Hi everyone,
My laptop is in the repair shop again unfortunately due to a jammed cooling
fan, so this e-mail may have indentation problems in the mailing list.
That aside, I'm taking the opportunity to look at some optimisations for
AArch64, and in the process I rediscovered the problem with labels having
incorrect reference counts, which ultimately means I cannot reliably
detect if a label only has a single reference.
After some investigation, I found out this is due to how AArch64 creates jumps
in the code generator:
----
procedure tcgaarch64.a_jmp_always(list: TAsmList; l: TAsmLabel);
var
ai: taicpu;
begin
ai:=TAiCpu.op_sym(A_B,current_asmdata.RefAsmSymbol(l.name,AT_FUNCTION));
ai.is_jmp:=true;
list.Concat(ai);
end;
----
Nothing too out of the ordinary. However, the problem comes from reading the
"name" property, since this calls the virtual method GetName, which is
overridden in TAsmLabel:
----
function TAsmLabel.getname:TSymStr;
begin
getname:=inherited getname;
increfs;
end;
----
Every time you want to get the name of the label, its reference count is
incremented. Coupled with tai_cpu_abstract.loadref incrementing the reference
count for a symbol, this has the result for increasing the
reference count by 2 for each jump. This is very undesirable because it has
the side-effect of increasing the reference count whenever you want to find out
the label's name, which also has the side-effect of
increasing the reference count when writing the assembly dumps (as evidenced
when DEBUG_LABEL Is defined).
At the moment I can't just simply remove the overridden method because other
parts of the compiler depend on the reference count being incremented, as
evidenced when I try to build for aarch64-linux:
----
as -o /home/pi/Documents/source/fpc/rtl/units/aarch64-linux/gprt0.o
aarch64/gprt0.as
/home/pi/Documents/source/fpc/compiler/ppc1 -Ur -Ur -Xs -O2 -n -Fi../inc
-Fi../aarch64 -Fi../unix -Fiaarch64 -FE.
-FU/home/pi/Documents/source/fpc/rtl/units/aarch64-linux
-Fl/usr/lib/gcc/aarch64-linux-gnu/8 -a -O4 -
DD2022/05/07 -daarch64 -dRELEASE -Us -Sg system.pp
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Assembler
messages:
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld17$strlab does not evaluate to a constant
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld41$strlab does not evaluate to a constant
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld42$strlab does not evaluate to a constant
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld43$strlab does not evaluate to a constant
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld44$strlab does not evaluate to a constant
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld45$strlab does not evaluate to a constant
/home/pi/Documents/source/fpc/rtl/units/aarch64-linux/system.s: Error: .size
expression for .Ld46$strlab does not evaluate to a constant
system.pp(737) Error: Error while assembling exitcode 1
system.pp(737) Fatal: There were 2 errors compiling module, stopping
Fatal: Compilation aborted
----
In conclusion, I would like to refactor the compiler so simply reading the
label name does not automatically increase its reference count, and correct the
parts of the compiler that depend on this behaviour.
Without accurate reference counts, it is very difficult to make reliable
optimisations for AArch64 related to branches, and probably other platforms too.
Gareth aka. Kit
_______________________________________________
fpc-devel maillist - [email protected]
https://lists.freepascal.org/cgi-bin/mailman/listinfo/fpc-devel
