diff mbox

[PPC/Darwin] some tidy-ups for save_world (and a prelude to splitting it out of the rs6000 code).

Message ID 2A206914-6A78-4F47-A7D0-F06BCC7CFD4A@sandoe-acoustics.co.uk
State New
Headers show

Commit Message

Iain Sandoe Dec. 4, 2011, 8:23 p.m. UTC
On 4 Dec 2011, at 06:35, Mike Stump wrote:

> On Nov 30, 2011, at 6:28 AM, Iain Sandoe wrote:
>> While trying to track down the vector unwind problems on ppc- 
>> darwin, I made some tidy-ups for "save_world()".
>> In the end, that was not where the main problem, lay - but I did  
>> find a few things wrong there on the way - they should be fixed,  
>> even if there's no specific bug filed at present.
>>
>> I'm also attaching a second patch which is purely cosmetic white- 
>> space/comment tidies I'd also like to apply.
>>
>> checked on powerpc-darwin{8-G4,9-G5} and crosses to powerpc-eabisim  
>> and powerpc-ibm-aix6.1.3.0
>
>> -  /* On Darwin, the unwind routines are compiled without
>> -     TARGET_ALTIVEC, and use save_world to save/restore the
>> -     altivec registers when necessary.  */
>> +  /* When generating code for earlier versions of Darwin, which  
>> might run on
>> +     hardware with or without Altivec, we use out-of-line save/ 
>> restores in
>> +     function prologues/epilogues that require it.  These routines  
>> determine
>> +     whether to save/restore Altivec at runtime.  */
>
> So, we need to first settle on how the library is compiled before  
> this becomes true...

The patch does not change any functionality here.
This was a ( clearly failed ;-) ) attempt to make the comment more  
useful to newcomers to the code ...
(withdrawn, can think again sometime later).

>> +save_world:
>> +	trap
>> +
>> +	.private_extern eh_rest_world_r10
>> +eh_rest_world_r10:
>> +	trap
>
> So, I can't see the necessity of doing this.  I think it is bad  
> style to generate code that will die at runtime.  I'd rather have it  
> not link.

done - there would now be a link fail if it were specified.

(although I would point out, FTR, that the code would have always  
crashed if it had ever been invoked - since it was only m32 capable).

[[FWIW, I made a working m64 version, which we could resurrect should  
there be some reason we decided that out-of-line world saves were  
useful everywhere]]

OK for trunk?
Iain

Comments

David Edelsohn Dec. 10, 2011, 11:57 p.m. UTC | #1
On Sun, Dec 4, 2011 at 3:23 PM, Iain Sandoe
<developer@sandoe-acoustics.co.uk> wrote:

gcc:

       * config/rs6000/rs6000.c (first_altivec_reg_to_save): Amend comment.
       (compute_vrsave_mask): Likewise.
       (rs6000_emit_prologue): Move update of VRSave mask to save_world()
       when that is in use.

libgcc:

       * config/rs6000/darwin-world.S (toplevel): Make it clear that
this function is
       not used for PPC64.  (save_world): Amend comments.  Update the VRsave
       mask to reflect the saved regs.
       (rest_world): Update comments, do not  clobber r10, do not use r8.
       (eh_rest_world_r10): Amend comments, do not use r8.
       (rest_world_eh_r7r8): Make local, move restore of CR and target address
       to end of routine, do not use r8.

> OK for trunk?

Okay.

It probably would be helpful to add a comment that rest_world_eh_r7r8
has "r8" in the name for backward compatibility.

Thanks, David
diff mbox

Patch

Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c	(revision 181990)
+++ gcc/config/rs6000/rs6000.c	(working copy)
@@ -19899,7 +19899,9 @@  rs6000_emit_prologue (void)
      used in this function, and do the corresponding magic in the
      epilogue.  */
 
-  if (TARGET_ALTIVEC && TARGET_ALTIVEC_VRSAVE
+  if (!WORLD_SAVE_P (info)
+      && TARGET_ALTIVEC
+      && TARGET_ALTIVEC_VRSAVE
       && info->vrsave_mask != 0)
     {
       rtx reg, mem, vrsave;
@@ -19915,15 +19917,12 @@  rs6000_emit_prologue (void)
       else
         emit_insn (gen_rtx_SET (VOIDmode, reg, vrsave));
 
-      if (!WORLD_SAVE_P (info))
-        {
-          /* Save VRSAVE.  */
-          offset = info->vrsave_save_offset + sp_offset;
-          mem = gen_frame_mem (SImode,
-                               gen_rtx_PLUS (Pmode, frame_reg_rtx,
-                                             GEN_INT (offset)));
-          insn = emit_move_insn (mem, reg);
-        }
+      /* Save VRSAVE.  */
+      offset = info->vrsave_save_offset + sp_offset;
+      mem = gen_frame_mem (SImode,
+			   gen_rtx_PLUS (Pmode, frame_reg_rtx, 
+					 GEN_INT (offset)));
+      insn = emit_move_insn (mem, reg);
 
       /* Include the registers in the mask.  */
       emit_insn (gen_iorsi3 (reg, reg, GEN_INT ((int) info->vrsave_mask)));
Index: libgcc/config/rs6000/darwin-world.S
===================================================================
--- libgcc/config/rs6000/darwin-world.S	(revision 181990)
+++ libgcc/config/rs6000/darwin-world.S	(working copy)
@@ -24,6 +24,8 @@ 
  * <http://www.gnu.org/licenses/>.
  */ 
 
+#ifndef __ppc64__
+
 	.machine ppc7400
 .data
 	.align 2
@@ -33,12 +35,7 @@ 
 .non_lazy_symbol_pointer
 L_has_vec$non_lazy_ptr:
 	.indirect_symbol __cpu_has_altivec
-#ifdef __ppc64__
-	.quad	0
-#else
 	.long	0
-#endif
-
 #else
 
 /* For static, "pretend" we have a non-lazy-pointer.  */
@@ -57,12 +54,11 @@  L_has_vec$non_lazy_ptr:
    provided by the System Framework to determine this.)
 
    SAVE_WORLD takes R0 (the caller`s caller`s return address) and R11
-   (the stack frame size) as parameters.  It returns VRsave in R0 if
-   we`re on a CPU with vector regs.
+   (the stack frame size) as parameters.  It returns the updated VRsave
+   in R0 if we`re on a CPU with vector regs.
 
-   With gcc3, we now need to save and restore CR as well, since gcc3's
-   scheduled prologs can cause comparisons to be moved before calls to
-   save_world!
+   For gcc3 onward, we need to save and restore CR as well, since scheduled
+   prologs can cause comparisons to be moved before calls to save_world.
 
    USES: R0 R11 R12  */
 
@@ -143,69 +139,62 @@  L$saveVMX:
 	stvx v30,r11,r12
 	mfspr r0,VRsave
 	li r11,-16
-	stvx v31,r11,r12
-				/* VRsave lives at -224(R1)  */
-	stw r0,0(r12)
+	stvx v31,r11,r12	
+	stw r0,0(r12)		/* VRsave lives at -224(R1).  */
+	ori r0,r0,0xfff		/* We just saved these.  */
+	mtspr VRsave,r0
 	blr
 
+/* rest_world  is jumped to, not called, so no need to worry about LR.
+   clobbers R0, R7, R11 and R12.  This just undoes the work done above.  */
 
+	.private_extern rest_world
+rest_world:
+				
+	lwz r11, 0(r1)		/* Pickup previous SP  */
+	li r7, 0		/* Stack offset is zero, r10 is ignored.  */
+	b Lrest_world_eh_r7r8
+
 /* eh_rest_world_r10 is jumped to, not called, so no need to worry about LR.
    R10 is the C++ EH stack adjust parameter, we return to the caller`s caller.
 
-   USES: R0 R10 R11 R12   and R7 R8
-   RETURNS: C++ EH Data registers (R3 - R6.)
+   clobbers: R0, R7, R11 and R12
+   uses    : R10
+   RETURNS : C++ EH Data registers (R3 - R6).  */
 
-   We now set up R7/R8 and jump to rest_world_eh_r7r8.
-
-   rest_world doesn't use the R10 stack adjust parameter, nor does it
-   pick up the R3-R6 exception handling stuff.  */
-
-.private_extern rest_world
-rest_world:
-				/* Pickup previous SP  */
-	lwz r11, 0(r1)
-	li r7, 0
-	lwz r8, 8(r11)
-	li r10, 0
-	b rest_world_eh_r7r8
-
-.private_extern eh_rest_world_r10
+	.private_extern eh_rest_world_r10
 eh_rest_world_r10:
-				/* Pickup previous SP  */
-	lwz r11, 0(r1)
-	mr  r7,r10
-	lwz r8, 8(r11)
-			/* pickup the C++ EH data regs (R3 - R6.)  */
+				
+	lwz r11, 0(r1)		/* Pickup previous SP  */
+	mr  r7,r10		/* Stack offset.  */
+	
+	/* pickup the C++ EH data regs (R3 - R6.)  */
 	lwz r6,-420(r11)
 	lwz r5,-424(r11)
 	lwz r4,-428(r11)
 	lwz r3,-432(r11)
 
-	b rest_world_eh_r7r8
+	/* Fall through to Lrest_world_eh_r7r8.  */
 
-/* rest_world_eh_r7r8 is jumped to -- not called! -- when we're doing
-   the exception-handling epilog.  R7 contains the offset to add to
-   the SP, and R8 contains the 'real' return address.
+/* When we are doing the exception-handling epilog, R7 contains the offset to
+   add to the SP.
 
-   USES: R0 R11 R12  [R7/R8]
-   RETURNS: C++ EH Data registers (R3 - R6.)  */
+   clobbers: R0, R11 and R12
+   uses    : R7.  */
 
-rest_world_eh_r7r8:
+Lrest_world_eh_r7r8:
+	/* See if we have Altivec.  */
 	bcl 20,31,Lr7r8$pb
 Lr7r8$pb: mflr r12
-	lwz r11,0(r1)
-				/* R11 := previous SP  */
+
 	addis r12,r12,ha16(L_has_vec$non_lazy_ptr-Lr7r8$pb)
 	lwz r12,lo16(L_has_vec$non_lazy_ptr-Lr7r8$pb)(r12)
-	lwz r0,4(r11)
-				/* R0 := old CR  */
-	lwz r12,0(r12)
-				/* R12 := HAS_VEC  */
-	mtcr r0	
+	lwz r12,0(r12)		/* R12 := HAS_VEC  */
 	cmpwi r12,0
 	lmw r13,-220(r11)
 	beq L.rest_world_fp_eh
-				/* restore VRsave and V20..V31  */
+
+	/* We have Altivec, restore VRsave and V20..V31  */
 	lwz r0,-224(r11)
 	li r12,-416
 	mtspr VRsave,r0
@@ -234,6 +223,7 @@  Lr7r8$pb: mflr r12
 	lvx v31,r11,r12
 
 L.rest_world_fp_eh:
+	lwz r0,4(r11)		/* recover saved CR  */
 	lfd f14,-144(r11)
 	lfd f15,-136(r11)
 	lfd f16,-128(r11)
@@ -251,9 +241,12 @@  L.rest_world_fp_eh:
 	lfd f28,-32(r11)
 	lfd f29,-24(r11)
 	lfd f30,-16(r11)
-			/* R8 is the exception-handler's address  */
-	mtctr r8
-	lfd f31,-8(r11)
-			/* set SP to original value + R7 offset  */
-	add r1,r11,r7
+	mtcr r0			/* restore the saved cr.  */
+	lwz r0, 8(r11)		/* Pick up the 'real' return address.  */
+	lfd f31,-8(r11)		
+	mtctr r0		/* exception-handler ret. address  */
+	add r1,r11,r7		/* set SP to original value + R7 offset  */
 	bctr
+#endif
+/* we should never be called on ppc64 for this ... */
+/* Done.  */