X-Git-Url: https://notaz.gp2x.de/cgi-bin/gitweb.cgi?a=blobdiff_plain;f=libpcsxcore%2Fnew_dynarec%2Flinkage_arm.S;h=7e0db2d7a75c3cc13b9e073a3b782b64b6c593aa;hb=9b495f6ec3f28cf5ed1d41f6af16a9967fcf3e64;hp=d409aff5e56997980c595d3cb596bee84c6ab646;hpb=7f94b097581d5833921f6db86f18c7ef1203aa62;p=pcsx_rearmed.git

diff --git a/libpcsxcore/new_dynarec/linkage_arm.S b/libpcsxcore/new_dynarec/linkage_arm.S
index d409aff5..7e0db2d7 100644
--- a/libpcsxcore/new_dynarec/linkage_arm.S
+++ b/libpcsxcore/new_dynarec/linkage_arm.S
@@ -26,14 +26,14 @@
 
 #ifdef __MACH__
 #define dynarec_local		ESYM(dynarec_local)
-#define add_jump_out		ESYM(add_jump_out)
-#define new_recompile_block	ESYM(new_recompile_block)
-#define get_addr		ESYM(get_addr)
-#define get_addr_ht		ESYM(get_addr_ht)
-#define clean_blocks		ESYM(clean_blocks)
+#define ndrc_add_jump_out	ESYM(ndrc_add_jump_out)
+#define ndrc_try_restore_block	ESYM(ndrc_try_restore_block)
+#define ndrc_get_addr_ht	ESYM(ndrc_get_addr_ht)
+#define ndrc_get_addr_ht_param	ESYM(ndrc_get_addr_ht_param)
+#define ndrc_write_invalidate_one ESYM(ndrc_write_invalidate_one)
 #define gen_interupt		ESYM(gen_interupt)
-#define invalidate_addr		ESYM(invalidate_addr)
 #define gteCheckStallRaw	ESYM(gteCheckStallRaw)
+#define psxException		ESYM(psxException)
 #endif
 
 	.bss
@@ -87,19 +87,6 @@ DRC_VAR(invc_ptr, 4)
 DRC_VAR(scratch_buf_ptr, 4)
 DRC_VAR(ram_offset, 4)
 DRC_VAR(mini_ht, 256)
-DRC_VAR(restore_candidate, 512)
-
-
-#ifdef TEXRELS_FORBIDDEN
-	.data
-	.align 2
-ptr_jump_in:
-	.word	ESYM(jump_in)
-ptr_jump_dirty:
-	.word	ESYM(jump_dirty)
-ptr_hash_table:
-	.word	ESYM(hash_table)
-#endif
 
 
 	.syntax unified
@@ -158,46 +145,28 @@ ptr_hash_table:
 #endif
 .endm
 
-/* r0 = virtual target address */
-/* r1 = instruction to patch */
-.macro dyna_linker_main
+FUNCTION(dyna_linker):
+	/* r0 = virtual target address */
+	/* r1 = pointer to an instruction to patch */
 #ifndef NO_WRITE_EXEC
-	load_varadr_ext r3, jump_in
-	/* get_page */
-	lsr     r2, r0, #12
-	mov	r6, #4096
-	bic	r2, r2, #0xe0000
-	sub	r6, r6, #1
-	cmp     r2, #0x1000
 	ldr	r7, [r1]
-	biclt   r2, #0x0e00
-	and	r6, r6, r2
-	cmp	r2, #2048
-	add	r12, r7, #2
-	orrcs	r2, r6, #2048
-	ldr	r5, [r3, r2, lsl #2]
-	lsl	r12, r12, #8
-	add	r6, r1, r12, asr #6  /* old target */
-	mov	r8, #0
-	/* jump_in lookup */
-1:
-	movs	r4, r5
-	beq	2f
-	ldr	r3, [r5]	 /* ll_entry .vaddr */
-	ldrd	r4, r5, [r4, #8] /* ll_entry .next, .addr */
-	teq	r3, r0
-	bne	1b
-	teq	r4, r6
-	moveq	pc, r4 /* Stale i-cache */
-	mov	r8, r4
-	b  	1b     /* jump_in may have dupes, continue search */
-2:
-	tst     r8, r8
-	beq	3f     /* r0 not in jump_in */
-
+	mov	r4, r0
+	add	r6, r7, #2
 	mov	r5, r1
+	lsl	r6, r6, #8
+	/* must not compile - that might expire the caller block */
+	mov	r1, #0
+	bl	ndrc_get_addr_ht_param
+
+	movs	r8, r0
+	beq	0f
+	add	r6, r5, r6, asr #6  /* old target */
+	teq	r0, r6
+	moveq	pc, r0 /* Stale i-cache */
+	mov	r0, r4
 	mov	r1, r6
-	bl	add_jump_out
+	bl	ndrc_add_jump_out
+
 	sub	r2, r8, r5
 	and	r1, r7, #0xff000000
 	lsl	r2, r2, #6
@@ -205,246 +174,74 @@ ptr_hash_table:
 	add	r1, r1, r2, lsr #8
 	str	r1, [r5]
 	mov	pc, r8
-3:
-	/* hash_table lookup */
-	cmp	r2, #2048
-	load_varadr_ext r3, jump_dirty
-	eor	r4, r0, r0, lsl #16
-	lslcc	r2, r0, #9
-	load_varadr_ext r6, hash_table
-	lsr	r4, r4, #12
-	lsrcc	r2, r2, #21
-	bic	r4, r4, #15
-	ldr	r5, [r3, r2, lsl #2]
-	ldr	r7, [r6, r4]!
-	teq	r7, r0
-	ldreq	pc, [r6, #8]
-	ldr	r7, [r6, #4]
-	teq	r7, r0
-	ldreq	pc, [r6, #12]
-	/* jump_dirty lookup */
-6:
-	movs	r4, r5
-	beq	8f
-	ldr	r3, [r5]
-	ldr	r5, [r4, #12]
-	teq	r3, r0
-	bne	6b
-7:
-	ldr	r1, [r4, #8]
-	/* hash_table insert */
-	ldr	r2, [r6]
-	ldr	r3, [r6, #8]
-	str	r0, [r6]
-	str	r1, [r6, #8]
-	str	r2, [r6, #4]
-	str	r3, [r6, #12]
-	mov	pc, r1
-8:
+0:
+	mov	r0, r4
 #else
 	/* XXX: should be able to do better than this... */
-	bl	get_addr_ht
-	mov	pc, r0
 #endif
-.endm
-
-
-FUNCTION(dyna_linker):
-	/* r0 = virtual target address */
-	/* r1 = instruction to patch */
-	dyna_linker_main
-
-	mov	r4, r0
-	mov	r5, r1
-	bl	new_recompile_block
-	tst	r0, r0
-	mov	r0, r4
-	mov	r1, r5
-	beq	dyna_linker
-	/* pagefault */
-	mov	r1, r0
-	mov	r2, #8
-	.size	dyna_linker, .-dyna_linker
-
-FUNCTION(exec_pagefault):
-	/* r0 = instruction pointer */
-	/* r1 = fault address */
-	/* r2 = cause */
-	ldr	r3, [fp, #LO_reg_cop0+48] /* Status */
-	mvn	r6, #0xF000000F
-	ldr	r4, [fp, #LO_reg_cop0+16] /* Context */
-	bic	r6, r6, #0x0F800000
-	str	r0, [fp, #LO_reg_cop0+56] /* EPC */
-	orr	r3, r3, #2
-	str	r1, [fp, #LO_reg_cop0+32] /* BadVAddr */
-	bic	r4, r4, r6
-	str	r3, [fp, #LO_reg_cop0+48] /* Status */
-	and	r5, r6, r1, lsr #9
-	str	r2, [fp, #LO_reg_cop0+52] /* Cause */
-	and	r1, r1, r6, lsl #9
-	str	r1, [fp, #LO_reg_cop0+40] /* EntryHi */
-	orr	r4, r4, r5
-	str	r4, [fp, #LO_reg_cop0+16] /* Context */
-	mov	r0, #0x80000000
-	bl	get_addr_ht
+	bl	ndrc_get_addr_ht
 	mov	pc, r0
-	.size	exec_pagefault, .-exec_pagefault
-
-/* Special dynamic linker for the case where a page fault
-   may occur in a branch delay slot */
-FUNCTION(dyna_linker_ds):
-	/* r0 = virtual target address */
-	/* r1 = instruction to patch */
-	dyna_linker_main
-
-	mov	r4, r0
-	bic	r0, r0, #7
-	mov	r5, r1
-	orr	r0, r0, #1
-	bl	new_recompile_block
-	tst	r0, r0
-	mov	r0, r4
-	mov	r1, r5
-	beq	dyna_linker_ds
-	/* pagefault */
-	bic	r1, r0, #7
-	mov	r2, #0x80000008 /* High bit set indicates pagefault in delay slot */
-	sub	r0, r1, #4
-	b	exec_pagefault
-	.size	dyna_linker_ds, .-dyna_linker_ds
+	.size	dyna_linker, .-dyna_linker
 
 	.align	2
-
-FUNCTION(jump_vaddr_r0):
-	eor	r2, r0, r0, lsl #16
-	b	jump_vaddr
-	.size	jump_vaddr_r0, .-jump_vaddr_r0
 FUNCTION(jump_vaddr_r1):
-	eor	r2, r1, r1, lsl #16
 	mov	r0, r1
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r1, .-jump_vaddr_r1
 FUNCTION(jump_vaddr_r2):
 	mov	r0, r2
-	eor	r2, r2, r2, lsl #16
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r2, .-jump_vaddr_r2
 FUNCTION(jump_vaddr_r3):
-	eor	r2, r3, r3, lsl #16
 	mov	r0, r3
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r3, .-jump_vaddr_r3
 FUNCTION(jump_vaddr_r4):
-	eor	r2, r4, r4, lsl #16
 	mov	r0, r4
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r4, .-jump_vaddr_r4
 FUNCTION(jump_vaddr_r5):
-	eor	r2, r5, r5, lsl #16
 	mov	r0, r5
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r5, .-jump_vaddr_r5
 FUNCTION(jump_vaddr_r6):
-	eor	r2, r6, r6, lsl #16
 	mov	r0, r6
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r6, .-jump_vaddr_r6
 FUNCTION(jump_vaddr_r8):
-	eor	r2, r8, r8, lsl #16
 	mov	r0, r8
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r8, .-jump_vaddr_r8
 FUNCTION(jump_vaddr_r9):
-	eor	r2, r9, r9, lsl #16
 	mov	r0, r9
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r9, .-jump_vaddr_r9
 FUNCTION(jump_vaddr_r10):
-	eor	r2, r10, r10, lsl #16
 	mov	r0, r10
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r10, .-jump_vaddr_r10
 FUNCTION(jump_vaddr_r12):
-	eor	r2, r12, r12, lsl #16
 	mov	r0, r12
-	b	jump_vaddr
+	b	jump_vaddr_r0
 	.size	jump_vaddr_r12, .-jump_vaddr_r12
 FUNCTION(jump_vaddr_r7):
-	eor	r2, r7, r7, lsl #16
 	add	r0, r7, #0
 	.size	jump_vaddr_r7, .-jump_vaddr_r7
-FUNCTION(jump_vaddr):
-	load_varadr_ext r1, hash_table
-	mvn	r3, #15
-	and	r2, r3, r2, lsr #12
-	ldr	r2, [r1, r2]!
-	teq	r2, r0
-	ldreq	pc, [r1, #8]
-	ldr	r2, [r1, #4]
-	teq	r2, r0
-	ldreq	pc, [r1, #12]
-	str	r10, [fp, #LO_cycle_count]
-	bl	get_addr
-	ldr	r10, [fp, #LO_cycle_count]
-	mov	pc, r0
-	.size	jump_vaddr, .-jump_vaddr
-
-	.align	2
-
-FUNCTION(verify_code_ds):
-	str	r8, [fp, #LO_branch_target]  @ preserve HOST_BTREG?
-FUNCTION(verify_code):
-	/* r1 = source */
-	/* r2 = target */
-	/* r3 = length */
-	tst	r3, #4
-	mov	r4, #0
-	add	r3, r1, r3
-	mov	r5, #0
-	ldrne	r4, [r1], #4
-	mov	r12, #0
-	ldrne	r5, [r2], #4
-	teq	r1, r3
-	beq	.D3
-.D2:
-	ldr	r7, [r1], #4
-	eor	r9, r4, r5
-	ldr	r8, [r2], #4
-	orrs	r9, r9, r12
-	bne	.D4
-	ldr	r4, [r1], #4
-	eor	r12, r7, r8
-	ldr	r5, [r2], #4
-	cmp	r1, r3
-	bcc	.D2
-	teq	r7, r8
-.D3:
-	teqeq	r4, r5
-.D4:
-	ldr	r8, [fp, #LO_branch_target]
-	moveq	pc, lr
-.D5:
-	bl	get_addr
+FUNCTION(jump_vaddr_r0):
+	bl	ndrc_get_addr_ht
 	mov	pc, r0
-	.size	verify_code, .-verify_code
-	.size	verify_code_ds, .-verify_code_ds
+	.size	jump_vaddr_r0, .-jump_vaddr_r0
 
 	.align	2
 FUNCTION(cc_interrupt):
 	ldr	r0, [fp, #LO_last_count]
 	mov	r1, #0
-	mov	r2, #0x1fc
 	add	r10, r0, r10
 	str	r1, [fp, #LO_pending_exception]
-	and	r2, r2, r10, lsr #17
-	add	r3, fp, #LO_restore_candidate
 	str	r10, [fp, #LO_cycle]		/* PCSX cycles */
-@@	str	r10, [fp, #LO_reg_cop0+36]	/* Count */
-	ldr	r4, [r2, r3]
+@@	str	r10, [fp, #LO_reg_cop0+36]	/* Count - not on PSX */
 	mov	r10, lr
-	tst	r4, r4
-	bne	.E4
-.E1:
+
 	bl	gen_interupt
 	mov	lr, r10
 	ldr	r10, [fp, #LO_cycle]
@@ -457,22 +254,9 @@ FUNCTION(cc_interrupt):
 	ldmfdne	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, pc}
 	tst	r1, r1
 	moveq	pc, lr
-.E2:
 	ldr	r0, [fp, #LO_pcaddr]
-	bl	get_addr_ht
+	bl	ndrc_get_addr_ht
 	mov	pc, r0
-.E4:
-	/* Move 'dirty' blocks to the 'clean' list */
-	lsl	r5, r2, #3
-	str	r1, [r2, r3]
-.E5:
-	lsrs	r4, r4, #1
-	mov	r0, r5
-	add	r5, r5, #1
-	blcs	clean_blocks
-	tst	r5, #31
-	bne	.E5
-	b	.E1
 	.size	cc_interrupt, .-cc_interrupt
 
 	.align	2
@@ -487,7 +271,7 @@ FUNCTION(fp_exception):
 	str	r1, [fp, #LO_reg_cop0+48] /* Status */
 	str	r2, [fp, #LO_reg_cop0+52] /* Cause */
 	add	r0, r3, #0x80
-	bl	get_addr_ht
+	bl	ndrc_get_addr_ht
 	mov	pc, r0
 	.size	fp_exception, .-fp_exception
 	.align	2
@@ -497,19 +281,28 @@ FUNCTION(fp_exception_ds):
 	.size	fp_exception_ds, .-fp_exception_ds
 
 	.align	2
+FUNCTION(jump_break_ds):
+	mov	r0, #0x24
+	mov	r1, #1
+	b	call_psxException
+FUNCTION(jump_break):
+	mov	r0, #0x24
+	mov	r1, #0
+	b	call_psxException
+FUNCTION(jump_syscall_ds):
+	mov	r0, #0x20
+	mov	r1, #1
+	b	call_psxException
 FUNCTION(jump_syscall):
-	ldr	r1, [fp, #LO_reg_cop0+48] /* Status */
-	mov	r3, #0x80000000
-	str	r0, [fp, #LO_reg_cop0+56] /* EPC */
-	orr	r1, #2
-	mov	r2, #0x20
-	str	r1, [fp, #LO_reg_cop0+48] /* Status */
-	str	r2, [fp, #LO_reg_cop0+52] /* Cause */
-	add	r0, r3, #0x80
-	bl	get_addr_ht
-	mov	pc, r0
-	.size	jump_syscall, .-jump_syscall
-	.align	2
+	mov	r0, #0x20
+	mov	r1, #0
+
+call_psxException:
+	ldr	r3, [fp, #LO_last_count]
+	str	r2, [fp, #LO_pcaddr]
+	add	r10, r3, r10
+	str	r10, [fp, #LO_cycle]		/* PCSX cycles */
+	bl	psxException
 
 	/* note: psxException might do recursive recompiler call from it's HLE code,
 	 * so be ready for this */
@@ -519,7 +312,7 @@ FUNCTION(jump_to_new_pc):
 	ldr	r0, [fp, #LO_pcaddr]
 	sub	r10, r10, r1
 	str	r1, [fp, #LO_last_count]
-	bl	get_addr_ht
+	bl	ndrc_get_addr_ht
 	mov	pc, r0
 	.size	jump_to_new_pc, .-jump_to_new_pc
 
@@ -608,7 +401,7 @@ invalidate_addr_call:
 	ldr	lr, [fp, #LO_inv_code_end]
 	cmp	r0, r12
 	cmpcs   lr, r0
-	blcc	invalidate_addr
+	blcc	ndrc_write_invalidate_one
 	ldmia	fp, {r0, r1, r2, r3, EXTRA_UNSAVED_REGS r12, pc}
 	.size	invalidate_addr_call, .-invalidate_addr_call
 
@@ -618,7 +411,7 @@ FUNCTION(new_dyna_start):
 	stmfd	sp!, {r4, r5, r6, r7, r8, r9, sl, fp, ip, lr}
 	mov	fp, r0 /* dynarec_local */
 	ldr	r0, [fp, #LO_pcaddr]
-	bl	get_addr_ht
+	bl	ndrc_get_addr_ht
 	ldr	r1, [fp, #LO_next_interupt]
 	ldr	r10, [fp, #LO_cycle]
 	str	r1, [fp, #LO_last_count]
@@ -838,4 +631,44 @@ FUNCTION(call_gteStall):
 	add	r10, r10, r0
 	bx	lr
 
+#ifdef HAVE_ARMV6
+
+FUNCTION(get_reg):
+	ldr	r12, [r0]
+	and	r1, r1, #0xff
+	ldr	r2, [r0, #4]
+	orr	r1, r1, r1, lsl #8
+	ldr	r3, [r0, #8]
+	orr	r1, r1, r1, lsl #16   @ searched char in every byte
+	ldrb    r0, [r0, #12]         @ last byte
+	eor	r12, r12, r1
+	eor	r2, r2, r1
+	eor	r3, r3, r1
+	cmp     r0, r1, lsr #24
+	mov	r0, #12
+	mvn	r1, #0                @ r1=~0
+	bxeq	lr
+	orr	r3, r3, #0xff000000   @ EXCLUDE_REG
+	uadd8	r0, r12, r1           @ add and set GE bits when not 0 (match)
+	mov	r12, #0
+	sel	r0, r12, r1           @ 0 if no match, else ff in some byte
+	uadd8	r2, r2, r1
+	sel	r2, r12, r1
+	uadd8	r3, r3, r1
+	sel	r3, r12, r1
+	mov	r12, #3
+	clz     r0, r0                @ 0, 8, 16, 24 or 32
+	clz     r2, r2
+	clz     r3, r3
+	sub	r0, r12, r0, lsr #3   @ 3, 2, 1, 0 or -1
+	sub	r2, r12, r2, lsr #3
+	sub	r3, r12, r3, lsr #3
+	orr	r2, r2, #4
+	orr	r3, r3, #8
+	and	r0, r0, r2
+	and	r0, r0, r3
+	bx	lr
+
+#endif /* HAVE_ARMV6 */
+
 @ vim:filetype=armasm