From 2bc3d9a3815a6cd172b3b4f7696c46a006021b13 Mon Sep 17 00:00:00 2001
From: Mirek Kratochvil <exa.exa@gmail.com>
Date: Tue, 31 Oct 2023 00:12:51 +0100
Subject: [PATCH] gc kinda almost there

---
 exit_123.s        |   1 +
 fibs.s            |   3 +
 include/data.s    | 189 +++++++++++++++++++++++++++++++++-------------
 include/gc.s      | 173 ++++++++++++++++++++++++++++++++++++------
 include/io.s      |   1 +
 include/listops.s |   2 +
 include/primops.s |   2 +
 include/uskel.s   |  37 +--------
 print_123.s       |   1 +
 9 files changed, 298 insertions(+), 111 deletions(-)

diff --git a/exit_123.s b/exit_123.s
index 1e7ba23..cf56f7d 100644
--- a/exit_123.s
+++ b/exit_123.s
@@ -5,6 +5,7 @@
 
 # || -> cont
 .thunkcode main
+	needs_alloc $0120
 	thunkto %r11, $INT_code, $100
 	thunkto %r12, $INT_code, $23
 	thunkto %r13, $plus, $2, %r11, %r12
diff --git a/fibs.s b/fibs.s
index 3d35544..f008dae 100644
--- a/fibs.s
+++ b/fibs.s
@@ -8,6 +8,7 @@
 
 # | lag1 | lag2 | -> cont
 .thunkcode fibs
+	needs_alloc $0140
 	# next value
 	thunkto %r11, $plus, $2, 020(%rbp), 030(%rbp)
 
@@ -28,12 +29,14 @@
 
 # || -> cont
 .thunkcode fibs0
+	needs_alloc $0100
 	thunkto %r12, $INT_code, $1
 	thunkto %r11, $INT_code, $0
 	thunkto %rbp, $fibs, $2, %r11, %r12
 	enter_rbp
 
 .thunkcode main
+	needs_alloc $0160
 	thunkto %r12, $fibs0, $0
 	thunkto %r11, $INT_code, $20
 	thunkto %r11, $list_int_index, $2, %r11, %r12
diff --git a/include/data.s b/include/data.s
index 43d2ba8..3ee8213 100644
--- a/include/data.s
+++ b/include/data.s
@@ -2,40 +2,104 @@
 .ifndef _data_s_file
 _data_s_file:
 
-# Simple values and boxed machine integers
-# | ptr | value |
-CON_evacuate1:
-	retq # TODO
-CON_scavenge1:
-	retq
-
 # Format of the info tables:
 # - code
 # ----- code pointer
 # - 8B helper information for eval/apply (generally this is 0, and only gets used for FUN/PAP)
 # - 8B pointer to scavenge
 # - 8B pointer to evacuate
+#
+# Evacuate interface:
+# in: %rsi continuation, %rbp what to evacuate
+# out: %rbp where the thing is now
+# Notes:
+#  - IND thunks skip themselves on evacuate
+#  - checking if whether stuff is already in write region are managed by _gc_evacuate prelude
+#
+# Scavenge interface:
+# in: %rsi continuation, %rbp what to scavenge
+# out: %rbp next thing to scavenge in memory
+#
+# Saved registers by evacuate and scavenge
+# - _uskel_gc needs to preserve %rdi now; that might increase
+# - scavenges use %r12-%r15
+# - %rax-%rdx is scratch and evacuate use
+
+# Simple values and boxed machine integers
+# | ptr | value |
+INT_evacuate:
+	pushq 010(%rbp)
+	pushq $INT_code
+	mov %rsp,%rbp
+	jmp *%rsi
+INT_scavenge:
+	add $020, %rbp
+	jmp *%rsi
 
 INT_info_table:
-	cell CON_evacuate1
-	cell CON_scavenge1
+	cell INT_evacuate
+	cell INT_scavenge
 	cell 0
 INT_code:
 	continue
 
+# Indirection
+# | ptr | indptr |
+IND_evacuate:
+	mov 010(%rbp), %rbp
+	jmp _gc_evacuate
+IND_scavenge:
+	jmp 0 # thou shalt not scavenge here
+
+IND_info:
+	cell IND_evacuate
+	cell IND_scavenge
+	cell 0
+IND_code:
+	enter 010(%rbp)
+
 # List
-# | ptr | 0 |
-# | ptr | 1 | a | b |
+# | ptr | 0 |           # [] case
+# | ptr | 1 | a | b |   # (a:b) case
 LIST_evacuate:
-	# [] | a : b
-	retq #TODO
+	cmpq $0, 010(%rbp)
+	je LIST_evacuate_nil
+	pushq 030(%rbp)
+	pushq 020(%rbp)
+	pushq $1
+	pushq $LIST_code
+	mov %rsp, %rbp
+	jmp *%rsi
+	LIST_evacuate_nil:
+	pushq $0
+	pushq $LIST_code
+	mov %rsp, %rbp
+	jmp *%rsi
 LIST_scavenge:
-	mov 010(%rbp), %rax
-	shl $1, %rax
-	add $2, %rax
-	shl $3, %rax
-	add %rax, %rsi
-	retq
+	cmpq $0, 010(%rbp)
+	je LIST_scavenge_nil
+	mov %rbp, %r15
+	mov %rsi, %r14
+	
+	mov $LIST_scavenge1, %rsi
+	mov 020(%r15), %rbp
+	jmp _gc_evacuate
+	LIST_scavenge1:
+	mov %rbp, 020(%r15)
+
+	mov $LIST_scavenge2, %rsi
+	mov 030(%r15), %rbp
+	jmp _gc_evacuate
+	LIST_scavenge2:
+	mov %rbp, 030(%r15)
+
+	mov %r15, %rbp
+	add $040, %rbp
+	jmp *%r14
+	LIST_scavenge_nil:
+	add $020, %rbp
+	jmp *%rsi
+
 LIST_info_table:
 	cell LIST_evacuate
 	cell LIST_scavenge
@@ -46,13 +110,14 @@ LIST_code:
 # FUN objects
 # | ptr | thunkptr | args | arg[0] | arg[1] | ... | arg[args] |
 FUN_evacuate:
-	retq #TODO
+	#TODO
 FUN_scavenge:
-	mov 020(%rbp), %rax
-	add $3, %rax
-	shl $3, %rax
-	add %rax, %rsi
-	retq
+	#TODO
+	#mov 020(%rbp), %rax
+	#add $3, %rax
+	#shl $3, %rax
+	#add %rax, %rsi
+	#retq
 
 # Info tables for FUN objects.
 FUN0_info_table:
@@ -89,39 +154,55 @@ FUN4_info_table:
 	cell 4
 FUN4_code:
 	continue
+# add more funN here as needed
 
-# TODO: add more funN here as needed
-
-# indirection (Q: how to recognize IND and THUNK on return?)
-# | ptr | indptr |
-IND_evacuate:
-	retq #TODO
-IND_scavenge:
-	add $020,%rsi
-	retq
-IND_info:
-	cell IND_evacuate
-	cell IND_scavenge
-	cell 0
-IND_code:
-	enter 010(%rbp)
-
-# THU objects (gc implementation only, actual THUs are defined by functions)
+# THU objects (gc implementation only, actual THU data are created by functions)
 # | ptr | args | arg[0] | arg[1] | ... | arg[args] |
 # args wouldn't need to be here but let's keep them for gc simplicity
 THU_evacuate:
-	retq #TODO
+	mov 010(%rbp), %rbx
+	mov %rbx,%rdx
+	mov %rbx,%rcx
+	add $2, %rdx
+	shl $3, %rdx
+	add %rbp, %rdx
+	THU_evacuate_one:
+	dec %rcx
+	cmp $0, %rcx
+	jl THU_evacuate_fini
+	sub $010, %rdx
+	pushq (%rdx)
+	jmp THU_evacuate_one
+	THU_evacuate_fini:
+	pushq %rbx
+	pushq 000(%rbp)
+	mov %rsp, %rbp
+	jmp *%rsi
 THU_scavenge:
-	mov 010(%rbp), %rax
-	add $2,%rax
-	shl $3,%rax
-	add %rax,%rsi
-	retq
+	mov 010(%rbp), %r12
+	mov %rbp, %r15
+	mov %rsi, %r14
+	mov %r12,%r13
+	add $2, %r13
+	shl $3, %r13
+	add %r15, %r13
+	THU_scavenge_one:
+	sub $010, %r13
+	sub $1, %r12
+	jl THU_scavenge_fini
+	mov (%r13), %rbp
+	mov $THU_scavenge_one_cont, %rsi
+	jmp _gc_evacuate
+	THU_scavenge_one_cont:
+	mov %rbp, (%r13)
+	jmp THU_scavenge_one
+
+	THU_scavenge_fini:
+	mov %r15, %rbp
+	mov 010(%rbp), %r13
+	add $2, %r13
+	shl $3, %r13
+	add %r13, %rbp
+	jmp *%r14
 
 .endif # _data_s_file
-
-# evacuate and scavenge:
-# - evacuate just copies the object
-# - scavenge evacuates all children (to the new location IF they are in the old
-#   location), changes the pointer, and moves the scavenge pointer to the next
-#   object (because everything needs to be scavenged)
diff --git a/include/gc.s b/include/gc.s
index bb32240..636a6a2 100644
--- a/include/gc.s
+++ b/include/gc.s
@@ -2,10 +2,53 @@
 .ifndef _gc_s_file
 _gc_s_file:
 
-.macro prealloc name amount
-	mov %rsp,%rax
+.section .bss
+_write_region_start:
+	# begin of the active memory area
+	cell 0
+_write_region_end:
+	# end of the active memory area (%rsp kinda starts here and goes down
+	# towars the start)
+	cell 0
+_gc_trigger:
+	# point in memory where the gc will trigger (we don't necessarily wait for the write region to fill up!)
+	cell 0
+
+_gc_last_size:
+	# how much data we evacuated last time
+	cell 0
+_gc_min_alloc:
+	# minimum possible allocation
+	cell 0 # tunable constant
+_gc_grow_ratio:
+	# 256th's of the minimal amount of memory increment compared to the
+	# last time. New minimal amount is compared as:
+	# (ratio * last size) >> 8
+	cell 0 # tunable constant
+_gc_shrink_ratio:
+	# 256th's of the ratio of post-gc still-free to-space that should be considered for discarding
+	cell 0
+
+_gc_region_start:
+	# in GC, this region is being evacuated and will eventually disappear
+	cell 0
+_gc_region_end:
+	# end of the disappear region
+	cell 0
+
+_gc_backup_thunk:
+	# backup of %rsi so that we can use the register for other nonsense
+	cell 0
+_gc_backup_cont:
+	# backup of %rbp for same reason
+	cell 0
+
+.section .text
+
+.macro needs_alloc amount
+	mov %rsp, %rax
 	sub _write_region_start, %rax
-	cmp $amount, %rax #TODO check direction
+	cmp \amount, %rax
 	jb _uskel_gc
 .endm
 
@@ -14,23 +57,24 @@ _uskel_alloc:
 
 	# calculate the desired size to %r14
 	mov _gc_min_alloc, %r14
+	#add _gc_region_end, %r14
+	#sub _gc_region_start, %r14
 
 	# check if the desired size isn't greater because of the last gc use
 	mov _gc_last_size, %rax
-	mul _gc_min_expect_ratio, %rax
-	shr 8, %rax
-	cmp %r14, %rax #TODO check direction
+	mulq _gc_grow_ratio
+	shr $8, %rax
+	add _gc_min_alloc, %rax
+	cmp %r14, %rax
 	cmova %rax, %r14
 
 	# check if we don't need even more space because we need to evacuate stuff
 	mov _gc_region_end, %rax
 	sub %rsp, %rax # trick -- if we counted from gc region start, allocated memory could never shrink
-	cmp %r14, %rax #TODO direction
+	cmp %r14, %rax
 	cmova %rax, %r14
-	#TODO add functionality to trigger the gc's a bit earlier than when
-	# they hit _write_region_start, to allow for faster compaction.
-	# maybe _write_region_trigger = 25%ish between _write_region_start and %rsp ?
 
+	alloc_goes_mmap:
 	mov $9, %rax # mmap
 	mov $0, %rdi # addr = NULL
 	mov %r14, %rsi # len = %r14
@@ -48,27 +92,110 @@ _uskel_alloc:
 
 	jmp *%r15
 
+_uskel_gc_init:
+	mov %rsi, %r13
+	movq $0x100, _gc_min_alloc
+	movq $0x180, _gc_grow_ratio
+	movq $0x40, _gc_shrink_ratio
+	mov $0, %rsp # fake original rsp for first alloc run
+	mov $_uskel_gc_init_cont, %rsi
+	jmp _uskel_alloc
+	_uskel_gc_init_cont:
+	mov _write_region_start, %rax
+	mov %rax, _gc_trigger
+	jmp *%r13
+
 _uskel_gc:
+	# save what we did before ending up here
 	mov %rbp, _gc_backup_thunk
 	mov %rsi, _gc_backup_cont
+
 	# first we need a new memory area
-	mov _write_region_start, _gc_region_start
-	mov _write_region_end, _gc_region_end
-	mov $_uskel_gc_evacuate, %r15
+	mov _write_region_start, %rbx
+	mov _write_region_end, %rcx
+	mov %rbx, _gc_region_start
+	mov %rcx, _gc_region_end
+	mov $_uskel_gc_evacuate, %rsi
 	jmp _uskel_alloc
 	_uskel_gc_evacuate:
-	# we may also need a scavenging queue! Or can that be solved by walking the write region?
-	# plan:
-	# evacuate thunk and cont, save them right away in _gc_backup_thunk and _cont
-	# scavenge from the top of the gc region, evacuating stuff en route
-	# (this is a little more complex because we grow down, so we have to
-	# scavenge up from %rsp to the last known scavenged position)
-	# finish when scavenge doesn't have anything more to scavenge
+
+	# point the writer to the new memory area
+	mov _write_region_end, %rsp
+	mov %rsp, %rdi # %rdi is "the last scavenged thing"
+	
+	# start by evacuating the thunk and cont
+	mov _gc_backup_thunk, %rbp
+	mov $_uskel_gc_evacuate_cont_thunk, %rsi
+	jmp _gc_evacuate
+	_uskel_gc_evacuate_cont_thunk:
+	mov %rbp, _gc_backup_thunk
+
+	mov _gc_backup_cont, %rbp
+	mov $_uskel_gc_evacuate_cont_cont, %rsi
+	jmp _gc_evacuate
+	_uskel_gc_evacuate_cont_cont:
+	mov %rbp, _gc_backup_cont
+
+	# scavenge everything
+	_uskel_gc_scavenge:
+	# start at what we wrote last
+	mov %rsp, %rbp
+
+	# if the thing is already scavenged, we didn't write anything, mark done.
+	cmp %rbp, %rdi
+	jbe _uskel_gc_scavenge_end
+
+	_uskel_gc_scavenge1:
+	# if all ok, scavenge one thing (moving %rbp) and recheck
+	mov $_uskel_gc_scavenge1_ret, %rsi
+	mov (%rbp), %rax
+	jmp *-020(%rax) # scavenge position in infotable
+	_uskel_gc_scavenge1_ret:
+	cmp %rbp, %rdi
+	ja _uskel_gc_scavenge1
+
+	# everything above rsp is now scavenged, continue with next round
+	mov %rsp, %rdi
+	jmp _uskel_gc_scavenge
+
+	_uskel_gc_scavenge_end:
+
+	# deallocate the old memory region
+	mov $11, %rax # munmap
+	mov _gc_region_end, %rsi
+	mov _gc_region_start, %rdi # addr = gc start
+	sub %rdi, %rsi # len = gc end - gc start
+	syscall
+	
+	# recalculate the gc trigger point
+	mov %rsp, %rax
+	sub _write_region_start, %rax
+	mulq _gc_shrink_ratio
+	shr $8, %rax
+	add _write_region_start, %rax
+	mov %rax, _gc_trigger
 
 	# save how much data we actually had at this point
-	mov _write_region_end, _gc_last_size
-	sub %rsp, _gc_last_size
+	mov _write_region_end, %rax
+	sub %rsp, %rax
+	mov %rax, _gc_last_size
 
-	enter_rbp # restart the thunk for simplicity
+	# restore what we were doing
+	mov _gc_backup_thunk, %rbp
+	mov _gc_backup_cont, %rsi
+	enter_rbp # for simplicity just restart the thunk
+
+_gc_evacuate:
+	# check if we are really out of the target region
+	cmp _write_region_start, %rbp
+	jb _gc_evacuate_ok
+	cmp _write_region_end, %rbp
+	jae _gc_evacuate_ok
+	# if not, let's just jump to cont and leave %rbp as result
+	jmp *%rsi
+	_gc_evacuate_ok:
+	# if we should evacuate, jump to the evac routine
+	mov (%rbp), %rax
+	jmp *-030(%rax)
 
 .endif #_gc_s_file
diff --git a/include/io.s b/include/io.s
index 205f144..0431dd0 100644
--- a/include/io.s
+++ b/include/io.s
@@ -4,6 +4,7 @@ _io_s_file:
 
 # | int | -> cont
 .thunkcode print
+	needs_alloc $040
 	thunkto %rsi, $print_fini, $2, %rbp, %rsi
 	enter 020(%rbp)
 
diff --git a/include/listops.s b/include/listops.s
index 447c125..0ee39ad 100644
--- a/include/listops.s
+++ b/include/listops.s
@@ -6,6 +6,8 @@ _listops_s_file:
 
 # | n | list | -> cont
 .primop2 list_int_index
+	needs_alloc $060
+
 	mov 010(%rsi), %rdx # the list constructor id, must be 1
 	cmp $1, %rdx
 	jne list_int_index_not_found
diff --git a/include/primops.s b/include/primops.s
index f532144..117bcca 100644
--- a/include/primops.s
+++ b/include/primops.s
@@ -9,6 +9,7 @@ _primops_s_file:
 .macro .primop1 name
 # | arg1 | -> cont
 .thunkcode \name
+	needs_alloc $040
 	# push a thunk for collecting the first arg and set it as continuation
 	thunkto %rsi, $\name\()_fini, $2, %rbp, %rsi
 	enter 020(%rbp) # evaluate arg1
@@ -39,6 +40,7 @@ _primops_s_file:
 .macro .primop2 name
 # | arg1 | arg2 | -> cont
 .thunkcode \name
+	needs_alloc $050
 	# push a thunk for collecting the first arg and set it as continuation
 	thunkto %rsi, $\name\()_step1, $3, 030(%rbp), %rbp, %rsi
 	enter 020(%rbp) # evaluate arg1
diff --git a/include/uskel.s b/include/uskel.s
index bca33c9..f59a1f7 100644
--- a/include/uskel.s
+++ b/include/uskel.s
@@ -8,47 +8,16 @@ _start:
 
 .include "include/macros.s"
 
-# this has globals
 .section .bss
 _unix_rsp:
 	# back-up of program entry rsp (aka the actual stack given by the
 	# actual OS; we might like to use it at some point, maybe)
 	cell 0
-_write_region_start:
-	# begin of the active memory area
-	cell 0
-_write_region_end:
-	# end of the active memory area (%rsp kinda starts here and goes down
-	# towars the start)
-	cell 0
-_gc_last_size:
-	# how much data we evacuated last time
-	cell 0
-_gc_min_alloc:
-	# minimum possible allocation
-	cell 0x100000 # tunable constant
-_gc_min_expect_ratio:
-	# 256th's of the minimal amount of memory increment compared to the
-	# last time. New minimal amount is compared as:
-	# (ratio * last size) >> 8
-	cell 0x200 # tunable constant
-_gc_region_start:
-	# in GC, this region is being evacuated and will eventually disappear
-	cell 0
-_gc_region_end:
-	# end of the disappear region
-	cell 0
-_gc_backup_thunk:
-	# backup of %rsi so that we can use the register for other nonsense
-	cell 0
-_gc_backup_cont:
-	# backup of %rbp for same reason
-	cell 0
-
-.section .text
 
 .include "include/gc.s"
 
+.section .text
+
 _uskel_start:
 	# we use the stack pointer for easy writing to the heap;
 	# back it up to memory state just if we ever needed it again.
@@ -56,7 +25,7 @@ _uskel_start:
 
 	# allocate the initial chunk of memory
 	mov $_uskel_start_main, %rsi
-	jmp _uskel_alloc
+	jmp _uskel_gc_init
 
 	_uskel_start_main:
 	# push a thunk for main
diff --git a/print_123.s b/print_123.s
index b87309b..35c9cbe 100644
--- a/print_123.s
+++ b/print_123.s
@@ -7,6 +7,7 @@
 
 # || -> cont
 .thunkcode main
+	needs_alloc $0160
 	# push a new integer
 	thunkto %r11, $INT_code, $100