Merge branch 'master' of git://factorcode.org/git/factor

basis/cpu/ppc/bootstrap.factor

@@ -9,8 +9,8 @@ IN: bootstrap.ppc
 4 \ cell set
 big-endian on
 
-CONSTANT: ds-reg 29
+CONSTANT: ds-reg 13
-CONSTANT: rs-reg 30
+CONSTANT: rs-reg 14
 
 : factor-area-size ( -- n ) 4 bootstrap-cells ;
 
@@ -138,6 +138,16 @@ CONSTANT: rs-reg 30
     jit-3r>
 ] jit-3dip jit-define
 
+: prepare-(execute) ( -- operand )
+    3 ds-reg 0 LWZ
+    ds-reg dup 4 SUBI
+    4 3 word-xt-offset LWZ
+    4 ;
+
+[ prepare-(execute) MTCTR BCTR ] jit-execute-jump jit-define
+
+[ prepare-(execute) MTLR BLRL ] jit-execute-call jit-define
+
 [
     0 1 lr-save stack-frame + LWZ
     1 1 stack-frame ADDI
@@ -146,7 +156,91 @@ CONSTANT: rs-reg 30
 
 [ BLR ] jit-return jit-define
 
-! Sub-primitives
+! ! ! Polymorphic inline caches
+
+! Load a value from a stack position
+[
+    4 ds-reg 0 LWZ rc-absolute-ppc-2 rt-untagged jit-rel
+] pic-load jit-define
+
+! Tag
+: load-tag ( -- )
+    4 4 tag-mask get ANDI
+    4 4 tag-bits get SLWI ;
+
+[ load-tag ] pic-tag jit-define
+
+! Hi-tag
+[
+    3 4 MR
+    load-tag
+    0 4 object tag-number tag-fixnum CMPI
+    2 BNE
+    4 3 object tag-number neg LWZ
+] pic-hi-tag jit-define
+
+! Tuple
+[
+    3 4 MR
+    load-tag
+    0 4 tuple tag-number tag-fixnum CMPI
+    2 BNE
+    4 3 tuple tag-number neg bootstrap-cell + LWZ
+] pic-tuple jit-define
+
+! Hi-tag and tuple
+[
+    3 4 MR
+    load-tag
+    ! If bits 2 and 3 are set, the tag is either 6 (object) or 7 (tuple)
+    0 4 BIN: 110 tag-fixnum CMPI
+    5 BLT
+    ! Untag r3
+    3 3 0 0 31 tag-bits get - RLWINM
+    ! Set r4 to 0 for objects, and bootstrap-cell for tuples
+    4 4 1 tag-fixnum ANDI
+    4 4 1 SRAWI
+    ! Load header cell or tuple layout cell
+    4 4 3 LWZX
+] pic-hi-tag-tuple jit-define
+
+[
+    0 4 0 CMPI rc-absolute-ppc-2 rt-immediate jit-rel
+] pic-check-tag jit-define
+
+[
+    0 5 LOAD32 rc-absolute-ppc-2/2 rt-immediate jit-rel
+    4 0 5 CMP
+] pic-check jit-define
+
+[ 2 BNE 0 B rc-relative-ppc-3 rt-xt jit-rel ] pic-hit jit-define
+
+! ! ! Megamorphic caches
+
+[
+    ! cache = ...
+    0 3 LOAD32 rc-absolute-ppc-2/2 rt-immediate jit-rel
+    ! key = class
+    5 4 MR
+    ! key &= cache.length - 1
+    5 5 mega-cache-size get 1- bootstrap-cell * ANDI
+    ! cache += array-start-offset
+    3 3 array-start-offset ADDI
+    ! cache += key
+    3 3 5 ADD
+    ! if(get(cache) == class)
+    6 3 0 LWZ
+    6 0 4 CMP
+    5 BNE
+    ! ... goto get(cache + bootstrap-cell)
+    3 3 4 LWZ
+    3 3 word-xt-offset LWZ
+    3 MTCTR
+    BCTR
+    ! fall-through on miss
+] mega-lookup jit-define
+
+! ! ! Sub-primitives
 
 ! Quotations and words
 [
@@ -157,14 +251,6 @@ CONSTANT: rs-reg 30
     BCTR
 ] \ (call) define-sub-primitive
 
-[
-    3 ds-reg 0 LWZ
-    ds-reg dup 4 SUBI
-    4 3 word-xt-offset LWZ
-    4 MTCTR
-    BCTR
-] \ (execute) define-sub-primitive
-
 ! Objects
 [
     3 ds-reg 0 LWZ

basis/cpu/ppc/ppc.factor

@@ -1,20 +1,19 @@
-! Copyright (C) 2005, 2008 Slava Pestov.
+! Copyright (C) 2005, 2009 Slava Pestov.
 ! See http://factorcode.org/license.txt for BSD license.
 USING: accessors assocs sequences kernel combinators make math
 math.order math.ranges system namespaces locals layouts words
-alien alien.c-types cpu.architecture cpu.ppc.assembler
+alien alien.c-types literals cpu.architecture cpu.ppc.assembler
-compiler.cfg.registers compiler.cfg.instructions
+literals compiler.cfg.registers compiler.cfg.instructions
 compiler.constants compiler.codegen compiler.codegen.fixup
 compiler.cfg.intrinsics compiler.cfg.stack-frame ;
 IN: cpu.ppc
 
 ! PowerPC register assignments:
-! r2-r27: integer vregs
+! r2-r12: integer vregs
-! r28: integer scratch
+! r15-r29
-! r29: data stack
+! r30: integer scratch
-! r30: retain stack
 ! f0-f29: float vregs
-! f30, f31: float scratch
+! f30: float scratch
 
 enable-float-intrinsics
 
@@ -23,11 +22,11 @@ enable-float-intrinsics
 
 M: ppc machine-registers
     {
-        { int-regs T{ range f 2 26 1 } }
+        { int-regs $[ 2 12 [a,b] 15 29 [a,b] append ] }
-        { double-float-regs T{ range f 0 29 1 } }
+        { double-float-regs $[ 0 29 [a,b] ] }
     } ;
 
-CONSTANT: scratch-reg 28
+CONSTANT: scratch-reg 30
 CONSTANT: fp-scratch-reg 30
 
 M: ppc two-operand? f ;
@@ -40,8 +39,8 @@ M: ppc %load-reference ( reg obj -- )
 M: ppc %alien-global ( register symbol dll -- )
     [ 0 swap LOAD32 ] 2dip rc-absolute-ppc-2/2 rel-dlsym ;
 
-CONSTANT: ds-reg 29
+CONSTANT: ds-reg 13
-CONSTANT: rs-reg 30
+CONSTANT: rs-reg 14
 
 GENERIC: loc-reg ( loc -- reg )
 

basis/cpu/x86/bootstrap.factor

@@ -194,7 +194,7 @@ big-endian off
     [
         ! Untag temp0
         temp0 tag-mask get bitnot AND
-        ! Set temp1 to 0 for objects, and 8 for tuples
+        ! Set temp1 to 0 for objects, and bootstrap-cell for tuples
         temp1 1 tag-fixnum AND
         bootstrap-cell 4 = [ temp1 1 SHR ] when
         ! Load header cell or tuple layout cell

basis/math/miller-rabin/miller-rabin.factor

@@ -8,6 +8,8 @@ IN: math.miller-rabin
 
 : >odd ( n -- int ) dup even? [ 1 + ] when ; foldable
 
+: >even ( n -- int ) 0 clear-bit ; foldable
+
 TUPLE: positive-even-expected n ;
 
 :: (miller-rabin) ( n trials -- ? )
@@ -97,6 +99,7 @@ PRIVATE>
     } 1&& ;
 
 : next-safe-prime ( n -- q )
+    1 - >even 2 /
     next-safe-prime-candidate
     dup >safe-prime-form
     dup miller-rabin

basis/tools/time/time.factor

@@ -1,6 +1,6 @@
 ! Copyright (C) 2003, 2008 Slava Pestov.
 ! See http://factorcode.org/license.txt for BSD license.
-USING: kernel math math.vectors memory io io.styles prettyprint
+USING: kernel math memory io io.styles prettyprint
 namespaces system sequences splitting grouping assocs strings
 generic.single combinators ;
 IN: tools.time

core/hashtables/hashtables.factor

@@ -139,14 +139,14 @@ M: hashtable set-at ( value key hash -- )
 PRIVATE>
 
 M: hashtable >alist
-    [ array>> [ length 2/ iota ] keep ] [ assoc-size <vector> ] bi [
+    [ array>> [ length 2/ ] keep ] [ assoc-size <vector> ] bi [
         [
             [
                 [ 1 fixnum-shift-fast ] dip
                 [ array-nth ] [ [ 1 fixnum+fast ] dip array-nth ] 2bi
             ] dip
             pick tombstone? [ 3drop ] [ [ 2array ] dip push-unsafe ] if
-        ] 2curry each
+        ] 2curry each-integer
     ] keep { } like ;
 
 M: hashtable clone

vm/cpu-ppc.S

@@ -2,7 +2,7 @@
 in the public domain. */
 #include "asm.h"
 
-#define DS_REG r29
+#define DS_REG r13
 
 DEF(void,primitive_fixnum_add,(void)):
 	lwz r3,0(DS_REG)
@@ -45,7 +45,7 @@ multiply_overflow:
 	
 /* Note that the XT is passed to the quotation in r11 */
 #define CALL_OR_JUMP_QUOT \
-	lwz r11,14(r3)	   /* load quotation-xt slot */ XX \
+	lwz r11,16(r3)	   /* load quotation-xt slot */ XX \
 
 #define CALL_QUOT \
 	CALL_OR_JUMP_QUOT XX \
@@ -100,22 +100,22 @@ the Factor compiler treats the entire register file as volatile. */
 DEF(void,c_to_factor,(CELL quot)):
 	PROLOGUE
 
-	SAVE_INT(r13,0)	   /* save GPRs */
+	SAVE_INT(r15,0)	   /* save GPRs */
-	SAVE_INT(r14,1)
+	SAVE_INT(r16,1)
-	SAVE_INT(r15,2)
+	SAVE_INT(r17,2)
-	SAVE_INT(r16,3)
+	SAVE_INT(r18,3)
-	SAVE_INT(r17,4)
+	SAVE_INT(r19,4)
-	SAVE_INT(r18,5)
+	SAVE_INT(r20,5)
-	SAVE_INT(r19,6)
+	SAVE_INT(r21,6)
-	SAVE_INT(r20,7)
+	SAVE_INT(r22,7)
-	SAVE_INT(r21,8)
+	SAVE_INT(r23,8)
-	SAVE_INT(r22,9)
+	SAVE_INT(r24,9)
-	SAVE_INT(r23,10)
+	SAVE_INT(r25,10)
-	SAVE_INT(r24,11)
+	SAVE_INT(r26,11)
-	SAVE_INT(r25,12)
+	SAVE_INT(r27,12)
-	SAVE_INT(r26,13)
+	SAVE_INT(r28,13)
-	SAVE_INT(r27,14)
+	SAVE_INT(r29,14)
-	SAVE_INT(r28,15)
+	SAVE_INT(r30,15)
 	SAVE_INT(r31,16)
 
 	SAVE_FP(f14,20)	/* save FPRs */
@@ -165,22 +165,22 @@ DEF(void,c_to_factor,(CELL quot)):
 	RESTORE_FP(f14,20)	/* save FPRs */
 
 	RESTORE_INT(r31,16)   /* restore GPRs */
-	RESTORE_INT(r28,15)
+	RESTORE_INT(r30,15)
-	RESTORE_INT(r27,14)
+	RESTORE_INT(r29,14)
-	RESTORE_INT(r26,13)
+	RESTORE_INT(r28,13)
-	RESTORE_INT(r25,12)
+	RESTORE_INT(r27,12)
-	RESTORE_INT(r24,11)
+	RESTORE_INT(r26,11)
-	RESTORE_INT(r23,10)
+	RESTORE_INT(r25,10)
-	RESTORE_INT(r22,9)
+	RESTORE_INT(r24,9)
-	RESTORE_INT(r21,8)
+	RESTORE_INT(r23,8)
-	RESTORE_INT(r20,7)
+	RESTORE_INT(r22,7)
-	RESTORE_INT(r19,6)
+	RESTORE_INT(r21,6)
-	RESTORE_INT(r18,5)
+	RESTORE_INT(r20,5)
-	RESTORE_INT(r17,4)
+	RESTORE_INT(r19,4)
-	RESTORE_INT(r16,3)
+	RESTORE_INT(r18,3)
-	RESTORE_INT(r15,2)
+	RESTORE_INT(r17,2)
-	RESTORE_INT(r14,1)
+	RESTORE_INT(r16,1)
-	RESTORE_INT(r13,0)
+	RESTORE_INT(r15,0)
 
 	EPILOGUE
 	blr
@@ -234,3 +234,11 @@ DEF(void,flush_icache,(void *start, int len)):
 	sync		   /* finish up */
 	isync
 	blr
+
+DEF(void,primitive_inline_cache_miss,(void)):
+    mflr r3
+    PROLOGUE
+    bl MANGLE(inline_cache_miss)
+    EPILOGUE
+    mtctr r3
+    bctr

vm/cpu-ppc.hpp

@@ -2,16 +2,58 @@ namespace factor
 {
 
 #define FACTOR_CPU_STRING "ppc"
-#define VM_ASM_API
+#define VM_ASM_API VM_C_API
 
-register cell ds asm("r29");
+register cell ds asm("r13");
-register cell rs asm("r30");
+register cell rs asm("r14");
 
-void c_to_factor(cell quot);
+inline static void check_call_site(cell return_address)
-void undefined(cell word);
+{
-void set_callstack(stack_frame *to, stack_frame *from, cell length, void *memcpy);
+#ifdef FACTOR_DEBUG
-void throw_impl(cell quot, stack_frame *rewind);
+	cell insn = *(cell *)return_address;
-void lazy_jit_compile(cell quot);
+	assert((insn & 0x3) == 0x1);
-void flush_icache(cell start, cell len);
+	assert((insn >> 26) == 0x12);
+#endif
+}
+
+#define B_MASK 0x3fffffc
+
+inline static void *get_call_target(cell return_address)
+{
+	return_address -= sizeof(cell);
+
+	check_call_site(return_address);
+	cell insn = *(cell *)return_address;
+	cell unsigned_addr = (insn & B_MASK);
+	fixnum signed_addr = (fixnum)(unsigned_addr << 6) >> 6;
+	return (void *)(signed_addr + return_address);
+}
+
+inline static void set_call_target(cell return_address, void *target)
+{
+	return_address -= sizeof(cell);
+
+#ifdef FACTOR_DEBUG
+	assert((return_address & ~B_MASK) == 0);
+	check_call_site(return_address);
+#endif
+	cell insn = *(cell *)return_address;
+	insn = ((insn & ~B_MASK) | (((cell)target - return_address) & B_MASK));
+	*(cell *)return_address = insn;
+
+	/* Flush the cache line containing the call we just patched */
+	__asm__ __volatile__ ("icbi 0, %0\n" "sync\n"::"r" (return_address):);
+}
+
+/* Defined in assembly */
+VM_ASM_API void c_to_factor(cell quot);
+VM_ASM_API void throw_impl(cell quot, stack_frame *rewind);
+VM_ASM_API void lazy_jit_compile(cell quot);
+VM_ASM_API void flush_icache(cell start, cell len);
+
+VM_ASM_API void set_callstack(stack_frame *to,
+			       stack_frame *from,
+			       cell length,
+			       void *(*memcpy)(void*,const void*, size_t));
 
 }

vm/cpu-x86.32.S

@@ -30,7 +30,6 @@ and the callstack top is passed in EDX */
 	pop %ebx
 
 #define QUOT_XT_OFFSET 16
-#define WORD_XT_OFFSET 30
 
 /* We pass a function pointer to memcpy to work around a Mac OS X
 ABI limitation which would otherwise require us to do a bizzaro PC-relative

vm/cpu-x86.64.S

@@ -62,7 +62,6 @@
 #endif
 
 #define QUOT_XT_OFFSET 36
-#define WORD_XT_OFFSET 66
 
 /* We pass a function pointer to memcpy to work around a Mac OS X
 ABI limitation which would otherwise require us to do a bizzaro PC-relative

vm/inline_cache.cpp

@@ -21,6 +21,8 @@ void deallocate_inline_cache(cell return_address)
 {
 	/* Find the call target. */
 	void *old_xt = get_call_target(return_address);
+	check_code_pointer((cell)old_xt);
+
 	code_block *old_block = (code_block *)old_xt - 1;
 	cell old_type = old_block->type;