* [PATCH] aarch64: Improve cost of `a ? {-,}1 : b`
@ 2023-11-27 4:33 Andrew Pinski
2023-11-27 15:35 ` Richard Sandiford
0 siblings, 1 reply; 5+ messages in thread
From: Andrew Pinski @ 2023-11-27 4:33 UTC (permalink / raw)
To: gcc-patches
While looking into PR 112454, I found the cost for
`(if_then_else (cmp) (const_int 1) (reg))` was being recorded as 8
(or `COSTS_N_INSNS (2)`) but it should have been 4 (or `COSTS_N_INSNS (1)`).
This improves the cost by not adding the cost of `(const_int 1)` to
the total cost.
It does not does not fix PR 112454 as that requires other changes to forwprop
the `(const_int 1)` earlier than combine.
Bootstrapped and tested on aarch64-linux-gnu with no regressions.
gcc/ChangeLog:
* config/aarch64/aarch64.cc (aarch64_if_then_else_costs):
Don't add the cost of `1` or `-1`.
Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
---
gcc/config/aarch64/aarch64.cc | 13 ++++++++++---
1 file changed, 10 insertions(+), 3 deletions(-)
diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
index f6f6f94bf43..63241c5aaa5 100644
--- a/gcc/config/aarch64/aarch64.cc
+++ b/gcc/config/aarch64/aarch64.cc
@@ -11642,9 +11642,16 @@ aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed)
/* CSINV/NEG with zero extend + const 0 (*csinv3_uxtw_insn3). */
op1 = XEXP (inner, 0);
}
-
- *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
- *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
+ if (op2 == constm1_rtx || op2 == const1_rtx)
+ *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
+ else if (op1 == constm1_rtx || op1 == const1_rtx)
+ *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
+ else
+ {
+ *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
+ *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 1, speed);
+ }
+
return true;
}
--
2.34.1
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] aarch64: Improve cost of `a ? {-,}1 : b`
2023-11-27 4:33 [PATCH] aarch64: Improve cost of `a ? {-,}1 : b` Andrew Pinski
@ 2023-11-27 15:35 ` Richard Sandiford
2023-11-27 15:44 ` Richard Sandiford
2023-11-27 16:11 ` Andrew Pinski (QUIC)
0 siblings, 2 replies; 5+ messages in thread
From: Richard Sandiford @ 2023-11-27 15:35 UTC (permalink / raw)
To: Andrew Pinski; +Cc: gcc-patches
Andrew Pinski <quic_apinski@quicinc.com> writes:
> While looking into PR 112454, I found the cost for
> `(if_then_else (cmp) (const_int 1) (reg))` was being recorded as 8
> (or `COSTS_N_INSNS (2)`) but it should have been 4 (or `COSTS_N_INSNS (1)`).
> This improves the cost by not adding the cost of `(const_int 1)` to
> the total cost.
>
> It does not does not fix PR 112454 as that requires other changes to forwprop
> the `(const_int 1)` earlier than combine.
>
> Bootstrapped and tested on aarch64-linux-gnu with no regressions.
>
> gcc/ChangeLog:
>
> * config/aarch64/aarch64.cc (aarch64_if_then_else_costs):
> Don't add the cost of `1` or `-1`.
>
> Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
> ---
> gcc/config/aarch64/aarch64.cc | 13 ++++++++++---
> 1 file changed, 10 insertions(+), 3 deletions(-)
>
> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
> index f6f6f94bf43..63241c5aaa5 100644
> --- a/gcc/config/aarch64/aarch64.cc
> +++ b/gcc/config/aarch64/aarch64.cc
> @@ -11642,9 +11642,16 @@ aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed)
> /* CSINV/NEG with zero extend + const 0 (*csinv3_uxtw_insn3). */
> op1 = XEXP (inner, 0);
> }
> -
> - *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> - *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
> + if (op2 == constm1_rtx || op2 == const1_rtx)
> + *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> + else if (op1 == constm1_rtx || op1 == const1_rtx)
> + *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
It looks like this is really an extra option on top of the previous
if-else chain, since it only applies when OP1 and OP2 are still the
operands of the if_then_else. So how about:
else if (op1 == constm1_rtx || op1 == const1_rtx)
{
/* Use CSINV. */
*cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
return true;
}
else if (op2 == constm1_rtx || op2 == const1_rtx)
{
/* Use CSINV. */
*cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
return true;
}
leaving the code to fall through to:
*cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
*cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
return true;
as it does currently. OK in that form if you agree.
Let me know if you don't. But in that case:
> + else
> + {
> + *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> + *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 1, speed);
should be 2, speed
> + }
> +
Thanks,
Richard
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] aarch64: Improve cost of `a ? {-,}1 : b`
2023-11-27 15:35 ` Richard Sandiford
@ 2023-11-27 15:44 ` Richard Sandiford
2023-11-27 16:11 ` Andrew Pinski (QUIC)
1 sibling, 0 replies; 5+ messages in thread
From: Richard Sandiford @ 2023-11-27 15:44 UTC (permalink / raw)
To: Andrew Pinski; +Cc: gcc-patches
Richard Sandiford <richard.sandiford@arm.com> writes:
> Andrew Pinski <quic_apinski@quicinc.com> writes:
>> While looking into PR 112454, I found the cost for
>> `(if_then_else (cmp) (const_int 1) (reg))` was being recorded as 8
>> (or `COSTS_N_INSNS (2)`) but it should have been 4 (or `COSTS_N_INSNS (1)`).
>> This improves the cost by not adding the cost of `(const_int 1)` to
>> the total cost.
>>
>> It does not does not fix PR 112454 as that requires other changes to forwprop
>> the `(const_int 1)` earlier than combine.
>>
>> Bootstrapped and tested on aarch64-linux-gnu with no regressions.
>>
>> gcc/ChangeLog:
>>
>> * config/aarch64/aarch64.cc (aarch64_if_then_else_costs):
>> Don't add the cost of `1` or `-1`.
>>
>> Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
>> ---
>> gcc/config/aarch64/aarch64.cc | 13 ++++++++++---
>> 1 file changed, 10 insertions(+), 3 deletions(-)
>>
>> diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
>> index f6f6f94bf43..63241c5aaa5 100644
>> --- a/gcc/config/aarch64/aarch64.cc
>> +++ b/gcc/config/aarch64/aarch64.cc
>> @@ -11642,9 +11642,16 @@ aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed)
>> /* CSINV/NEG with zero extend + const 0 (*csinv3_uxtw_insn3). */
>> op1 = XEXP (inner, 0);
>> }
>> -
>> - *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
>> - *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
>> + if (op2 == constm1_rtx || op2 == const1_rtx)
>> + *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
>> + else if (op1 == constm1_rtx || op1 == const1_rtx)
>> + *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
>
> It looks like this is really an extra option on top of the previous
> if-else chain, since it only applies when OP1 and OP2 are still the
> operands of the if_then_else. So how about:
>
> else if (op1 == constm1_rtx || op1 == const1_rtx)
> {
> /* Use CSINV. */
eh, of course I meant CSINV or CSINC...
> *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
> return true;
> }
> else if (op2 == constm1_rtx || op2 == const1_rtx)
> {
> /* Use CSINV. */
> *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> return true;
> }
>
> leaving the code to fall through to:
>
> *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
> return true;
>
> as it does currently. OK in that form if you agree.
>
> Let me know if you don't. But in that case:
>
>> + else
>> + {
>> + *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
>> + *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 1, speed);
>
> should be 2, speed
>
>> + }
>> +
>
> Thanks,
> Richard
^ permalink raw reply [flat|nested] 5+ messages in thread
* RE: [PATCH] aarch64: Improve cost of `a ? {-,}1 : b`
2023-11-27 15:35 ` Richard Sandiford
2023-11-27 15:44 ` Richard Sandiford
@ 2023-11-27 16:11 ` Andrew Pinski (QUIC)
1 sibling, 0 replies; 5+ messages in thread
From: Andrew Pinski (QUIC) @ 2023-11-27 16:11 UTC (permalink / raw)
To: Richard Sandiford, Andrew Pinski (QUIC); +Cc: gcc-patches
> -----Original Message-----
> From: Richard Sandiford <richard.sandiford@arm.com>
> Sent: Monday, November 27, 2023 7:35 AM
> To: Andrew Pinski (QUIC) <quic_apinski@quicinc.com>
> Cc: gcc-patches@gcc.gnu.org
> Subject: Re: [PATCH] aarch64: Improve cost of `a ? {-,}1 : b`
>
> Andrew Pinski <quic_apinski@quicinc.com> writes:
> > While looking into PR 112454, I found the cost for `(if_then_else
> > (cmp) (const_int 1) (reg))` was being recorded as 8 (or `COSTS_N_INSNS
> > (2)`) but it should have been 4 (or `COSTS_N_INSNS (1)`).
> > This improves the cost by not adding the cost of `(const_int 1)` to
> > the total cost.
> >
> > It does not does not fix PR 112454 as that requires other changes to
> > forwprop the `(const_int 1)` earlier than combine.
> >
> > Bootstrapped and tested on aarch64-linux-gnu with no regressions.
> >
> > gcc/ChangeLog:
> >
> > * config/aarch64/aarch64.cc (aarch64_if_then_else_costs):
> > Don't add the cost of `1` or `-1`.
> >
> > Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
> > ---
> > gcc/config/aarch64/aarch64.cc | 13 ++++++++++---
> > 1 file changed, 10 insertions(+), 3 deletions(-)
> >
> > diff --git a/gcc/config/aarch64/aarch64.cc
> > b/gcc/config/aarch64/aarch64.cc index f6f6f94bf43..63241c5aaa5 100644
> > --- a/gcc/config/aarch64/aarch64.cc
> > +++ b/gcc/config/aarch64/aarch64.cc
> > @@ -11642,9 +11642,16 @@ aarch64_if_then_else_costs (rtx op0, rtx op1,
> rtx op2, int *cost, bool speed)
> > /* CSINV/NEG with zero extend + const 0 (*csinv3_uxtw_insn3). */
> > op1 = XEXP (inner, 0);
> > }
> > -
> > - *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> > - *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
> > + if (op2 == constm1_rtx || op2 == const1_rtx)
> > + *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> > + else if (op1 == constm1_rtx || op1 == const1_rtx)
> > + *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
>
> It looks like this is really an extra option on top of the previous if-else chain,
> since it only applies when OP1 and OP2 are still the operands of the
> if_then_else. So how about:
>
> else if (op1 == constm1_rtx || op1 == const1_rtx)
> {
> /* Use CSINV. */
> *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
> return true;
> }
> else if (op2 == constm1_rtx || op2 == const1_rtx)
> {
> /* Use CSINV. */
> *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> return true;
> }
>
> leaving the code to fall through to:
>
> *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
> return true;
>
> as it does currently. OK in that form if you agree.
Yes I think this is the correct way of implementing this, Let me test it and get back to you.
Thanks,
Andrew
>
> Let me know if you don't. But in that case:
>
> > + else
> > + {
> > + *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
> > + *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 1, speed);
>
> should be 2, speed
>
> > + }
> > +
>
> Thanks,
> Richard
^ permalink raw reply [flat|nested] 5+ messages in thread
* [PATCH] aarch64: Improve cost of `a ? {-,}1 : b`
@ 2023-11-27 23:08 Andrew Pinski
0 siblings, 0 replies; 5+ messages in thread
From: Andrew Pinski @ 2023-11-27 23:08 UTC (permalink / raw)
To: gcc-patches
While looking into PR 112454, I found the cost for
`(if_then_else (cmp) (const_int 1) (reg))` was being recorded as 8
(or `COSTS_N_INSNS (2)`) but it should have been 4 (or `COSTS_N_INSNS (1)`).
This improves the cost by not adding the cost of `(const_int 1)` to
the total cost.
It does not does not fully fix PR 112454 as that requires other changes to forwprop
the `(const_int 1)` earlier than combine. Though we do fix the loop case where the
constant was only used once.
Committed as approved after bootstrapped and tested on aarch64-linux-gnu with no regressions.
gcc/ChangeLog:
* config/aarch64/aarch64.cc (aarch64_if_then_else_costs):
Handle csinv/csinc case of 1/-1.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/csinc-3.c: New test.
Signed-off-by: Andrew Pinski <quic_apinski@quicinc.com>
---
gcc/config/aarch64/aarch64.cc | 12 ++++++++++++
gcc/testsuite/gcc.target/aarch64/csinc-3.c | 10 ++++++++++
2 files changed, 22 insertions(+)
create mode 100644 gcc/testsuite/gcc.target/aarch64/csinc-3.c
diff --git a/gcc/config/aarch64/aarch64.cc b/gcc/config/aarch64/aarch64.cc
index b2093430937..4fd8c2de43a 100644
--- a/gcc/config/aarch64/aarch64.cc
+++ b/gcc/config/aarch64/aarch64.cc
@@ -11607,6 +11607,18 @@ aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed)
/* CSINV/NEG with zero extend + const 0 (*csinv3_uxtw_insn3). */
op1 = XEXP (inner, 0);
}
+ else if (op1 == constm1_rtx || op1 == const1_rtx)
+ {
+ /* Use CSINV or CSINC. */
+ *cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
+ return true;
+ }
+ else if (op2 == constm1_rtx || op2 == const1_rtx)
+ {
+ /* Use CSINV or CSINC. */
+ *cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
+ return true;
+ }
*cost += rtx_cost (op1, VOIDmode, IF_THEN_ELSE, 1, speed);
*cost += rtx_cost (op2, VOIDmode, IF_THEN_ELSE, 2, speed);
diff --git a/gcc/testsuite/gcc.target/aarch64/csinc-3.c b/gcc/testsuite/gcc.target/aarch64/csinc-3.c
new file mode 100644
index 00000000000..bde131a584e
--- /dev/null
+++ b/gcc/testsuite/gcc.target/aarch64/csinc-3.c
@@ -0,0 +1,10 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -fno-tree-vectorize" } */
+
+int f(int *a, int n, int *b, int d)
+{
+ for(int i = 0; i < n; i++)
+ b[i] = a[i] == 100 ? 1 : d;
+ /* { dg-final { scan-assembler "csinc\tw\[0-9\].*wzr" } } */
+ return 0;
+}
--
2.34.1
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2023-11-27 23:09 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-11-27 4:33 [PATCH] aarch64: Improve cost of `a ? {-,}1 : b` Andrew Pinski
2023-11-27 15:35 ` Richard Sandiford
2023-11-27 15:44 ` Richard Sandiford
2023-11-27 16:11 ` Andrew Pinski (QUIC)
2023-11-27 23:08 Andrew Pinski
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).