On 04/28/17 20:46, Jeff Law wrote:
> On 04/28/2017 11:27 AM, Bernd Edlinger wrote:
>>>
>>
>> Yes I agree, that is probably not worth it. So I could try to remove
>> the special handling of PIC+const and see what happens.
>>
>> However the SYMBOL_REF_FUNCTION_P is another story, that part I would
>> like to keep: It happens quite often, already w/o -fpic that call
>> statements are using SYMBOL_REFs to ordinary (not weak) function
>> symbols, and may_trap returns 1 for these call statements wihch is IMHO
>> wrong.
> Hmm, thinking more about this, wasn't the original case a PIC referrence
> for something like &x[BIGNUM].
>
> Perhaps we could consider a PIC reference without other arithmetic as
> safe. That would likely pick up the SYMBOL_REF_FUNCTION_P case you want
> as well good deal many more PIC references as non-trapping.
>
Yes, I like this idea.
I tried to compile openssl with -m32 -fpic as an example, and counted
how often the mem[pic+const] is hit: that was 2353 times, all kind of
object refs.
Then I tried your idea, and only 54 unhandled pic refs remained, all of
them looking like this:
(plus:SI (reg:SI 107)
(const:SI (plus:SI (unspec:SI [
(symbol_ref:SI ("bf_init") [flags 0x2] <var_decl
0x2ac00f7bac60 bf_init>)
] UNSPEC_GOTOFF)
(const_int 4164 [0x1044]))))
I believe that is a negligible fall out from such a big code base.
Although the pic references do no longer reach the
SYMBOL_REF_FUNCTION_P in this version of the patch, I still see
that happening without -fpic option, so I left it as is.
Attached is the new version of my patch.
Bootstrapped and reg-tested on x86_64-pc-linux-gnu.
Is it OK for trunk?
Thanks
Bernd.
2017-04-29 Bernd Edlinger <bernd.edlin...@hotmail.de>
rtl-optimizatoin/79286
* ira.c (update_equiv_regs): Revert to using may_tap_p again.
* rtlanal.c (rtx_addr_can_trap_p_1): SYMBOL_REF_FUNCTION_P can never
trap. PIC register plus a const unspec without offset can never trap.
Index: gcc/ira.c
===================================================================
--- gcc/ira.c (revision 247397)
+++ gcc/ira.c (working copy)
@@ -3551,7 +3551,8 @@ update_equiv_regs (void)
if (DF_REG_DEF_COUNT (regno) == 1
&& note
&& !rtx_varies_p (XEXP (note, 0), 0)
- && def_dominates_uses (regno))
+ && (!may_trap_or_fault_p (XEXP (note, 0))
+ || def_dominates_uses (regno)))
{
rtx note_value = XEXP (note, 0);
remove_note (insn, note);
Index: gcc/rtlanal.c
===================================================================
--- gcc/rtlanal.c (revision 247397)
+++ gcc/rtlanal.c (working copy)
@@ -485,7 +485,7 @@ rtx_addr_can_trap_p_1 (const_rtx x, HOST_WIDE_INT
case SYMBOL_REF:
if (SYMBOL_REF_WEAK (x))
return 1;
- if (!CONSTANT_POOL_ADDRESS_P (x))
+ if (!CONSTANT_POOL_ADDRESS_P (x) && !SYMBOL_REF_FUNCTION_P (x))
{
tree decl;
HOST_WIDE_INT decl_size;
@@ -644,8 +644,11 @@ rtx_addr_can_trap_p_1 (const_rtx x, HOST_WIDE_INT
case PLUS:
/* An address is assumed not to trap if:
- - it is the pic register plus a constant. */
- if (XEXP (x, 0) == pic_offset_table_rtx && CONSTANT_P (XEXP (x, 1)))
+ - it is the pic register plus a const unspec without offset. */
+ if (XEXP (x, 0) == pic_offset_table_rtx
+ && GET_CODE (XEXP (x, 1)) == CONST
+ && GET_CODE (XEXP (XEXP (x, 1), 0)) == UNSPEC
+ && offset == 0)
return 0;
/* - or it is an address that can't trap plus a constant integer. */
