_offset = in_bytes(VirtualCallData::receiver_offset(row));
test_mdp_data_at(mdp, recvr_offset, receiver,
(test_for_null_also ? reg2 : noreg),
next_test);
// (Reg2 now contains the receiver from the CallData.)

// The receiver is receiver[n]. Increment count[n].
int count_offset = in_bytes(VirtualCallData::receiver_count_offset(row));
increment_mdp_data_at(mdp, count_offset);
z_bru(done);
bind(next_test);

if (test_for_null_also) {
Label found_null;
// Failed the equality check on receiver[n]... Test for null.
z_ltgr(reg2, reg2);
if (start_row == last_row) {
// The only thing left to do is handle the null case.
z_brz(found_null);
// Receiver did not match any saved receiver and there is no empty row for it.
// Increment total counter to indicate polymorphic case.
increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
z_bru(done);
bind(found_null);
break;
}
// Since null is rare, make it be the branch-taken case.
z_brz(found_null);

// Put all the "Case 3" tests here.
record_klass_in_profile_helper(receiver, mdp, reg2, start_row + 1, done);

// Found a null. Keep searching for a matching receiver,
// but remember that this is an empty (unused) slot.
bind(found_null);
}
}

// In the fall-through case, we found no matching receiver, but we
// observed the receiver[start_row] is null.

// Fill in the receiver field and increment the count.
int recvr_offset = in_bytes(VirtualCallData::receiver_offset(start_row));
set_mdp_data_at(mdp, recvr_offset, receiver);
int count_offset = in_bytes(VirtualCallData::receiver_count_offset(start_row));
load_const_optimized(reg2, DataLayout::counter_increment);
set_mdp_data_at(mdp, count_offset, reg2);
if (start_row > 0) {
z_bru(done);
}
}

// Example state machine code for three profile rows:
// // main copy of decision tree, rooted at row[1]
// if (row[0].rec == rec) { row[0].incr(); goto done; }
// if (row[0].rec != nullptr) {
// // inner copy of decision tree, rooted at row[1]
// if (row[1].rec == rec) { row[1].incr(); goto done; }
// if (row[1].rec != nullptr) {
// // degenerate decision tree, rooted at row[2]
// if (row[2].rec == rec) { row[2].incr(); goto done; }
// if (row[2].rec != nullptr) { count.incr(); goto done; } // overflow
// row[2].init(rec); goto done;
// } else {
// // remember row[1] is empty
// if (row[2].rec == rec) { row[2].incr(); goto done; }
// row[1].init(rec); goto done;
// }
// } else {
// // remember row[0] is empty
// if (row[1].rec == rec) { row[1].incr(); goto done; }
// if (row[2].rec == rec) { row[2].incr(); goto done; }
// row[0].init(rec); goto done;
// }
// done:

void InterpreterMacroAssembler::record_klass_in_profile(Register receiver,
Register mdp, Register reg2) {
assert(ProfileInterpreter, "must be profiling");
Label done;

record_klass_in_profile_helper(receiver, mdp, reg2, 0, done);

bind (done);
}

void InterpreterMacroAssembler::profile_ret(Register return_bci, Register mdp) {
if (ProfileInterpreter) {
NearLabel profile_continue;
Expand Down Expand Up @@ -1462,7 +1343,7 @@ void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass,
mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());

// Record the object type.
record_klass_in_profile(klass, mdp, reg2);
profile_receiver_type(klass, mdp, 0, reg2);
}
update_mdp_by_constant(mdp, mdp_delta);

Expand Down
6 changes: 0 additions & 6 deletions src/hotspot/cpu/s390/interp_masm_s390.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -280,12 +280,6 @@ class InterpreterMacroAssembler: public MacroAssembler {
Register test_value_out,
Label& not_equal_continue);

void record_klass_in_profile(Register receiver, Register mdp,
Register reg2);
void record_klass_in_profile_helper(Register receiver, Register mdp,
Register reg2, int start_row,
Label& done);

void update_mdp_by_offset(Register mdp_in, int offset_of_offset);
void update_mdp_by_offset(Register mdp_in, Register dataidx, int offset_of_disp);
void update_mdp_by_constant(Register mdp_in, int constant);
Expand Down
Loading