Compiled code size optimizations for x86-64 resulting in development image size reduction of about 350kb

- RIP-relative LEA is now used to load program counter in prolog and tail calls; saves 3 bytes of code and 4 bytes of relocation information over an immediate MOV
- VM field accesses no longer involve a redundant addition

basis/cpu/architecture/architecture.factor

@@ -434,6 +434,7 @@ HOOK: %set-alien-double    cpu ( ptr offset value -- )
 HOOK: %set-alien-vector    cpu ( ptr offset value rep -- )
 
 HOOK: %alien-global cpu ( dst symbol library -- )
+HOOK: %vm-field cpu ( dst fieldname -- )
 HOOK: %vm-field-ptr cpu ( dst fieldname -- )
 
 HOOK: %allot cpu ( dst size class temp -- )

basis/cpu/ppc/ppc.factor

@@ -57,10 +57,11 @@ CONSTANT: vm-reg 15
 
 : %load-vm-addr ( reg -- ) vm-reg MR ;
 
-: %load-vm-field-addr ( reg symbol -- )
+M: ppc %vm-field ( dst field -- )
-    [ vm-reg ] dip vm-field-offset ADDI ;
+    [ vm-reg ] dip vm-field-offset LWZ ;
 
-M: ppc %vm-field-ptr ( dst field -- ) %load-vm-field-addr ;
+M: ppc %vm-field-ptr ( dst field -- )
+    [ vm-reg ] dip vm-field-offset ADDI ;
 
 GENERIC: loc-reg ( loc -- reg )
 
@@ -601,26 +602,19 @@ M: ppc %push-stack ( -- )
     ds-reg ds-reg 4 ADDI
     int-regs return-reg ds-reg 0 STW ;
 
-:: %load-context-datastack ( dst -- )
-    ! Load context struct
-    dst "ctx" %vm-field-ptr
-    dst dst 0 LWZ
-    ! Load context datastack pointer
-    dst dst "datastack" context-field-offset ADDI ;
-
 M: ppc %push-context-stack ( -- )
-    11 %load-context-datastack
+    11 "ctx" %vm-field
-    12 11 0 LWZ
+    12 11 "datastack" context-field-offset LWZ
     12 12 4 ADDI
-    12 11 0 STW
+    12 11 "datastack" context-field-offset STW
     int-regs return-reg 12 0 STW ;
 
 M: ppc %pop-context-stack ( -- )
-    11 %load-context-datastack
+    11 "ctx" %vm-field
-    12 11 0 LWZ
+    12 11 "datastack" context-field-offset LWZ
     int-regs return-reg 12 0 LWZ
     12 12 4 SUBI
-    12 11 0 STW ;
+    12 11 "datastack" context-field-offset STW ;
 
 M: ppc %unbox ( n rep func -- )
     ! Value must be in r3

basis/cpu/x86/32/32.factor

@@ -27,6 +27,9 @@ M: x86.32 temp-reg ECX ;
 M: x86.32 %mov-vm-ptr ( reg -- )
     0 MOV 0 rc-absolute-cell rel-vm ;
 
+M: x86.32 %vm-field ( dst field -- )
+    [ 0 [] MOV ] dip vm-field-offset rc-absolute-cell rel-vm ;
+
 M: x86.32 %vm-field-ptr ( dst field -- )
     [ 0 MOV ] dip vm-field-offset rc-absolute-cell rel-vm ;
 
@@ -102,6 +105,9 @@ M: x86.32 %prologue ( n -- )
     0 PUSH rc-absolute-cell rel-this
     3 cells - decr-stack-reg ;
 
+M: x86.32 %prepare-jump
+    pic-tail-reg 0 MOV xt-tail-pic-offset rc-absolute-cell rel-here ;
+
 M: x86.32 %load-param-reg
     stack-params assert=
     [ [ EAX ] dip local@ MOV ] dip
@@ -160,10 +166,10 @@ M: x86.32 %pop-stack ( n -- )
     EAX swap ds-reg reg-stack MOV ;
 
 M: x86.32 %pop-context-stack ( -- )
-    temp-reg %load-context-datastack
+    temp-reg "ctx" %vm-field
-    EAX temp-reg [] MOV
+    EAX temp-reg "datastack" context-field-offset [+] MOV
     EAX EAX [] MOV
-    temp-reg [] bootstrap-cell SUB ;
+    temp-reg "datastack" context-field-offset [+] bootstrap-cell SUB ;
 
 : call-unbox-func ( func -- )
     4 save-vm-ptr

basis/cpu/x86/32/bootstrap.factor

@@ -36,6 +36,11 @@ IN: bootstrap.x86
     ESP stack-frame-size 3 bootstrap-cells - SUB
 ] jit-prolog jit-define
 
+[
+    temp3 0 MOV rc-absolute-cell rt-here jit-rel
+    0 JMP rc-relative rt-entry-point-pic-tail jit-rel
+] jit-word-jump jit-define
+
 : jit-load-vm ( -- )
     vm-reg 0 MOV 0 rc-absolute-cell jit-vm ;
 

basis/cpu/x86/64/64.factor

@@ -42,17 +42,23 @@ M: x86.64 machine-registers
 M: x86.64 %mov-vm-ptr ( reg -- )
     vm-reg MOV ;
 
+M: x86.64 %vm-field ( dst field -- )
+    [ vm-reg ] dip vm-field-offset [+] MOV ;
+
 M: x86.64 %vm-field-ptr ( dst field -- )
     [ vm-reg ] dip vm-field-offset [+] LEA ;
 
 : param@ ( n -- op ) reserved-stack-space + stack@ ;
 
 M: x86.64 %prologue ( n -- )
-    temp-reg 0 MOV rc-absolute-cell rel-this
+    temp-reg -7 [] LEA
     dup PUSH
     temp-reg PUSH
     stack-reg swap 3 cells - SUB ;
 
+M: x86.64 %prepare-jump
+    pic-tail-reg xt-tail-pic-offset [] LEA ;
+
 : load-cards-offset ( dst -- )
     0 MOV rc-absolute-cell rel-cards-offset ;
 
@@ -104,10 +110,10 @@ M: x86.64 %pop-stack ( n -- )
     param-reg-0 swap ds-reg reg-stack MOV ;
 
 M: x86.64 %pop-context-stack ( -- )
-    temp-reg %load-context-datastack
+    temp-reg "ctx" %vm-field
-    param-reg-0 temp-reg [] MOV
+    param-reg-0 temp-reg "datastack" context-field-offset [+] MOV
     param-reg-0 param-reg-0 [] MOV
-    temp-reg [] bootstrap-cell SUB ;
+    temp-reg "datastack" context-field-offset [+] bootstrap-cell SUB ;
 
 M:: x86.64 %unbox ( n rep func -- )
     param-reg-1 %mov-vm-ptr

basis/cpu/x86/64/bootstrap.factor

@@ -28,7 +28,7 @@ IN: bootstrap.x86
 
 [
     ! load entry point
-    safe-reg 0 MOV rc-absolute-cell rt-this jit-rel
+    safe-reg -7 [] LEA
     ! save stack frame size
     stack-frame-size PUSH
     ! push entry point
@@ -37,6 +37,11 @@ IN: bootstrap.x86
     RSP stack-frame-size 3 bootstrap-cells - SUB
 ] jit-prolog jit-define
 
+[
+    temp3 5 [] LEA
+    0 JMP rc-relative rt-entry-point-pic-tail jit-rel
+] jit-word-jump jit-define
+
 : jit-load-context ( -- )
     ctx-reg vm-reg vm-context-offset [+] MOV ;
 

basis/cpu/x86/bootstrap.factor

@@ -76,11 +76,6 @@ big-endian off
     ds-reg [] temp0 MOV
 ] jit-push jit-define
 
-[
-    temp3 0 MOV rc-absolute-cell rt-here jit-rel
-    0 JMP rc-relative rt-entry-point-pic-tail jit-rel
-] jit-word-jump jit-define
-
 [
     0 CALL rc-relative rt-entry-point-pic jit-rel
 ] jit-word-call jit-define

basis/cpu/x86/x86.factor

@@ -88,8 +88,10 @@ M: x86 %call ( word -- ) 0 CALL rc-relative rel-word-pic ;
     #! See the comment in vm/cpu-x86.hpp
     4 1 + ; inline
 
+HOOK: %prepare-jump cpu ( -- )
+
 M: x86 %jump ( word -- )
-    pic-tail-reg 0 MOV xt-tail-pic-offset rc-absolute-cell rel-here
+    %prepare-jump
     0 JMP rc-relative rel-word-pic-tail ;
 
 M: x86 %jump-label ( label -- ) 0 JMP rc-relative label-fixup ;
@@ -474,17 +476,10 @@ M: x86 %push-stack ( -- )
     ds-reg cell ADD
     ds-reg [] int-regs return-reg MOV ;
 
-:: %load-context-datastack ( dst -- )
-    ! Load context struct
-    dst "ctx" %vm-field-ptr
-    dst dst [] MOV
-    ! Load context datastack pointer
-    dst "datastack" context-field-offset ADD ;
-
 M: x86 %push-context-stack ( -- )
-    temp-reg %load-context-datastack
+    temp-reg "ctx" %vm-field
-    temp-reg [] bootstrap-cell ADD
+    temp-reg "datastack" context-field-offset [+] bootstrap-cell ADD
-    temp-reg temp-reg [] MOV
+    temp-reg temp-reg "datastack" context-field-offset [+] MOV
     temp-reg [] int-regs return-reg MOV ;
 
 M: x86 %epilogue ( n -- ) cell - incr-stack-reg ;
@@ -1409,8 +1404,7 @@ M: x86 %loop-entry 16 code-alignment [ NOP ] times ;
 M:: x86 %restore-context ( temp1 temp2 -- )
     #! Load Factor stack pointers on entry from C to Factor.
     #! Also save callstack bottom!
-    temp1 "ctx" %vm-field-ptr
+    temp1 "ctx" %vm-field
-    temp1 temp1 [] MOV
     temp2 stack-reg stack-frame get total-size>> cell - [+] LEA
     temp1 "callstack-bottom" context-field-offset [+] temp2 MOV
     ds-reg temp1 "datastack" context-field-offset [+] MOV
@@ -1420,8 +1414,7 @@ M:: x86 %save-context ( temp1 temp2 -- )
     #! Save Factor stack pointers in case the C code calls a
     #! callback which does a GC, which must reliably trace
     #! all roots.
-    temp1 "ctx" %vm-field-ptr
+    temp1 "ctx" %vm-field
-    temp1 temp1 [] MOV
     temp2 stack-reg cell neg [+] LEA
     temp1 "callstack-top" context-field-offset [+] temp2 MOV
     temp1 "datastack" context-field-offset [+] ds-reg MOV

vm/jit.hpp

@@ -33,7 +33,9 @@ struct jit {
 	void word_jump(cell word_)
 	{
 		data_root<word> word(word_,parent);
+#ifndef FACTOR_AMD64
 		literal(tag_fixnum(xt_tail_pic_offset));
+#endif
 		literal(word.value());
 		emit(parent->special_objects[JIT_WORD_JUMP]);
 	}