* Handle multi-word regsiters in REG_CFA_RESTORE notes
@ 2011-09-15 2:42 Bernd Schmidt
2011-09-15 20:41 ` Richard Henderson
0 siblings, 1 reply; 7+ messages in thread
From: Bernd Schmidt @ 2011-09-15 2:42 UTC (permalink / raw)
To: GCC Patches
[-- Attachment #1: Type: text/plain, Size: 407 bytes --]
While testing the altest iteration of shrink-wrapping on mips-elf, a new
failure showed up in gcc.dg/pr43139.c. When restoring floating-point
registers, we attach REG_CFA_RESTORE notes for DFmode registers, but the
dwarf2cfi code only records a single regno for such a multiword hard reg.
Fixed with this patch, tested with shrink-wrapping on mips-elf (default
multilib plus two with -mips16). Ok?
Bernd
[-- Attachment #2: cfa-multiword.diff --]
[-- Type: text/plain, Size: 865 bytes --]
* dwarf2cfi.c (dwarf2out_frame_debug_cfa_expression): Handle
multi-word registers.
Index: gcc/dwarf2cfi.c
===================================================================
--- gcc/dwarf2cfi.c (revision 178734)
+++ gcc/dwarf2cfi.c (working copy)
@@ -1236,10 +1236,15 @@ dwarf2out_frame_debug_cfa_expression (rt
static void
dwarf2out_frame_debug_cfa_restore (rtx reg)
{
- unsigned int regno = dwf_regno (reg);
+ unsigned int orig_regno = REGNO (reg);
+ int nregs = hard_regno_nregs[orig_regno][GET_MODE (reg)];
+ while (nregs-- > 0)
+ {
+ unsigned int regno = DWARF_FRAME_REGNUM (orig_regno + nregs);
- add_cfi_restore (regno);
- update_row_reg_save (cur_row, regno, NULL);
+ add_cfi_restore (regno);
+ update_row_reg_save (cur_row, regno, NULL);
+ }
}
/* A subroutine of dwarf2out_frame_debug, process a REG_CFA_WINDOW_SAVE.
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Handle multi-word regsiters in REG_CFA_RESTORE notes
2011-09-15 2:42 Handle multi-word regsiters in REG_CFA_RESTORE notes Bernd Schmidt
@ 2011-09-15 20:41 ` Richard Henderson
2011-09-21 15:14 ` Bernd Schmidt
0 siblings, 1 reply; 7+ messages in thread
From: Richard Henderson @ 2011-09-15 20:41 UTC (permalink / raw)
To: Bernd Schmidt; +Cc: GCC Patches
On 09/14/2011 06:12 PM, Bernd Schmidt wrote:
> + unsigned int orig_regno = REGNO (reg);
> + int nregs = hard_regno_nregs[orig_regno][GET_MODE (reg)];
> + while (nregs-- > 0)
The rest of the file seems to use targetm.dwarf_register_span.
This probably ought to do the same.
r~
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Handle multi-word regsiters in REG_CFA_RESTORE notes
2011-09-15 20:41 ` Richard Henderson
@ 2011-09-21 15:14 ` Bernd Schmidt
2011-09-21 17:47 ` Richard Henderson
0 siblings, 1 reply; 7+ messages in thread
From: Bernd Schmidt @ 2011-09-21 15:14 UTC (permalink / raw)
To: Richard Henderson; +Cc: GCC Patches
On 09/15/11 21:42, Richard Henderson wrote:
> On 09/14/2011 06:12 PM, Bernd Schmidt wrote:
>> + unsigned int orig_regno = REGNO (reg);
>> + int nregs = hard_regno_nregs[orig_regno][GET_MODE (reg)];
>> + while (nregs-- > 0)
>
> The rest of the file seems to use targetm.dwarf_register_span.
> This probably ought to do the same.
I guess I am confused what dwarf_register_span is supposed to accomplish
that a loop over nregs couldn't. For little endian MIPS, it returns
NULL, so I don't think it would solve the problem I'm trying to fix.
Bernd
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Handle multi-word regsiters in REG_CFA_RESTORE notes
2011-09-21 15:14 ` Bernd Schmidt
@ 2011-09-21 17:47 ` Richard Henderson
2011-09-22 22:09 ` Bernd Schmidt
0 siblings, 1 reply; 7+ messages in thread
From: Richard Henderson @ 2011-09-21 17:47 UTC (permalink / raw)
To: Bernd Schmidt; +Cc: GCC Patches
On 09/21/2011 07:55 AM, Bernd Schmidt wrote:
> On 09/15/11 21:42, Richard Henderson wrote:
>> On 09/14/2011 06:12 PM, Bernd Schmidt wrote:
>>> + unsigned int orig_regno = REGNO (reg);
>>> + int nregs = hard_regno_nregs[orig_regno][GET_MODE (reg)];
>>> + while (nregs-- > 0)
>>
>> The rest of the file seems to use targetm.dwarf_register_span.
>> This probably ought to do the same.
>
> I guess I am confused what dwarf_register_span is supposed to accomplish
> that a loop over nregs couldn't. For little endian MIPS, it returns
> NULL, so I don't think it would solve the problem I'm trying to fix.
Why, then, is this the only place in dwarf2cfi that needs to handle
registers via a loop over nregs? It seems to me that we should either
be handling multi-register spans everywhere or nowhere.
Because alternately, this could be a bug in your backend that you
failed to add two RESTORE notes instead of just one...
r~
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Handle multi-word regsiters in REG_CFA_RESTORE notes
2011-09-21 17:47 ` Richard Henderson
@ 2011-09-22 22:09 ` Bernd Schmidt
2011-09-25 19:54 ` Richard Sandiford
0 siblings, 1 reply; 7+ messages in thread
From: Bernd Schmidt @ 2011-09-22 22:09 UTC (permalink / raw)
To: Richard Henderson; +Cc: GCC Patches, Richard Sandiford
[-- Attachment #1: Type: text/plain, Size: 433 bytes --]
On 09/21/11 19:33, Richard Henderson wrote:
> Why, then, is this the only place in dwarf2cfi that needs to handle
> registers via a loop over nregs? It seems to me that we should either
> be handling multi-register spans everywhere or nowhere.
>
> Because alternately, this could be a bug in your backend that you
> failed to add two RESTORE notes instead of just one...
Well, changing the backend works too. Patch below.
Bernd
[-- Attachment #2: mips-ep64.diff --]
[-- Type: text/plain, Size: 1855 bytes --]
* mips.c (mips_restore_reg): Split multiword registers for
REG_CFA_RESTORE notes.
Index: gcc/config/mips/mips.c
===================================================================
--- gcc/config/mips/mips.c (revision 178847)
+++ gcc/config/mips/mips.c (working copy)
@@ -10286,16 +10286,28 @@ mips_epilogue_set_cfa (rtx reg, HOST_WID
static void
mips_restore_reg (rtx reg, rtx mem)
{
+ enum machine_mode mode = GET_MODE (reg);
+ unsigned regno = REGNO (reg);
+
/* There's no MIPS16 instruction to load $31 directly. Load into
$7 instead and adjust the return insn appropriately. */
- if (TARGET_MIPS16 && REGNO (reg) == RETURN_ADDR_REGNUM)
- reg = gen_rtx_REG (GET_MODE (reg), GP_REG_FIRST + 7);
+ if (TARGET_MIPS16 && regno == RETURN_ADDR_REGNUM)
+ reg = gen_rtx_REG (mode, GP_REG_FIRST + 7);
+ else if (GET_MODE_SIZE (mode) != 8 || !mips_split_64bit_move_p (reg, mem))
+ mips_epilogue.cfa_restores
+ = alloc_reg_note (REG_CFA_RESTORE, reg, mips_epilogue.cfa_restores);
else
- mips_epilogue.cfa_restores = alloc_reg_note (REG_CFA_RESTORE, reg,
- mips_epilogue.cfa_restores);
+ {
+ rtx word1 = mips_subword (reg, true);
+ rtx word2 = mips_subword (reg, false);
+ mips_epilogue.cfa_restores
+ = alloc_reg_note (REG_CFA_RESTORE, word1, mips_epilogue.cfa_restores);
+ mips_epilogue.cfa_restores
+ = alloc_reg_note (REG_CFA_RESTORE, word2, mips_epilogue.cfa_restores);
+ }
- mips_emit_save_slot_move (reg, mem, MIPS_EPILOGUE_TEMP (GET_MODE (reg)));
- if (REGNO (reg) == REGNO (mips_epilogue.cfa_reg))
+ mips_emit_save_slot_move (reg, mem, MIPS_EPILOGUE_TEMP (mode));
+ if (regno == REGNO (mips_epilogue.cfa_reg))
/* The CFA is currently defined in terms of the register whose
value we have just restored. Redefine the CFA in terms of
the stack pointer. */
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Handle multi-word regsiters in REG_CFA_RESTORE notes
2011-09-22 22:09 ` Bernd Schmidt
@ 2011-09-25 19:54 ` Richard Sandiford
2011-09-27 12:57 ` Bernd Schmidt
0 siblings, 1 reply; 7+ messages in thread
From: Richard Sandiford @ 2011-09-25 19:54 UTC (permalink / raw)
To: Bernd Schmidt; +Cc: Richard Henderson, GCC Patches
Bernd Schmidt <bernds@codesourcery.com> writes:
> On 09/21/11 19:33, Richard Henderson wrote:
>> Why, then, is this the only place in dwarf2cfi that needs to handle
>> registers via a loop over nregs? It seems to me that we should either
>> be handling multi-register spans everywhere or nowhere.
>>
>> Because alternately, this could be a bug in your backend that you
>> failed to add two RESTORE notes instead of just one...
>
> Well, changing the backend works too. Patch below.
>
>
> Bernd
>
> * mips.c (mips_restore_reg): Split multiword registers for
> REG_CFA_RESTORE notes.
>
> Index: gcc/config/mips/mips.c
> ===================================================================
> --- gcc/config/mips/mips.c (revision 178847)
> +++ gcc/config/mips/mips.c (working copy)
> @@ -10286,16 +10286,28 @@ mips_epilogue_set_cfa (rtx reg, HOST_WID
> static void
> mips_restore_reg (rtx reg, rtx mem)
> {
> + enum machine_mode mode = GET_MODE (reg);
> + unsigned regno = REGNO (reg);
> +
> /* There's no MIPS16 instruction to load $31 directly. Load into
> $7 instead and adjust the return insn appropriately. */
> - if (TARGET_MIPS16 && REGNO (reg) == RETURN_ADDR_REGNUM)
> - reg = gen_rtx_REG (GET_MODE (reg), GP_REG_FIRST + 7);
> + if (TARGET_MIPS16 && regno == RETURN_ADDR_REGNUM)
> + reg = gen_rtx_REG (mode, GP_REG_FIRST + 7);
> + else if (GET_MODE_SIZE (mode) != 8 || !mips_split_64bit_move_p (reg, mem))
> + mips_epilogue.cfa_restores
> + = alloc_reg_note (REG_CFA_RESTORE, reg, mips_epilogue.cfa_restores);
> else
> - mips_epilogue.cfa_restores = alloc_reg_note (REG_CFA_RESTORE, reg,
> - mips_epilogue.cfa_restores);
> + {
> + rtx word1 = mips_subword (reg, true);
> + rtx word2 = mips_subword (reg, false);
> + mips_epilogue.cfa_restores
> + = alloc_reg_note (REG_CFA_RESTORE, word1, mips_epilogue.cfa_restores);
> + mips_epilogue.cfa_restores
> + = alloc_reg_note (REG_CFA_RESTORE, word2, mips_epilogue.cfa_restores);
> + }
I think the condition ought to match that in mips_save_reg:
static void
mips_save_reg (rtx reg, rtx mem)
{
if (GET_MODE (reg) == DFmode && !TARGET_FLOAT64)
{
rtx x1, x2;
if (mips_split_64bit_move_p (mem, reg))
mips_split_doubleword_move (mem, reg);
else
mips_emit_move (mem, reg);
x1 = mips_frame_set (mips_subword (mem, false),
mips_subword (reg, false));
x2 = mips_frame_set (mips_subword (mem, true),
mips_subword (reg, true));
mips_set_frame_expr (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, x1, x2)));
}
else
mips_emit_save_slot_move (mem, reg, MIPS_PROLOGUE_TEMP (GET_MODE (reg)));
}
The store itself can still be a single SDC1 instruction, so we should
generate the same notes regardless of mips_split_64bit_move_p.
If that's right, then how about the patch below (tested on
mips64-linux-gnu, but without the shrink-wrap patches)?
Richard
gcc/
2011-09-25 Bernd Schmidt <bernds@codesourcery.com>
Richard Sandiford <rdsandiford@googlemail.com>
* config/mips/mips.c (mips_add_cfa_restore): New function.
(mips16e_save_restore_reg): Use it.
(mips_restore_reg): Likewise. Split double FPRs for
REG_CFA_RESTORE notes.
Index: gcc/config/mips/mips.c
===================================================================
--- gcc/config/mips/mips.c 2011-09-25 18:16:38.000000000 +0100
+++ gcc/config/mips/mips.c 2011-09-25 18:19:15.000000000 +0100
@@ -8202,6 +8202,15 @@ mips_frame_set (rtx mem, rtx reg)
return set;
}
+
+/* Record that the epilogue has restored call-saved register REG. */
+
+static void
+mips_add_cfa_restore (rtx reg)
+{
+ mips_epilogue.cfa_restores = alloc_reg_note (REG_CFA_RESTORE, reg,
+ mips_epilogue.cfa_restores);
+}
\f
/* If a MIPS16e SAVE or RESTORE instruction saves or restores register
mips16e_s2_s8_regs[X], it must also save the registers in indexes
@@ -8393,8 +8402,7 @@ mips16e_save_restore_reg (bool restore_p
reg = gen_rtx_REG (SImode, regno);
if (restore_p)
{
- mips_epilogue.cfa_restores = alloc_reg_note (REG_CFA_RESTORE, reg,
- mips_epilogue.cfa_restores);
+ mips_add_cfa_restore (reg);
return gen_rtx_SET (VOIDmode, reg, mem);
}
if (reg_parm_p)
@@ -10290,9 +10298,13 @@ mips_restore_reg (rtx reg, rtx mem)
$7 instead and adjust the return insn appropriately. */
if (TARGET_MIPS16 && REGNO (reg) == RETURN_ADDR_REGNUM)
reg = gen_rtx_REG (GET_MODE (reg), GP_REG_FIRST + 7);
+ else if (GET_MODE (reg) == DFmode && !TARGET_FLOAT64)
+ {
+ mips_add_cfa_restore (mips_subword (reg, true));
+ mips_add_cfa_restore (mips_subword (reg, false));
+ }
else
- mips_epilogue.cfa_restores = alloc_reg_note (REG_CFA_RESTORE, reg,
- mips_epilogue.cfa_restores);
+ mips_add_cfa_restore (reg);
mips_emit_save_slot_move (reg, mem, MIPS_EPILOGUE_TEMP (GET_MODE (reg)));
if (REGNO (reg) == REGNO (mips_epilogue.cfa_reg))
^ permalink raw reply [flat|nested] 7+ messages in thread
* Re: Handle multi-word regsiters in REG_CFA_RESTORE notes
2011-09-25 19:54 ` Richard Sandiford
@ 2011-09-27 12:57 ` Bernd Schmidt
0 siblings, 0 replies; 7+ messages in thread
From: Bernd Schmidt @ 2011-09-27 12:57 UTC (permalink / raw)
To: Richard Henderson, GCC Patches, rdsandiford
On 09/25/11 19:27, Richard Sandiford wrote:
> The store itself can still be a single SDC1 instruction, so we should
> generate the same notes regardless of mips_split_64bit_move_p.
>
> If that's right, then how about the patch below (tested on
> mips64-linux-gnu, but without the shrink-wrap patches)?
You know mips better than I do... Tested with shrink-wrapping, and it
seems to work as well. Please go ahead.
Bernd
^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2011-09-27 11:59 UTC | newest]
Thread overview: 7+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2011-09-15 2:42 Handle multi-word regsiters in REG_CFA_RESTORE notes Bernd Schmidt
2011-09-15 20:41 ` Richard Henderson
2011-09-21 15:14 ` Bernd Schmidt
2011-09-21 17:47 ` Richard Henderson
2011-09-22 22:09 ` Bernd Schmidt
2011-09-25 19:54 ` Richard Sandiford
2011-09-27 12:57 ` Bernd Schmidt
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).