Module Name: src
Committed By: matt
Date: Fri Mar 21 14:03:30 UTC 2014
Modified Files:
src/libexec/ld.elf_so/arch/vax: rtld_start.S
Log Message:
Simplify. If we got called via a calls $n, *pcrel32, instead of constructing
a new stack frame, back up the PC by 7 and return back to the calls so it
will be reinvoked. (This is by far the most common way it gets invoked).
Otherwise rebuild a new callframe and jump to the routine.
To generate a diff of this commit:
cvs rdiff -u -r1.20 -r1.21 src/libexec/ld.elf_so/arch/vax/rtld_start.S
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/libexec/ld.elf_so/arch/vax/rtld_start.S
diff -u src/libexec/ld.elf_so/arch/vax/rtld_start.S:1.20 src/libexec/ld.elf_so/arch/vax/rtld_start.S:1.21
--- src/libexec/ld.elf_so/arch/vax/rtld_start.S:1.20 Wed Mar 19 21:52:00 2014
+++ src/libexec/ld.elf_so/arch/vax/rtld_start.S Fri Mar 21 14:03:30 2014
@@ -1,4 +1,4 @@
-/* $NetBSD: rtld_start.S,v 1.20 2014/03/19 21:52:00 joerg Exp $ */
+/* $NetBSD: rtld_start.S,v 1.21 2014/03/21 14:03:30 matt Exp $ */
/*
* Copyright 1996 Matt Thomas <[email protected]>
@@ -57,102 +57,218 @@ END(_rtld_start)
/*
* Lazy binding entry point, called via PLT via JMP into pltgot[1].
+ * SP+4: address to relocation offset
* SP+0: obj entry points
- * SP+4: address to relocation index
- *
- * Note: Some functions rely on there not being an additional call frame;
- * hence the `optimization' to avoid the callg opportunistically.
*/
ALTENTRY(_rtld_bind_start)
- movab -64(%sp),%sp /* reserve some space */
- pushr $0x3f /* save R0-R5 */
movq -8(%fp),%r0 /* get addresses of plt.got & reloc index */
- pushl (%r1) /* push relocation index */
+ pushl (%r1) /* push relocation offset */
pushl %r0 /* push address of obj entry */
calls $2,_rtld_bind
- addl3 $2,%r0,%r3 /* save routine address */
- extzv $0,$12,(%r0),%r1 /* get entry mask */
- extzv $0,$12,6(%fp),%r2 /* get saved mask */
- cmpw %r1,%r2 /* compare them */
- bneq 12f /* if they are different, rebuild */
- movl %r3,-4(%fp) /* save routine address */
- popr $0x3f /* pop registers */
- movab 68(%sp),%sp /* restore sp */
- rsb /* and jump to it */
-
- /*
- * We need to rebuild the callframe. Save the current one in case
- * we might overwrite it.
- */
-12: movq 4(%fp),-(%sp) /* save PSW and AP */
- movq 12(%fp),-(%sp) /* save FP and return address */
- /*
- * Find out where this this call frame ends.
- */
- movl %ap,%r0 /* get past callframe and registers */
- bbs $29,4(%fp),22f /* calls is easy, it's where AP is */
- /*
- * Callg not so much
- */
- movab 20(%fp),%r0 /* past fixed callframe */
- tstw %r2 /* no saved registers? */
- beql 22f /* none, so we are done. */
- movl $11,%r4 /* start with register 11 */
-20: bbc %r4,%r2,21f /* save this register? */
- addl2 $4,%r0 /* yes, adjust for saved register */
-21: sobgeq %r4,20b /* try next register */
-
-22:
- /*
- * First "push" the caller saved registers (if there any that
- * need to saved.)
- */
- tstw %r1 /* if there are no registers to save */
- beql 1f /* just push the callframe */
- cmpw %r1,$63 /* if there are no caller-saved registers */
- blequ 5f /* skip them */
- bbc $11,%r1,10f /* does it need to be saved? */
- movl %r11,-(%r0)
-10: bbc $10,%r1,9f /* does it need to be saved? */
- movl %r10,-(%r0)
-9: bbc $9,%r1,8f /* does it need to be saved? */
- movl %r9,-(%r0)
-8: bbc $8,%r1,7f /* does it need to be saved? */
- movl %r8,-(%r0)
-7: bbc $7,%r1,6f /* does it need to be saved? */
- movl %r7,-(%r0)
-6: bbc $6,%r1,5f /* does it need to be saved? */
- movl %r6,-(%r0)
-5:
- /*
- * r0-r5 are not normally preserved so we should be done.
- */
- cmpw %r1,$63
- bgtru 1f
- /*
- * For some reason, we have to preserve these.
- */
- movab 16(%sp),%r2
- bbc $5,%r1,4f /* does it need to be saved? */
- movl 20(%r2),-(%r0)
-4: bbc $4,%r1,3f /* does it need to be saved? */
- movl 16(%r2),-(%r0)
-3: bbc $3,%r1,2f /* does it need to be saved? */
- movl 12(%r2),-(%r0)
-2: bbc $2,%r1,1f /* does it need to be saved? */
- movl 8(%r2),-(%r0)
-
- /*
- * Now we save the fixed part of the callframe.
- */
-1: clrl %r4 /* clear condition handler slot */
- movq (%sp)+,-(%r0) /* move FP and PC into place */
- movq (%sp)+,-(%r0) /* move PSW/save-mask/etc + AP into place */
- movq %r3,-(%r0) /* move routine address + cond handle slot */
- addl3 $4,%r0,%fp /* get start of new callframe */
- insv %r1,$0,$12,6(%fp) /* insert new saved mask */
- popr $0x3f /* restore R0-R5 (cond flags not modified) */
- subl3 $4,%fp,%sp /* sp needs to be equal to fp */
- rsb /* and jmp to the routine */
+ /*
+ * This code checks to see if we got called via a call{s,g} $n,*pcrel32
+ * This is by far the most common case (a call indirectly via the PLT).
+ */
+ subl3 $7,16(%fp),%r1 /* return address */
+ bicb3 $1,(%r1),%r2 /* fetch opcode of instruction */
+ cmpb $0xfa,%r2 /* is it calls/callg */
+ jneq 20f /* no it isn't */
+ cmpb $0xff,2(%r1) /* and deferred 32-bit PC displacement? */
+ jneq 20f /* no it isn't */
+
+ /*
+ * This makes sure the longword with the PLT's address has been updated
+ * to point to the routine's address. If it hasn't, then returning
+ * would put us in an infinite loop. Instead we punt and fake up a
+ * callframe.
+ */
+ movl 3(%r1),%r3 /* get displacement */
+ addl2 16(%fp),%r3 /* add ending location */
+ cmpl (%r3),%r0 /* does it contain the routine address? */
+#ifdef DEBUG
+ jneq 30f /* no it doesn't, die */
+#else
+ jneq 20f /* no it doesn't, go fake a new callframe */
+#endif
+
+11: movl %r1,16(%fp) /* backup to the calls/callg */
+ jbc $29,4(%fp),12f /* skip if this was a callg */
+ clrl (%ap) /* clear argument count */
+12: ret /* return and redo the call */
+
+#if 1
+20:
+ /*
+ * Since the calling standard says only r6-r11 should be saved,
+ * that simplies things for us. That means we can use r0-r5 as
+ * temporaries without worrying about preserving them. This means
+ * can hold the current fixed callframe in r2-r5 as we build the
+ * callframe without having to worry about overwriting the existing
+ * callframe.
+ */
+ extzv $0,$12,(%r0),%r1/* get routine's save mask */
+ bitw $0x3f,%r1 /* does the routine use r0-r5? */
+ jneq 30f /* yes, that sucks */
+ jbc $29,4(%fp),27f /* handle callg */
+ movq 4(%fp),%r2 /* fetch callframe status & saved AP */
+ movq 12(%fp),%r4 /* fetch callframe saved FP & PC */
+ insv %r1,$16,$12,%r2 /* update save mask */
+ movl %ap,%sp /* reset stack to top of callframe */
+22: pushr %r1 /* push registers */
+ movq %r4,-(%sp) /* push callframe saved FP & PC */
+ movq %r2,-(%sp) /* push callframe status & saved AP */
+ pushl $0 /* push condition handler */
+ movl %sp,%fp /* sp == fp now */
+#if 1
+ jmp 2(%r0) /* jump past entry mask */
+#else
+ /*
+ * More correct but IV/DV are never set so ignore doing this for now.
+ */
+ movpsl -(%sp) /* push PSL */
+ clrb (%sp) /* clear user flags */
+ jbc $14,(%r0),24f /* IV need to be set? */
+ bisb2 $0x20,(%sp) /* yes, set it. */
+24: jbc $15,(%r0),25f /* DV need to be set? */
+ bisb2 $0x80,(%sp) /* yes, set it. */
+25: pushab 2(%r0) /* push address of first instruction */
+ rei /* and go to it (updating PSW) */
+#endif
+
+ /*
+ * Count how many registers are being used for callg.
+ */
+27: movl $0x32212110,%r3 /* bit counts */
+ extzv $6,$3,%r1,%r2 /* extract bits 6-8 */
+ ashl $2,%r2,%r2 /* shift by 2 */
+ extzv %r2,$4,%r3,%r4 /* extract count */
+ extzv $9,$3,%r1,%r2 /* extract bits 9-11 */
+ ashl $2,%r2,%r2 /* shift by 2 */
+ extzv %r2,$4,%r3,%r5 /* extract count */
+ movq 4(%fp),%r2 /* fetch callframe status & saved AP */
+ insv %r1,$16,$12,%r2 /* update save mask */
+ addl3 %r3,r4,%r1 /* add counts and discard them */
+ movq 12(%fp),%r4 /* fetch callframe saved FP & PC */
+ moval 20(%fp)[%r1],%sp/* pop callframe */
+ extzv $16,$12,%r2,%r1 /* get save mask back */
+ jbr 22b /* now build the new callframe */
+
+30:
+ calls $0,_C_LABEL(_rtld_die)
+#else
+ /*
+ * Check to see if called via call? $n,w^off(reg)
+ */
+20: addl2 $2,%r1 /* 16-bit displacement */
+ bicb3 $1,(%r1),%r2 /* fetch opcode of instruction */
+ cmpb $0xfa,%r2 /* is it calls/callg */
+ jneq 30f /* no it isn't */
+ bicb3 $0x1f,2(%r1),%r3/* extract addressing mode */
+ cmpb $0xc0,%r3 /* 16-bit displacement? */
+ jeql 11b /* yes, redo the call */
+ halt
+
+ /*
+ * Check to see if called via call? $n,b^off(reg)
+ */
+30: incl %r1 /* 8-bit displacement */
+ bicb3 $1,(%r1),%r2 /* fetch opcode of instruction */
+ cmpb $0xfa,%r2 /* is it calls/callg */
+ jneq 40f /* no it isn't */
+ bicb3 $0x1f,2(%r1),%r3/* extract addressing mode */
+ cmpb $0xa0,%r3 /* 8-bit displacement? */
+ jeql 11b /* yes, redo the call */
+ halt
+
+ /*
+ * Check to see if called via call? $n,(reg)
+ */
+40: incl %r1 /* no displacement */
+ bicb3 $1,(%r1),%r2 /* fetch opcode of instruction */
+ cmpb $0xfa,%r2 /* is it calls/callg */
+ jeql 41f /* yes it is */
+ halt /* no, die die die */
+41: bicb3 $0x0f,2(%r1),%r2/* extract addressing mode */
+ bicb3 $0xf0,2(%r1),%r3/* extract register */
+ extzv $0,$12,6(%fp),%r4/* extract saved mask */
+ cmpb $0x60,%r2 /* register deferred? */
+ jeql 42f /* yes, deal with it */
+ cmpb $0x90,%r2 /* autoincrement deferred? */
+ jeql 70f /* yes, deal with it */
+ halt /* no, die die die */
+
+42: cmpw %r4,$0xffc /* did we save r2-r11? */
+ jneq 50f /* no, deal with it */
+ jbc %r3,%r4,43f /* is the register in the saved mask? */
+
+ /*
+ * We saved r2-r11, so it's easy to replace the saved register with
+ * the right value by indexing into saved register (offset by 8).
+ */
+ movl %r0,(20-8)(%fp)[%r3] /* replace address in saved registers */
+ jbr 11b /* go back and redo call */
+ /*
+ * Must have been called via r0 or r1 which are saved locally.
+ * So move the routine address in the appropriate slot on the stack.
+ */
+43: movl %r0,(%sp)[%r3]
+ jbr 11b /* go back and redo call */
+
+50: jbs %r3,%r4,60f /* is the register in the saved mask? */
+ jbs %r3,$0x3f,43b /* is it r0-r5? */
+ /*
+ * The register used for the call was not saved so we need to move
+ * the new function address into it so the re-call will use the new
+ * address.
+ */
+ pushl %r0 /* save function address on the stack */
+ ashl %r5,$1,%r0 /* create a bitmask for the register */
+ popr %r0 /* pop it off the stack. */
+ jbr 11b /* and redo the call */
+
+60: clrl %r2 /* starting offset into saved registers */
+ clrl %r5 /* start with register 0 */
+
+61: cmpl %r2,%r3 /* is the register to save? */
+ jneq 62f /* no, advance to next */
+ movl %r0,20(%fp)[%r5]/* yes, save return address in saved reg */
+ jbr 11b /* and return the call */
+62: jbc %r5,%r4,63f /* is this register saved? */
+ incl %r5 /* yes, account for it */
+63: incl %r2 /* increment register number */
+ jbr 61b /* and loop */
+
+70: cmpb %r3,$12
+ blss 71f
+ halt
+
+71: cmpw %r4,$0xffc /* did we save r2-r11? */
+ jneq 72f /* no, deal with it */
+ subl2 $4,(20-8)(%fp)[%r3] /* backup incremented register */
+ jbr 11b /* and redo the call.
+
+72: jbs %r3,%r4,80f
+ jbs %r3,%3f,74f
+ ashl %r5,$1,%r0 /* create a bitmask for the register */
+ pushr %r0 /* pop it off the stack. */
+ subl2 $4,(%sp) /* backup incremented register */
+ popr %r0 /* pop it off the stack. */
+ jbr 11b /* and redo the call.
+
+73: subl2 %4,(%sp)[%r3] /* backup incremented register */
+ jbr 11b /* and redo the call.
+
+80: clrl %r2 /* starting offset into saved registers */
+ clrl %r5 /* start with register 0 */
+
+81: cmpl %r2,%r3 /* is the register to save? */
+ jneq 82f /* no, advance to next */
+ subl $4,20(%fp)[%r5] /* yes, backup incremented register */
+ jbr 11b /* and return the call */
+82: jbc %r5,%r4,83f /* is this register saved? */
+ incl %r5 /* yes, account for it */
+83: incl %r2 /* increment register number */
+ jbr 81b /* and loop */
+#endif
END(_rtld_bind_start)
