This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH i386 AVX512] [22/n] Extend unaligned loads & stores.
- From: Kirill Yukhin <kirill dot yukhin at gmail dot com>
- To: Uros Bizjak <ubizjak at gmail dot com>
- Cc: Jakub Jelinek <jakub at redhat dot com>, Richard Henderson <rth at redhat dot com>, GCC Patches <gcc-patches at gcc dot gnu dot org>
- Date: Tue, 26 Aug 2014 23:55:47 +0400
- Subject: Re: [PATCH i386 AVX512] [22/n] Extend unaligned loads & stores.
- Authentication-results: sourceware.org; auth=none
- References: <20140822115114 dot GB47539 at msticlxl57 dot ims dot intel dot com> <CAFULd4bP97be2MdiM=11fKQKGzMEDTJGigsqSKxam8mWrdwrNQ at mail dot gmail dot com>
Hello UroÅ,
On 23 Aug 09:44, Uros Bizjak wrote:
> On Fri, Aug 22, 2014 at 1:51 PM, Kirill Yukhin <kirill.yukhin@gmail.com> wrote:
>
> > This patch extends unaligned loads and stores patterns.
> At this stage, I'd still prefer simple constraints (the solution,
> proposed above), even for the price of additional patterns. Looking at
> the patterns, it is quite hard to calculate final condition for the
> particular mode/target combo, even without enable attribute and
> conditional operand constraints/predicates. With the solution above,
> the complexity is conveniently pushed to mask define_subst attribute.
In the bottom patch which splits unaligned ld/st patterns.
Bootstrapped and avx512-regtested on simulator.
gcc/
* config/i386/sse.md
(define_mode_iterator VI48_AVX512VL): New.
(define_mode_iterator VI_UNALIGNED_LOADSTORE): Delete.
(define_mode_iterator VI_ULOADSTORE_AVX512BW): New.
(define_mode_iterator VI_ULOADSTORE_AVX512F): Ditto.
(define_expand "<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
with VI1): Change mode iterator.
(define_expand "<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
with VI_ULOADSTORE_AVX512BW): New.
(define_expand "<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
with VI_ULOADSTORE_AVX512F): Ditto.
(define_insn "*<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
with VI1): Change mode iterator.
(define_insn "*<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
with VI_ULOADSTORE_AVX512BW): New.
(define_insn "*<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
with VI_ULOADSTORE_AVX512F): Ditto.
(define_insn "<sse2_avx_avx512f>_storedqu<mode>
with VI1): Change mode iterator.
(define_insn "<sse2_avx_avx512f>_storedqu<mode>
with VI_ULOADSTORE_AVX512BW): New.
(define_insn "<sse2_avx_avx512f>_storedqu<mode>
with VI_ULOADSTORE_AVX512F): Ditto.
(define_insn "avx512f_storedqu<mode>_mask"): Delete.
(define_insn "<avx512>_storedqu<mode>_mask" with
VI48_AVX512VL): New.
(define_insn "<avx512>_storedqu<mode>_mask" with
VI12_AVX512VL): Ditto.
Is it ok for trunk?
--
Thanks, K
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 0624582..0245ec4 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -235,6 +235,10 @@
(define_mode_iterator VF_512
[V16SF V8DF])
+(define_mode_iterator VI48_AVX512VL
+ [V16SI (V8SI "TARGET_AVX512VL") (V4SI "TARGET_AVX512VL")
+ V8DI (V4DI "TARGET_AVX512VL") (V2DI "TARGET_AVX512VL")])
+
(define_mode_iterator VF2_AVX512VL
[V8DF (V4DF "TARGET_AVX512VL") (V2DF "TARGET_AVX512VL")])
@@ -259,9 +263,13 @@
(define_mode_iterator VI1
[(V32QI "TARGET_AVX") V16QI])
-(define_mode_iterator VI_UNALIGNED_LOADSTORE
- [(V32QI "TARGET_AVX") V16QI
- (V16SI "TARGET_AVX512F") (V8DI "TARGET_AVX512F")])
+(define_mode_iterator VI_ULOADSTORE_AVX512BW
+ [V64QI
+ V32HI (V8HI "TARGET_AVX512VL") (V16HI "TARGET_AVX512VL")])
+
+(define_mode_iterator VI_ULOADSTORE_AVX512F
+ [V16SI (V8SI "TARGET_AVX512VL") (V4SI "TARGET_AVX512VL")
+ V8DI (V4DI "TARGET_AVX512VL") (V2DI "TARGET_AVX512VL")])
;; All DImode vector integer modes
(define_mode_iterator VI8
@@ -1172,18 +1180,18 @@
(set_attr "prefix" "evex")
(set_attr "mode" "<sseinsnmode>")])
+/* For AVX, normal *mov<mode>_internal pattern will handle unaligned loads
+ just fine if misaligned_operand is true, and without the UNSPEC it can
+ be combined with arithmetic instructions. If misaligned_operand is
+ false, still emit UNSPEC_LOADU insn to honor user's request for
+ misaligned load. */
(define_expand "<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
- [(set (match_operand:VI_UNALIGNED_LOADSTORE 0 "register_operand")
- (unspec:VI_UNALIGNED_LOADSTORE
- [(match_operand:VI_UNALIGNED_LOADSTORE 1 "nonimmediate_operand")]
+ [(set (match_operand:VI1 0 "register_operand")
+ (unspec:VI1
+ [(match_operand:VI1 1 "nonimmediate_operand")]
UNSPEC_LOADU))]
- "TARGET_SSE2 && <mask_mode512bit_condition>"
+ "TARGET_SSE2 && <mask_avx512vl_condition> && <mask_avx512bw_condition>"
{
- /* For AVX, normal *mov<mode>_internal pattern will handle unaligned loads
- just fine if misaligned_operand is true, and without the UNSPEC it can
- be combined with arithmetic instructions. If misaligned_operand is
- false, still emit UNSPEC_LOADU insn to honor user's request for
- misaligned load. */
if (TARGET_AVX
&& misaligned_operand (operands[1], <MODE>mode))
{
@@ -1197,25 +1205,61 @@
}
})
+(define_expand "<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
+ [(set (match_operand:VI_ULOADSTORE_AVX512BW 0 "register_operand")
+ (unspec:VI_ULOADSTORE_AVX512BW
+ [(match_operand:VI_ULOADSTORE_AVX512BW 1 "nonimmediate_operand")]
+ UNSPEC_LOADU))]
+ "TARGET_AVX512BW"
+{
+ if (misaligned_operand (operands[1], <MODE>mode))
+ {
+ rtx src = operands[1];
+ if (<mask_applied>)
+ src = gen_rtx_VEC_MERGE (<MODE>mode, operands[1],
+ operands[2 * <mask_applied>],
+ operands[3 * <mask_applied>]);
+ emit_insn (gen_rtx_SET (VOIDmode, operands[0], src));
+ DONE;
+ }
+})
+
+(define_expand "<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
+ [(set (match_operand:VI_ULOADSTORE_AVX512F 0 "register_operand")
+ (unspec:VI_ULOADSTORE_AVX512F
+ [(match_operand:VI_ULOADSTORE_AVX512F 1 "nonimmediate_operand")]
+ UNSPEC_LOADU))]
+ "TARGET_AVX512F"
+{
+ if (misaligned_operand (operands[1], <MODE>mode))
+ {
+ rtx src = operands[1];
+ if (<mask_applied>)
+ src = gen_rtx_VEC_MERGE (<MODE>mode, operands[1],
+ operands[2 * <mask_applied>],
+ operands[3 * <mask_applied>]);
+ emit_insn (gen_rtx_SET (VOIDmode, operands[0], src));
+ DONE;
+ }
+})
+
(define_insn "*<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
- [(set (match_operand:VI_UNALIGNED_LOADSTORE 0 "register_operand" "=v")
- (unspec:VI_UNALIGNED_LOADSTORE
- [(match_operand:VI_UNALIGNED_LOADSTORE 1 "nonimmediate_operand" "vm")]
+ [(set (match_operand:VI1 0 "register_operand" "=v")
+ (unspec:VI1
+ [(match_operand:VI1 1 "nonimmediate_operand" "vm")]
UNSPEC_LOADU))]
- "TARGET_SSE2 && <mask_mode512bit_condition>"
+ "TARGET_SSE2 && <mask_avx512vl_condition> && <mask_avx512bw_condition>"
{
switch (get_attr_mode (insn))
{
case MODE_V8SF:
case MODE_V4SF:
return "%vmovups\t{%1, %0|%0, %1}";
- case MODE_XI:
- if (<MODE>mode == V8DImode)
- return "vmovdqu64\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}";
- else
- return "vmovdqu32\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}";
default:
- return "%vmovdqu\t{%1, %0|%0, %1}";
+ if (!(TARGET_AVX512VL && TARGET_AVX512BW))
+ return "%vmovdqu\t{%1, %0|%0, %1}";
+ else
+ return "vmovdqu<ssescalarsize>\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}";
}
}
[(set_attr "type" "ssemov")
@@ -1238,10 +1282,34 @@
]
(const_string "<sseinsnmode>")))])
+(define_insn "*<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
+ [(set (match_operand:VI_ULOADSTORE_AVX512BW 0 "register_operand" "=v")
+ (unspec:VI_ULOADSTORE_AVX512BW
+ [(match_operand:VI_ULOADSTORE_AVX512BW 1 "nonimmediate_operand" "vm")]
+ UNSPEC_LOADU))]
+ "TARGET_AVX512BW"
+ "vmovdqu<ssescalarsize>\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}";
+ [(set_attr "type" "ssemov")
+ (set_attr "movu" "1")
+ (set_attr "ssememalign" "8")
+ (set_attr "prefix" "maybe_evex")])
+
+(define_insn "*<sse2_avx_avx512f>_loaddqu<mode><mask_name>"
+ [(set (match_operand:VI_ULOADSTORE_AVX512F 0 "register_operand" "=v")
+ (unspec:VI_ULOADSTORE_AVX512F
+ [(match_operand:VI_ULOADSTORE_AVX512F 1 "nonimmediate_operand" "vm")]
+ UNSPEC_LOADU))]
+ "TARGET_AVX512F"
+ "vmovdqu<ssescalarsize>\t{%1, %0<mask_operand2>|%0<mask_operand2>, %1}";
+ [(set_attr "type" "ssemov")
+ (set_attr "movu" "1")
+ (set_attr "ssememalign" "8")
+ (set_attr "prefix" "maybe_evex")])
+
(define_insn "<sse2_avx_avx512f>_storedqu<mode>"
- [(set (match_operand:VI_UNALIGNED_LOADSTORE 0 "memory_operand" "=m")
- (unspec:VI_UNALIGNED_LOADSTORE
- [(match_operand:VI_UNALIGNED_LOADSTORE 1 "register_operand" "v")]
+ [(set (match_operand:VI1 0 "memory_operand" "=m")
+ (unspec:VI1
+ [(match_operand:VI1 1 "register_operand" "v")]
UNSPEC_STOREU))]
"TARGET_SSE2"
{
@@ -1251,13 +1319,16 @@
case MODE_V8SF:
case MODE_V4SF:
return "%vmovups\t{%1, %0|%0, %1}";
- case MODE_XI:
- if (<MODE>mode == V8DImode)
- return "vmovdqu64\t{%1, %0|%0, %1}";
- else
- return "vmovdqu32\t{%1, %0|%0, %1}";
default:
- return "%vmovdqu\t{%1, %0|%0, %1}";
+ switch (<MODE>mode)
+ {
+ case V32QImode:
+ case V16QImode:
+ if (!(TARGET_AVX512VL && TARGET_AVX512BW))
+ return "%vmovdqu\t{%1, %0|%0, %1}";
+ default:
+ return "vmovdqu<ssescalarsize>\t{%1, %0|%0, %1}";
+ }
}
}
[(set_attr "type" "ssemov")
@@ -1281,21 +1352,56 @@
]
(const_string "<sseinsnmode>")))])
-(define_insn "avx512f_storedqu<mode>_mask"
- [(set (match_operand:VI48_512 0 "memory_operand" "=m")
- (vec_merge:VI48_512
- (unspec:VI48_512
- [(match_operand:VI48_512 1 "register_operand" "v")]
+(define_insn "<sse2_avx_avx512f>_storedqu<mode>"
+ [(set (match_operand:VI_ULOADSTORE_AVX512BW 0 "memory_operand" "=m")
+ (unspec:VI_ULOADSTORE_AVX512BW
+ [(match_operand:VI_ULOADSTORE_AVX512BW 1 "register_operand" "v")]
+ UNSPEC_STOREU))]
+ "TARGET_AVX512BW"
+ "vmovdqu<ssescalarsize>\t{%1, %0|%0, %1}"
+ [(set_attr "type" "ssemov")
+ (set_attr "movu" "1")
+ (set_attr "ssememalign" "8")
+ (set_attr "prefix" "maybe_evex")])
+
+(define_insn "<sse2_avx_avx512f>_storedqu<mode>"
+ [(set (match_operand:VI_ULOADSTORE_AVX512F 0 "memory_operand" "=m")
+ (unspec:VI_ULOADSTORE_AVX512F
+ [(match_operand:VI_ULOADSTORE_AVX512F 1 "register_operand" "v")]
+ UNSPEC_STOREU))]
+ "TARGET_AVX512F"
+ "vmovdqu<ssescalarsize>\t{%1, %0|%0, %1}"
+ [(set_attr "type" "ssemov")
+ (set_attr "movu" "1")
+ (set_attr "ssememalign" "8")
+ (set_attr "prefix" "maybe_vex")])
+
+(define_insn "<avx512>_storedqu<mode>_mask"
+ [(set (match_operand:VI48_AVX512VL 0 "memory_operand" "=m")
+ (vec_merge:VI48_AVX512VL
+ (unspec:VI48_AVX512VL
+ [(match_operand:VI48_AVX512VL 1 "register_operand" "v")]
UNSPEC_STOREU)
(match_dup 0)
(match_operand:<avx512fmaskmode> 2 "register_operand" "Yk")))]
"TARGET_AVX512F"
-{
- if (<MODE>mode == V8DImode)
- return "vmovdqu64\t{%1, %0%{%2%}|%0%{%2%}, %1}";
- else
- return "vmovdqu32\t{%1, %0%{%2%}|%0%{%2%}, %1}";
-}
+ "vmovdqu<ssescalarsize>\t{%1, %0%{%2%}|%0%{%2%}, %1}"
+ [(set_attr "type" "ssemov")
+ (set_attr "movu" "1")
+ (set_attr "memory" "store")
+ (set_attr "prefix" "evex")
+ (set_attr "mode" "<sseinsnmode>")])
+
+(define_insn "<avx512>_storedqu<mode>_mask"
+ [(set (match_operand:VI12_AVX512VL 0 "memory_operand" "=m")
+ (vec_merge:VI12_AVX512VL
+ (unspec:VI12_AVX512VL
+ [(match_operand:VI12_AVX512VL 1 "register_operand" "v")]
+ UNSPEC_STOREU)
+ (match_dup 0)
+ (match_operand:<avx512fmaskmode> 2 "register_operand" "Yk")))]
+ "TARGET_AVX512BW"
+ "vmovdqu<ssescalarsize>\t{%1, %0%{%2%}|%0%{%2%}, %1}"
[(set_attr "type" "ssemov")
(set_attr "movu" "1")
(set_attr "memory" "store")