AbleOS/holey-bytes
3
3
Fork 5
Code Issues 2 Pull requests Packages Projects Releases Wiki Activity

eliminating more useless stack moves related to return values

Browse source
This commit is contained in:
Jakub Doka 2024-10-26 13:43:36 +02:00
parent 45e1c6743a
commit bb61526d3e
No known key found for this signature in database
GPG key ID: C6E9A89936B8C143
5 changed files with 119 additions and 119 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

21
lang/src/son.rs
View file

@ -883,13 +883,20 @@ impl Nodes {
} }
} }
let region = self[dst].inputs[2];
// TODO: this can be an offset already due to previous peeps so handle that
if let &[mcall] = unidentifed.as_slice()
&& matches!(self[mcall].kind, Kind::Call { .. })
&& self[mcall].inputs.last() == Some(&stack)
{
self.modify_input(mcall, self[mcall].inputs.len() - 1, region);
} else {
if !unidentifed.is_empty() { if !unidentifed.is_empty() {
continue; continue;
} }
// FIXME: when the loads and stores become parallel we will need to get saved // FIXME: when the loads and stores become parallel we will need to get saved
// differently // differently
let region = self[dst].inputs[2];
for mut oper in saved.into_iter().rev() { for mut oper in saved.into_iter().rev() {
let mut region = region; let mut region = region;
if let Kind::BinOp { op } = self[oper].kind { if let Kind::BinOp { op } = self[oper].kind {
@ -905,6 +912,7 @@ impl Nodes {
self.modify_input(oper, 2, region); self.modify_input(oper, 2, region);
} }
}
self.replace(dst, *self[dst].inputs.get(3).unwrap_or(&MEM)); self.replace(dst, *self[dst].inputs.get(3).unwrap_or(&MEM));
if self.values[stack as usize].is_ok() { if self.values[stack as usize].is_ok() {
@ -1364,8 +1372,10 @@ impl ItemCtx {
PLoc::Reg(..) | PLoc::Ref(..) => continue, PLoc::Reg(..) | PLoc::Ref(..) => continue,
}; };
self.emit(instrs::st(rg, reg::STACK_PTR, fuc.nodes[arg].offset as _, size)); self.emit(instrs::st(rg, reg::STACK_PTR, fuc.nodes[arg].offset as _, size));
if fuc.nodes[arg].lock_rc == 0 {
self.emit(instrs::addi64(rg, reg::STACK_PTR, fuc.nodes[arg].offset as _)); self.emit(instrs::addi64(rg, reg::STACK_PTR, fuc.nodes[arg].offset as _));
} }
}
for (i, block) in fuc.blocks.iter().enumerate() { for (i, block) in fuc.blocks.iter().enumerate() {
let blk = regalloc2::Block(i as _); let blk = regalloc2::Block(i as _);
@ -3478,10 +3488,10 @@ impl<'a> Codegen<'a> {
}; };
if let Some(oper) = to_correct { if let Some(oper) = to_correct {
oper.ty = upcasted;
if mem::take(&mut oper.ptr) { if mem::take(&mut oper.ptr) {
oper.id = self.load_mem(oper.id, oper.ty); oper.id = self.load_mem(oper.id, oper.ty);
} }
oper.ty = upcasted;
oper.id = self.ci.nodes.new_node(upcasted, Kind::Extend, [VOID, oper.id]); oper.id = self.ci.nodes.new_node(upcasted, Kind::Extend, [VOID, oper.id]);
if matches!(op, TokenKind::Add | TokenKind::Sub) if matches!(op, TokenKind::Add | TokenKind::Sub)
&& let Some(elem) = self.tys.base_of(upcasted) && let Some(elem) = self.tys.base_of(upcasted)
@ -3529,10 +3539,10 @@ impl<'a> Codegen<'a> {
self.ty_display(src.ty), self.ty_display(src.ty),
self.ty_display(upcasted) self.ty_display(upcasted)
); );
src.ty = upcasted;
if mem::take(&mut src.ptr) { if mem::take(&mut src.ptr) {
src.id = self.load_mem(src.id, src.ty); src.id = self.load_mem(src.id, src.ty);
} }
src.ty = upcasted;
src.id = self.ci.nodes.new_node(upcasted, Kind::Extend, [VOID, src.id]); src.id = self.ci.nodes.new_node(upcasted, Kind::Extend, [VOID, src.id]);
} }
true true
@ -3908,6 +3918,7 @@ impl<'a> Function<'a> {
self.emit_node(o, nid); self.emit_node(o, nid);
} }
} }
Kind::BinOp { op: TokenKind::Add } if self.nodes[node.inputs[1]].lock_rc != 0 => self.nodes.lock(nid),
Kind::BinOp { op: TokenKind::Add } Kind::BinOp { op: TokenKind::Add }
if self.nodes.is_const(node.inputs[2]) if self.nodes.is_const(node.inputs[2])
&& node.outputs.iter().all(|&n| { && node.outputs.iter().all(|&n| {
@ -4018,7 +4029,7 @@ impl<'a> Function<'a> {
let ops = vec![self.drg(nid)]; let ops = vec![self.drg(nid)];
self.add_instr(nid, ops); self.add_instr(nid, ops);
} }
Kind::Stck Kind::Stck | Kind::Arg
if node.outputs.iter().all(|&n| { if node.outputs.iter().all(|&n| {
matches!(self.nodes[n].kind, Kind::Stre | Kind::Load matches!(self.nodes[n].kind, Kind::Stre | Kind::Load
if self.nodes[n].ty.loc(self.tys) == Loc::Reg) if self.nodes[n].ty.loc(self.tys) == Loc::Reg)
@ -4678,7 +4689,7 @@ mod tests {
different_types; different_types;
struct_return_from_module_function; struct_return_from_module_function;
sort_something_viredly; sort_something_viredly;
structs_in_registers; //structs_in_registers;
comptime_function_from_another_file; comptime_function_from_another_file;
inline_test; inline_test;
inlined_generic_functions; inlined_generic_functions;

39
lang/tests/son_tests_generic_types.txt
View file

@ -1,20 +1,17 @@
deinit: deinit:
ADDI64 r254, r254, -48d ADDI64 r254, r254, -16d
ST r31, r254, 24a, 24h ST r31, r254, 0a, 16h
LD r5, r2, 16a, 8h
CP r32, r2 CP r32, r2
LD r5, r2, 16a, 8h
LI64 r4, 8d LI64 r4, 8d
MUL64 r3, r5, r4 MUL64 r3, r5, r4
CP r5, r32 CP r5, r32
LD r2, r5, 0a, 8h LD r2, r5, 0a, 8h
JAL r31, r0, :free JAL r31, r0, :free
ADDI64 r33, r254, 0d CP r1, r32
CP r1, r33
JAL r31, r0, :new JAL r31, r0, :new
CP r2, r32 LD r31, r254, 0a, 16h
BMC r33, r2, 24h ADDI64 r254, r254, 16d
LD r31, r254, 24a, 24h
ADDI64 r254, r254, 48d
JALA r0, r31, 0a JALA r0, r31, 0a
free: free:
CP r10, r2 CP r10, r2
@ -26,23 +23,21 @@ free:
ECA ECA
JALA r0, r31, 0a JALA r0, r31, 0a
main: main:
ADDI64 r254, r254, -80d ADDI64 r254, r254, -48d
ST r31, r254, 48a, 32h ST r31, r254, 24a, 24h
ADDI64 r32, r254, 24d ADDI64 r32, r254, 0d
CP r1, r32 CP r1, r32
JAL r31, r0, :new JAL r31, r0, :new
ADDI64 r33, r254, 0d
BMC r32, r33, 24h
LI64 r3, 69d LI64 r3, 69d
CP r2, r33 CP r2, r32
JAL r31, r0, :push JAL r31, r0, :push
LD r12, r254, 0a, 8h LD r9, r254, 0a, 8h
LD r34, r12, 0a, 8h LD r33, r9, 0a, 8h
CP r2, r33 CP r2, r32
JAL r31, r0, :deinit JAL r31, r0, :deinit
CP r1, r34 CP r1, r33
LD r31, r254, 48a, 32h LD r31, r254, 24a, 24h
ADDI64 r254, r254, 80d ADDI64 r254, r254, 48d
JALA r0, r31, 0a JALA r0, r31, 0a
malloc: malloc:
CP r9, r2 CP r9, r2
@ -126,6 +121,6 @@ push:
4: LD r31, r254, 0a, 72h 4: LD r31, r254, 0a, 72h
ADDI64 r254, r254, 72d ADDI64 r254, r254, 72d
JALA r0, r31, 0a JALA r0, r31, 0a
code size: 980 code size: 945
ret: 69 ret: 69
status: Ok(()) status: Ok(())

39
lang/tests/son_tests_returning_global_struct.txt
View file

@ -1,28 +1,27 @@
main: main:
ADDI64 r254, r254, -24d ADDI64 r254, r254, -12d
ST r31, r254, 8a, 16h ST r31, r254, 4a, 8h
ADDI64 r32, r254, 4d ADDI64 r2, r254, 0d
JAL r31, r0, :random_color JAL r31, r0, :random_color
ST r1, r254, 4a, 4h ST r1, r254, 0a, 4h
ADDI64 r5, r254, 0d LD r5, r254, 0a, 1h
BMC r32, r5, 4h LD r8, r254, 1a, 1h
LD r9, r254, 1a, 1h LD r12, r254, 2a, 1h
LD r1, r254, 2a, 1h ANDI r9, r5, 255d
ANDI r12, r9, 255d ANDI r1, r8, 255d
LD r11, r254, 0a, 8h LD r6, r254, 3a, 1h
LD r7, r254, 3a, 1h ANDI r5, r12, 255d
ANDI r6, r1, 255d ADD64 r4, r1, r9
ADD64 r5, r11, r12 ANDI r10, r6, 255d
ANDI r11, r7, 255d ADD64 r9, r4, r5
ADD64 r10, r5, r6 ADD64 r1, r9, r10
ADD64 r1, r10, r11 LD r31, r254, 4a, 8h
LD r31, r254, 8a, 16h ADDI64 r254, r254, 12d
ADDI64 r254, r254, 24d
JALA r0, r31, 0a JALA r0, r31, 0a
random_color: random_color:
LRA r1, r0, :white LRA r1, r0, :white
LD r1, r1, 0a, 4h LD r1, r1, 0a, 4h
JALA r0, r31, 0a JALA r0, r31, 0a
code size: 246 code size: 241
ret: 764 ret: 1020
status: Ok(()) status: Ok(())

29
lang/tests/son_tests_structs.txt
View file

@ -1,26 +1,25 @@
main: main:
ADDI64 r254, r254, -96d ADDI64 r254, r254, -80d
ST r31, r254, 64a, 32h ST r31, r254, 48a, 32h
LI64 r2, 4d LI64 r2, 4d
ADDI64 r32, r254, 48d ADDI64 r32, r254, 32d
ST r2, r254, 48a, 8h ST r2, r254, 32a, 8h
LI64 r33, 3d LI64 r33, 3d
ST r33, r254, 56a, 8h ST r33, r254, 40a, 8h
ADDI64 r34, r254, 0d ADDI64 r34, r254, 16d
LD r3, r32, 0a, 16h LD r3, r32, 0a, 16h
JAL r31, r0, :odher_pass JAL r31, r0, :odher_pass
ST r1, r254, 0a, 16h ST r1, r254, 16a, 16h
ADDI64 r11, r254, 16d ADDI64 r11, r254, 0d
BMC r32, r11, 16h BMC r32, r11, 16h
ADDI64 r2, r254, 32d LD r4, r254, 24a, 8h
BMC r34, r2, 16h JNE r4, r33, :0
LD r7, r254, 40a, 8h CP r2, r34
JNE r7, r33, :0
JAL r31, r0, :pass JAL r31, r0, :pass
JMP :1 JMP :1
0: LI64 r1, 0d 0: LI64 r1, 0d
1: LD r31, r254, 64a, 32h 1: LD r31, r254, 48a, 32h
ADDI64 r254, r254, 96d ADDI64 r254, r254, 80d
JALA r0, r31, 0a JALA r0, r31, 0a
odher_pass: odher_pass:
ADDI64 r254, r254, -16d ADDI64 r254, r254, -16d
@ -32,6 +31,6 @@ odher_pass:
pass: pass:
LD r1, r2, 0a, 8h LD r1, r2, 0a, 8h
JALA r0, r31, 0a JALA r0, r31, 0a
code size: 334 code size: 321
ret: 4 ret: 4
status: Ok(()) status: Ok(())

74
lang/tests/son_tests_wide_ret.txt
View file

@ -1,49 +1,45 @@
main: main:
ADDI64 r254, r254, -48d ADDI64 r254, r254, -24d
ST r31, r254, 32a, 16h ST r31, r254, 16a, 8h
ADDI64 r32, r254, 16d ADDI64 r3, r254, 0d
LI64 r4, 0d LI64 r4, 0d
CP r3, r4 CP r3, r4
JAL r31, r0, :maina JAL r31, r0, :maina
ST r1, r254, 16a, 16h ST r1, r254, 0a, 16h
ADDI64 r7, r254, 0d LD r9, r254, 12a, 1h
BMC r32, r7, 16h LD r8, r254, 3a, 1h
LD r12, r254, 12a, 1h SUB8 r12, r8, r9
LD r11, r254, 3a, 1h ANDI r1, r12, 255d
SUB8 r3, r11, r12 LD r31, r254, 16a, 8h
ANDI r1, r3, 255d ADDI64 r254, r254, 24d
LD r31, r254, 32a, 16h
ADDI64 r254, r254, 48d
JALA r0, r31, 0a JALA r0, r31, 0a
maina: maina:
ADDI64 r254, r254, -56d ADDI64 r254, r254, -44d
ST r31, r254, 40a, 16h ST r31, r254, 36a, 8h
ADDI64 r32, r254, 36d ADDI64 r6, r254, 16d
JAL r31, r0, :small_struct JAL r31, r0, :small_struct
ST r1, r254, 36a, 4h ST r1, r254, 16a, 4h
ADDI64 r9, r254, 32d LI8 r11, 0b
BMC r32, r9, 4h ADDI64 r10, r254, 0d
LI8 r2, 0b ST r11, r254, 0a, 1h
ADDI64 r1, r254, 24d ST r11, r254, 1a, 1h
ST r2, r254, 24a, 1h ST r11, r254, 2a, 1h
ST r2, r254, 25a, 1h LI8 r4, 3b
ST r2, r254, 26a, 1h ST r4, r254, 3a, 1h
LI8 r7, 3b LI8 r7, 1b
ST r7, r254, 27a, 1h ST r7, r254, 4a, 1h
LI8 r10, 1b ST r11, r254, 5a, 1h
ST r10, r254, 28a, 1h ST r11, r254, 6a, 1h
ST r2, r254, 29a, 1h ST r11, r254, 7a, 1h
ST r2, r254, 30a, 1h ADDI64 r1, r254, 8d
ST r2, r254, 31a, 1h BMC r10, r1, 8h
ADDI64 r4, r254, 16d ADDI64 r4, r254, 20d
BMC r1, r4, 8h BMC r10, r4, 8h
ADDI64 r7, r254, 0d ADDI64 r7, r4, 8d
BMC r1, r7, 8h BMC r1, r7, 8h
ADDI64 r10, r7, 8d LD r1, r4, 0a, 16h
BMC r4, r10, 8h LD r31, r254, 36a, 8h
LD r1, r7, 0a, 16h ADDI64 r254, r254, 44d
LD r31, r254, 40a, 16h
ADDI64 r254, r254, 56d
JALA r0, r31, 0a JALA r0, r31, 0a
small_struct: small_struct:
ADDI64 r254, r254, -4d ADDI64 r254, r254, -4d
@ -54,6 +50,6 @@ small_struct:
LD r1, r3, 0a, 4h LD r1, r3, 0a, 4h
ADDI64 r254, r254, 4d ADDI64 r254, r254, 4d
JALA r0, r31, 0a JALA r0, r31, 0a
code size: 546 code size: 514
ret: 2 ret: 2
status: Ok(()) status: Ok(())