* [PATCH][AArch64] Properly cost zero_extend+ashift forms of ubfi[xz]
@ 2015-12-04 9:30 Kyrill Tkachov
2015-12-11 9:53 ` Kyrill Tkachov
2015-12-16 14:59 ` James Greenhalgh
0 siblings, 2 replies; 3+ messages in thread
From: Kyrill Tkachov @ 2015-12-04 9:30 UTC (permalink / raw)
To: GCC Patches; +Cc: Marcus Shawcroft, Richard Earnshaw, James Greenhalgh
[-- Attachment #1: Type: text/plain, Size: 862 bytes --]
Hi all,
We don't handle properly the patterns for the [us]bfiz and [us]bfx instructions when they
have an extend+ashift form. For example, the *<ANY_EXTEND:optab><GPI:mode>_ashl<SHORT:mode> pattern.
This leads to rtx costs recuring into the extend and assigning a cost to these patterns that is too
large.
This patch fixes that oversight.
I stumbled across this when working on a different combine patch and ended up matching the above
pattern, only to have it rejected for -mcpu=cortex-a53 due to the erroneous cost.
Bootstrapped and tested on aarch64.
Ok for trunk?
Thanks,
Kyrill
2015-12-04 Kyrylo Tkachov <kyrylo.tkachov@arm.com>
* config/aarch64/aarch64.c (aarch64_extend_bitfield_pattern_p):
New function.
(aarch64_rtx_costs, ZERO_EXTEND, SIGN_EXTEND cases): Use the above
to handle extend+shift rtxes.
[-- Warning: decoded text below may be mangled, UTF-8 assumed --]
[-- Attachment #2: aarch64-extend-bfiz.patch --]
[-- Type: text/x-patch; name=aarch64-extend-bfiz.patch, Size: 2489 bytes --]
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index c97ecdc0859e0a24792a57aeb18b2e4ea35918f4..d180f6f2d37a280ad77f34caad8496ddaa6e01b2 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -5833,6 +5833,50 @@ aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed)
return false;
}
+/* Check whether X is a bitfield operation of the form shift + extend that
+ maps down to a UBFIZ/SBFIZ/UBFX/SBFX instruction. If so, return the
+ operand to which the bitfield operation is applied to. Otherwise return
+ NULL_RTX. */
+
+static rtx
+aarch64_extend_bitfield_pattern_p (rtx x)
+{
+ rtx_code outer_code = GET_CODE (x);
+ machine_mode outer_mode = GET_MODE (x);
+
+ if (outer_code != ZERO_EXTEND && outer_code != SIGN_EXTEND
+ && outer_mode != SImode && outer_mode != DImode)
+ return NULL_RTX;
+
+ rtx inner = XEXP (x, 0);
+ rtx_code inner_code = GET_CODE (inner);
+ machine_mode inner_mode = GET_MODE (inner);
+ rtx op = NULL_RTX;
+
+ switch (inner_code)
+ {
+ case ASHIFT:
+ if (CONST_INT_P (XEXP (inner, 1))
+ && (inner_mode == QImode || inner_mode == HImode))
+ op = XEXP (inner, 0);
+ break;
+ case LSHIFTRT:
+ if (outer_code == ZERO_EXTEND && CONST_INT_P (XEXP (inner, 1))
+ && (inner_mode == QImode || inner_mode == HImode))
+ op = XEXP (inner, 0);
+ break;
+ case ASHIFTRT:
+ if (outer_code == SIGN_EXTEND && CONST_INT_P (XEXP (inner, 1))
+ && (inner_mode == QImode || inner_mode == HImode))
+ op = XEXP (inner, 0);
+ break;
+ default:
+ break;
+ }
+
+ return op;
+}
+
/* Calculate the cost of calculating X, storing it in *COST. Result
is true if the total cost of the operation has now been calculated. */
static bool
@@ -6521,6 +6565,14 @@ cost_plus:
return true;
}
+ op0 = aarch64_extend_bitfield_pattern_p (x);
+ if (op0)
+ {
+ *cost += rtx_cost (op0, mode, ZERO_EXTEND, 0, speed);
+ if (speed)
+ *cost += extra_cost->alu.bfx;
+ return true;
+ }
if (speed)
{
if (VECTOR_MODE_P (mode))
@@ -6552,6 +6604,14 @@ cost_plus:
return true;
}
+ op0 = aarch64_extend_bitfield_pattern_p (x);
+ if (op0)
+ {
+ *cost += rtx_cost (op0, mode, SIGN_EXTEND, 0, speed);
+ if (speed)
+ *cost += extra_cost->alu.bfx;
+ return true;
+ }
if (speed)
{
if (VECTOR_MODE_P (mode))
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH][AArch64] Properly cost zero_extend+ashift forms of ubfi[xz]
2015-12-04 9:30 [PATCH][AArch64] Properly cost zero_extend+ashift forms of ubfi[xz] Kyrill Tkachov
@ 2015-12-11 9:53 ` Kyrill Tkachov
2015-12-16 14:59 ` James Greenhalgh
1 sibling, 0 replies; 3+ messages in thread
From: Kyrill Tkachov @ 2015-12-11 9:53 UTC (permalink / raw)
To: GCC Patches; +Cc: Marcus Shawcroft, Richard Earnshaw, James Greenhalgh
Ping.
https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00526.html
Thanks,
Kyrill
On 04/12/15 09:30, Kyrill Tkachov wrote:
> Hi all,
>
> We don't handle properly the patterns for the [us]bfiz and [us]bfx instructions when they
> have an extend+ashift form. For example, the *<ANY_EXTEND:optab><GPI:mode>_ashl<SHORT:mode> pattern.
> This leads to rtx costs recuring into the extend and assigning a cost to these patterns that is too
> large.
>
> This patch fixes that oversight.
> I stumbled across this when working on a different combine patch and ended up matching the above
> pattern, only to have it rejected for -mcpu=cortex-a53 due to the erroneous cost.
>
> Bootstrapped and tested on aarch64.
>
> Ok for trunk?
>
> Thanks,
> Kyrill
>
> 2015-12-04 Kyrylo Tkachov <kyrylo.tkachov@arm.com>
>
> * config/aarch64/aarch64.c (aarch64_extend_bitfield_pattern_p):
> New function.
> (aarch64_rtx_costs, ZERO_EXTEND, SIGN_EXTEND cases): Use the above
> to handle extend+shift rtxes.
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: [PATCH][AArch64] Properly cost zero_extend+ashift forms of ubfi[xz]
2015-12-04 9:30 [PATCH][AArch64] Properly cost zero_extend+ashift forms of ubfi[xz] Kyrill Tkachov
2015-12-11 9:53 ` Kyrill Tkachov
@ 2015-12-16 14:59 ` James Greenhalgh
1 sibling, 0 replies; 3+ messages in thread
From: James Greenhalgh @ 2015-12-16 14:59 UTC (permalink / raw)
To: Kyrill Tkachov; +Cc: GCC Patches, Marcus Shawcroft, Richard Earnshaw
On Fri, Dec 04, 2015 at 09:30:45AM +0000, Kyrill Tkachov wrote:
> Hi all,
>
> We don't handle properly the patterns for the [us]bfiz and [us]bfx instructions when they
> have an extend+ashift form. For example, the *<ANY_EXTEND:optab><GPI:mode>_ashl<SHORT:mode> pattern.
> This leads to rtx costs recuring into the extend and assigning a cost to these patterns that is too
> large.
>
> This patch fixes that oversight.
> I stumbled across this when working on a different combine patch and ended up matching the above
> pattern, only to have it rejected for -mcpu=cortex-a53 due to the erroneous cost.
>
> Bootstrapped and tested on aarch64.
>
> Ok for trunk?
>
> Thanks,
> Kyrill
>
> 2015-12-04 Kyrylo Tkachov <kyrylo.tkachov@arm.com>
>
> * config/aarch64/aarch64.c (aarch64_extend_bitfield_pattern_p):
> New function.
> (aarch64_rtx_costs, ZERO_EXTEND, SIGN_EXTEND cases): Use the above
> to handle extend+shift rtxes.
> diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
> index c97ecdc0859e0a24792a57aeb18b2e4ea35918f4..d180f6f2d37a280ad77f34caad8496ddaa6e01b2 100644
> --- a/gcc/config/aarch64/aarch64.c
> +++ b/gcc/config/aarch64/aarch64.c
> @@ -5833,6 +5833,50 @@ aarch64_if_then_else_costs (rtx op0, rtx op1, rtx op2, int *cost, bool speed)
> return false;
> }
>
> +/* Check whether X is a bitfield operation of the form shift + extend that
> + maps down to a UBFIZ/SBFIZ/UBFX/SBFX instruction. If so, return the
> + operand to which the bitfield operation is applied to. Otherwise return
No need for that second "to" at the end of the sentence.
> + NULL_RTX. */
> +
> +static rtx
> +aarch64_extend_bitfield_pattern_p (rtx x)
> +{
> + rtx_code outer_code = GET_CODE (x);
> + machine_mode outer_mode = GET_MODE (x);
> +
> + if (outer_code != ZERO_EXTEND && outer_code != SIGN_EXTEND
> + && outer_mode != SImode && outer_mode != DImode)
> + return NULL_RTX;
> +
> + rtx inner = XEXP (x, 0);
> + rtx_code inner_code = GET_CODE (inner);
> + machine_mode inner_mode = GET_MODE (inner);
> + rtx op = NULL_RTX;
> +
> + switch (inner_code)
> + {
> + case ASHIFT:
> + if (CONST_INT_P (XEXP (inner, 1))
> + && (inner_mode == QImode || inner_mode == HImode))
> + op = XEXP (inner, 0);
> + break;
> + case LSHIFTRT:
> + if (outer_code == ZERO_EXTEND && CONST_INT_P (XEXP (inner, 1))
> + && (inner_mode == QImode || inner_mode == HImode))
> + op = XEXP (inner, 0);
> + break;
> + case ASHIFTRT:
> + if (outer_code == SIGN_EXTEND && CONST_INT_P (XEXP (inner, 1))
> + && (inner_mode == QImode || inner_mode == HImode))
> + op = XEXP (inner, 0);
> + break;
> + default:
> + break;
> + }
> +
> + return op;
> +}
> +
> /* Calculate the cost of calculating X, storing it in *COST. Result
> is true if the total cost of the operation has now been calculated. */
> static bool
> @@ -6521,6 +6565,14 @@ cost_plus:
> return true;
> }
>
> + op0 = aarch64_extend_bitfield_pattern_p (x);
> + if (op0)
> + {
> + *cost += rtx_cost (op0, mode, ZERO_EXTEND, 0, speed);
> + if (speed)
> + *cost += extra_cost->alu.bfx;
> + return true;
> + }
Newline here.
> if (speed)
> {
> if (VECTOR_MODE_P (mode))
> @@ -6552,6 +6604,14 @@ cost_plus:
> return true;
> }
>
> + op0 = aarch64_extend_bitfield_pattern_p (x);
> + if (op0)
> + {
> + *cost += rtx_cost (op0, mode, SIGN_EXTEND, 0, speed);
> + if (speed)
> + *cost += extra_cost->alu.bfx;
> + return true;
> + }
And here.
> if (speed)
> {
> if (VECTOR_MODE_P (mode))
OK with those changes.
Thanks,
James
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2015-12-16 14:59 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2015-12-04 9:30 [PATCH][AArch64] Properly cost zero_extend+ashift forms of ubfi[xz] Kyrill Tkachov
2015-12-11 9:53 ` Kyrill Tkachov
2015-12-16 14:59 ` James Greenhalgh
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).