This is the mail archive of the
binutils@sourceware.org
mailing list for the binutils project.
SPU overlay manager fix and new ld param
- From: Alan Modra <amodra at bigpond dot net dot au>
- To: binutils at sourceware dot org
- Date: Tue, 17 Jun 2008 01:53:45 +0930
- Subject: SPU overlay manager fix and new ld param
This patch
a) replaces the fixed 2000 byte red zone stack allowance for
--auto-overlay with a linker command line parameter,
b) makes the overlay manager interrupt safe,
c) and does not attempt to put any functions called by the spu
interrupt handler into non-overlay local store, since we can't do
that reliably in the presence of calls via function pointer.
bfd/
* elf32-spu.c (struct spu_link_hash_table): Add extra_stack_space.
(spu_elf_check_vma): Add extra_stack_space param, copy to htab.
(spu_elf_auto_overlay): Use it.
(RECURSE_UNMARK): Define as 0.
(unmark_overlay_section): Heed RECURSE_UNMARK.
* elf32-spu.h (spu_elf_check_vma): Update prototype.
ld/
* emultempl/spuelf.em (extra_stack_space): New variable.
(gld${EMULATION_NAME}_finish): Pass it to spu_elf_check_vma.
(PARSE_AND_LIST_LONGOPTS, PARSE_AND_LIST_OPTIONS,
PARSE_AND_LIST_ARGS_CASES): Handle --extra-stack-space.
* emultempl/spu_ovl.S: Mask interrupts during dma and update of
overlay manager structures.
* emultempl/spu_ovl.o: Regenerate.
diff -urp src.old/bfd/elf32-spu.c src/bfd/elf32-spu.c
--- src.old/bfd/elf32-spu.c 2008-06-07 21:50:20.000000000 +0930
+++ src/bfd/elf32-spu.c 2008-06-12 15:41:02.000000000 +0930
@@ -325,6 +325,11 @@ struct spu_link_hash_table
unsigned int overlay_fixed;
/* Local store --auto-overlay should reserve for stack and heap. */
unsigned int reserved;
+ /* If reserved is not specified, stack analysis will calculate a value
+ for the stack. This parameter adjusts that value to allow for
+ negative sp access (the ABI says 2000 bytes below sp are valid,
+ and the overlay manager uses some of this area). */
+ int extra_stack_space;
/* Count of overlay stubs needed in non-overlay area. */
unsigned int non_ovly_stub;
@@ -1548,6 +1553,7 @@ spu_elf_check_vma (struct bfd_link_info
unsigned int hi,
unsigned int overlay_fixed,
unsigned int reserved,
+ int extra_stack_space,
void (*spu_elf_load_ovl_mgr) (void),
FILE *(*spu_elf_open_overlay_script) (void),
void (*spu_elf_relink) (void))
@@ -1562,6 +1568,7 @@ spu_elf_check_vma (struct bfd_link_info
htab->local_store = hi + 1 - lo;
htab->overlay_fixed = overlay_fixed;
htab->reserved = reserved;
+ htab->extra_stack_space = extra_stack_space;
htab->spu_elf_load_ovl_mgr = spu_elf_load_ovl_mgr;
htab->spu_elf_open_overlay_script = spu_elf_open_overlay_script;
htab->spu_elf_relink = spu_elf_relink;
@@ -2923,6 +2930,11 @@ mark_overlay_section (struct function_in
return TRUE;
}
+/* If non-zero then unmark functions called from those within sections
+ that we need to unmark. Unfortunately this isn't reliable since the
+ call graph cannot know the destination of function pointer calls. */
+#define RECURSE_UNMARK 0
+
struct _uos_param {
asection *exclude_input_section;
asection *exclude_output_section;
@@ -2950,9 +2962,10 @@ unmark_overlay_section (struct function_
|| fun->sec->output_section == uos_param->exclude_output_section)
excluded = 1;
- uos_param->clearing += excluded;
+ if (RECURSE_UNMARK)
+ uos_param->clearing += excluded;
- if (uos_param->clearing)
+ if (RECURSE_UNMARK ? uos_param->clearing : excluded)
{
fun->sec->linker_mark = 0;
if (fun->rodata)
@@ -2963,7 +2976,8 @@ unmark_overlay_section (struct function_
if (!unmark_overlay_section (call->fun, info, param))
return FALSE;
- uos_param->clearing -= excluded;
+ if (RECURSE_UNMARK)
+ uos_param->clearing -= excluded;
return TRUE;
}
@@ -3574,7 +3588,7 @@ spu_elf_auto_overlay (struct bfd_link_in
sum_stack_param.overall_stack = 0;
if (!for_each_node (sum_stack, info, &sum_stack_param, TRUE))
goto err_exit;
- htab->reserved = sum_stack_param.overall_stack + 2000;
+ htab->reserved = sum_stack_param.overall_stack + htab->extra_stack_space;
}
fixed_size += htab->reserved;
fixed_size += htab->non_ovly_stub * OVL_STUB_SIZE;
diff -urp src.old/bfd/elf32-spu.h src/bfd/elf32-spu.h
--- src.old/bfd/elf32-spu.h 2008-03-28 17:24:57.000000000 +1030
+++ src/bfd/elf32-spu.h 2008-06-10 17:18:02.000000000 +0930
@@ -62,5 +62,5 @@ extern int spu_elf_size_stubs (struct bf
extern bfd_boolean spu_elf_build_stubs (struct bfd_link_info *, int);
extern asection *spu_elf_check_vma (struct bfd_link_info *, int,
unsigned int, unsigned int, unsigned int,
- unsigned int, void (*) (void),
+ unsigned int, int, void (*) (void),
FILE *(*) (void), void (*) (void));
diff -urp src.old/ld/emultempl/spuelf.em src/ld/emultempl/spuelf.em
--- src.old/ld/emultempl/spuelf.em 2008-06-07 11:27:23.000000000 +0930
+++ src/ld/emultempl/spuelf.em 2008-06-12 15:22:57.000000000 +0930
@@ -50,6 +50,7 @@ static int auto_overlay = 0;
static char *auto_overlay_file = 0;
static unsigned int auto_overlay_fixed = 0;
static unsigned int auto_overlay_reserved = 0;
+static int extra_stack_space = 2000;
int my_argc;
char **my_argv;
@@ -330,6 +331,7 @@ gld${EMULATION_NAME}_finish (void)
s = spu_elf_check_vma (&link_info, auto_overlay,
local_store_lo, local_store_hi,
auto_overlay_fixed, auto_overlay_reserved,
+ extra_stack_space,
spu_elf_load_ovl_mgr,
spu_elf_open_overlay_script,
spu_elf_relink);
@@ -523,7 +525,8 @@ PARSE_AND_LIST_PROLOGUE='
#define OPTION_SPU_OVERLAY_RODATA (OPTION_SPU_AUTO_RELINK + 1)
#define OPTION_SPU_FIXED_SPACE (OPTION_SPU_OVERLAY_RODATA + 1)
#define OPTION_SPU_RESERVED_SPACE (OPTION_SPU_FIXED_SPACE + 1)
-#define OPTION_SPU_NO_AUTO_OVERLAY (OPTION_SPU_RESERVED_SPACE + 1)
+#define OPTION_SPU_EXTRA_STACK (OPTION_SPU_RESERVED_SPACE + 1)
+#define OPTION_SPU_NO_AUTO_OVERLAY (OPTION_SPU_EXTRA_STACK + 1)
'
PARSE_AND_LIST_LONGOPTS='
@@ -539,6 +542,7 @@ PARSE_AND_LIST_LONGOPTS='
{ "overlay-rodata", no_argument, NULL, OPTION_SPU_OVERLAY_RODATA },
{ "fixed-space", required_argument, NULL, OPTION_SPU_FIXED_SPACE },
{ "reserved-space", required_argument, NULL, OPTION_SPU_RESERVED_SPACE },
+ { "extra-stack-space", required_argument, NULL, OPTION_SPU_EXTRA_STACK },
{ "no-auto-overlay", optional_argument, NULL, OPTION_SPU_NO_AUTO_OVERLAY },
'
@@ -557,7 +561,10 @@ PARSE_AND_LIST_OPTIONS='
--overlay-rodata Place read-only data with associated function\n\
code in overlays.\n\
--fixed-space=bytes Local store for non-overlay code and data.\n\
- --reserved-space=bytes Local store for stack and heap.\n"
+ --reserved-space=bytes Local store for stack and heap. If not specified\n\
+ ld will estimate stack size and assume no heap.\n\
+ --extra-stack-space=bytes Space for negative sp access (default 2000) if\n\
+ --reserved-space not given.\n"
));
'
@@ -635,6 +642,15 @@ PARSE_AND_LIST_ARGS_CASES='
}
break;
+ case OPTION_SPU_EXTRA_STACK:
+ {
+ char *end;
+ extra_stack_space = strtol (optarg, &end, 0);
+ if (*end != 0)
+ einfo (_("%P%F: invalid --extra-stack-space value `%s'\''\n"), optarg);
+ }
+ break;
+
case OPTION_SPU_NO_AUTO_OVERLAY:
auto_overlay = 0;
if (optarg != NULL)
diff -urp src.old/ld/emultempl/spu_ovl.S src/ld/emultempl/spu_ovl.S
--- src.old/ld/emultempl/spu_ovl.S 2008-02-07 11:56:23.000000000 +1030
+++ src/ld/emultempl/spu_ovl.S 2008-06-12 15:42:54.000000000 +0930
@@ -67,11 +67,13 @@
#define rv5 reserved4
#define cgshuf reserved4
#define newovl reserved4
+#define irqtmp1 reserved4
+#define irqtmp2 reserved4
#define reserved5 $79
#define target reserved5
-#define save1 $72
+#define save1 $74
#define rv4 save1
#define rv7 save1
#define tagid save1
@@ -87,7 +89,7 @@
#define oldovl save2
#define newvma save2
-#define save3 $74
+#define save3 $72
#define rv1 save3
#define ea64 save3
#define buf3 save3
@@ -95,6 +97,8 @@
#define newmap save3
#define oldmask save3
+#define save4 $71
+#define irq_stat save4
.text
.align 4
@@ -140,7 +144,12 @@ __ovly_return:
#nop; lnop
#nop
lqx vma, tab1, off1 # 1,6 8
+#ifdef OVLY_IRQ_SAVE
+ nop
+ stqd save4, -64($sp) # 1,6 9
+#else
#nop; lnop
+#endif
#nop; lnop
#nop; lnop
#nop; lnop
@@ -168,7 +177,9 @@ ovly_ret9:
* On entry $75 points to a word consisting of the overlay index in
* the top 14 bits, and the target address in the bottom 18 bits.
*
- * Sets up $lr to return via __ovly_return.
+ * Sets up $lr to return via __ovly_return. If $lr is already set
+ * to return via __ovly_return, don't change it. In that case we
+ * have a tail call from one overlay function to another.
* Updates __ovly_current.
*/
.align 3
@@ -245,7 +256,11 @@ __ovly_load:
or rv7, rv4, rv6 # 0,2 16
lqd save2, -32($sp) # 1,6 16
andi present2, size2, 1 # 0,2 17
+#ifdef OVLY_IRQ_SAVE
+ stqd save4, -64($sp) # 1,6 17
+#else
lnop # 1,0 17
+#endif
selb $lr, rv7, $lr, rv5 # 0,2 18
lqd save1, -16($sp) # 1,6 18
#nop
@@ -268,9 +283,19 @@ ovly_load9:
.type __ovly_load_event, @function
__ovly_load_event:
do_load:
+#ifdef OVLY_IRQ_SAVE
+ ila irqtmp1, do_load10 # 0,2 -5
+ rotqbyi sz, vma, 8 # 1,4 -5
#nop
- rotqbyi sz, vma, 8 # 1,4 0
+ rdch irq_stat, $SPU_RdMachStat # 1,6 -4
#nop
+ bid irqtmp1 # 1,4 -3
+do_load10:
+ nop
+#else
+#nop
+ rotqbyi sz, vma, 8 # 1,4 0
+#endif
rotqbyi osize, vma, 4 # 1,4 1
#nop
lqa ea64, _EAR_ # 1,6 2
@@ -410,14 +435,30 @@ __ovly_xfer_loop:
#nop; lnop
andc pbit, pbit, zovl # 0,2 74
lqd save2, -32($sp) # 1,6 74
+#ifdef OVLY_IRQ_SAVE
+ ila irqtmp2, do_load90 # 0,2 75
+#lnop
+ andi irq_stat, irq_stat, 1 # 0,2 76
+#lnop
+#else
#nop; lnop
#nop; lnop
+#endif
andc oldvma, oldvma, pbit # 0,2 77
lqd save1, -16($sp) # 1,6 77
-#nop; lnop
- nop
+ nop # 0,0 78
+#lnop
+#nop
stqx oldvma, tab5, off5 # 1,6 79
+#nop
+#ifdef OVLY_IRQ_SAVE
+ binze irq_stat, irqtmp2 # 1,4 80
+do_load90:
+#nop
+ lqd save4, -64($sp) # 1,6 84
+#else
#nop; lnop
+#endif
.global _ovly_debug_event
.type _ovly_debug_event, @function
@@ -425,6 +466,6 @@ _ovly_debug_event:
nop
/* Branch to target address. */
do_load99:
- bi target # 1,4 81
+ bi target # 1,4 81/85
.size __ovly_load, . - __ovly_load
--
Alan Modra
Australia Development Lab, IBM