From: "David Sherwood" <david.sherwood@arm.com>
To: <gcc-patches@gcc.gnu.org>
Subject: New patch: [AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe.
Date: Thu, 13 Nov 2014 10:11:00 -0000 [thread overview]
Message-ID: <000701cfff29$ea4fbcf0$beef36d0$@arm.com> (raw)
In-Reply-To:
[-- Attachment #1: Type: text/plain, Size: 1795 bytes --]
Hi All,
I have successfully rebased this and tested in conjunction with a patch from
Alan Hayward ([AArch64] [BE] Fix vector load/stores to not use ld1/st1), who
should be submitting a new version shortly. Built and tested on:
aarch64-none-elf
aarch64_be-none-elf
x86_64-linux-gnu
Regards,
David Sherwood.
-----Original Message-----
From: David Sherwood [mailto:david.sherwood@arm.com]
Sent: 28 October 2014 08:55
To: 'gcc-patches@gcc.gnu.org'
Subject: RE: [AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe.
Hi,
Sorry to bother you again. Could someone take a look at this change
please if they have time?
Thanks!
David.
-----Original Message-----
From: David Sherwood [mailto:david.sherwood@arm.com]
Sent: 10 October 2014 15:48
To: gcc-patches@gcc.gnu.org
Subject: [AArch64] [BE] [1/2] Make large opaque integer modes endianness-safe.
Hi,
I have a fix (originally written by Tejas Belagod) for the following bug:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=59810
Could someone take a look please?
Thanks!
David Sherwood.
ChangeLog:
gcc/:
2014-11-13 David Sherwood <david.sherwood@arm.com>
* config/aarch64/aarch64-protos.h (aarch64_simd_attr_length_rglist,
aarch64_reverse_mask): New decls.
* config/aarch64/iterators.md (UNSPEC_REV_REGLIST): New enum.
* config/aarch64/iterators.md (insn_count): New mode_attr.
* config/aarch64/aarch64-simd.md (vec_store_lanes(o/c/x)i,
vec_load_lanes(o/c/x)i): Fixed to work for Big Endian.
* config/aarch64/aarch64-simd.md (aarch64_rev_reglist,
aarch64_simd_(ld/st)(2/3/4)): Added.
* config/aarch64/aarch64.c (aarch64_simd_attr_length_rglist,
aarch64_reverse_mask): Added.
[-- Attachment #2: ccmc_v1_rebase.patch --]
[-- Type: application/octet-stream, Size: 9995 bytes --]
diff --git a/gcc/config/aarch64/aarch64-protos.h b/gcc/config/aarch64/aarch64-protos.h
index 470b9eb..494a1ae 100644
--- a/gcc/config/aarch64/aarch64-protos.h
+++ b/gcc/config/aarch64/aarch64-protos.h
@@ -196,6 +196,8 @@ bool aarch64_modes_tieable_p (machine_mode mode1,
bool aarch64_move_imm (HOST_WIDE_INT, machine_mode);
bool aarch64_mov_operand_p (rtx, enum aarch64_symbol_context,
machine_mode);
+int aarch64_simd_attr_length_rglist (enum machine_mode);
+rtx aarch64_reverse_mask (enum machine_mode);
bool aarch64_offset_7bit_signed_scaled_p (machine_mode, HOST_WIDE_INT);
char *aarch64_output_scalar_simd_mov_immediate (rtx, machine_mode);
char *aarch64_output_simd_mov_immediate (rtx, machine_mode, unsigned);
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index ef196e4..d3aed80 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -4035,7 +4035,7 @@
;; Patterns for vector struct loads and stores.
-(define_insn "vec_load_lanesoi<mode>"
+(define_insn "aarch64_simd_ld2<mode>"
[(set (match_operand:OI 0 "register_operand" "=w")
(unspec:OI [(match_operand:OI 1 "aarch64_simd_struct_operand" "Utv")
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
@@ -4067,7 +4067,26 @@
[(set_attr "type" "neon_load2_one_lane")]
)
-(define_insn "vec_store_lanesoi<mode>"
+(define_expand "vec_load_lanesoi<mode>"
+ [(set (match_operand:OI 0 "register_operand" "=w")
+ (unspec:OI [(match_operand:OI 1 "aarch64_simd_struct_operand" "Utv")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ UNSPEC_LD2))]
+ "TARGET_SIMD"
+{
+ if (BYTES_BIG_ENDIAN)
+ {
+ rtx tmp = gen_reg_rtx (OImode);
+ rtx mask = aarch64_reverse_mask (<MODE>mode);
+ emit_insn (gen_aarch64_simd_ld2<mode> (tmp, operands[1]));
+ emit_insn (gen_aarch64_rev_reglistoi (operands[0], tmp, mask));
+ }
+ else
+ emit_insn (gen_aarch64_simd_ld2<mode> (operands[0], operands[1]));
+ DONE;
+})
+
+(define_insn "aarch64_simd_st2<mode>"
[(set (match_operand:OI 0 "aarch64_simd_struct_operand" "=Utv")
(unspec:OI [(match_operand:OI 1 "register_operand" "w")
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
@@ -4088,7 +4107,26 @@
[(set_attr "type" "neon_store3_one_lane<q>")]
)
-(define_insn "vec_load_lanesci<mode>"
+(define_expand "vec_store_lanesoi<mode>"
+ [(set (match_operand:OI 0 "aarch64_simd_struct_operand" "=Utv")
+ (unspec:OI [(match_operand:OI 1 "register_operand" "w")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ UNSPEC_ST2))]
+ "TARGET_SIMD"
+{
+ if (BYTES_BIG_ENDIAN)
+ {
+ rtx tmp = gen_reg_rtx (OImode);
+ rtx mask = aarch64_reverse_mask (<MODE>mode);
+ emit_insn (gen_aarch64_rev_reglistoi (tmp, operands[1], mask));
+ emit_insn (gen_aarch64_simd_st2<mode> (operands[0], tmp));
+ }
+ else
+ emit_insn (gen_aarch64_simd_st2<mode> (operands[0], operands[1]));
+ DONE;
+})
+
+(define_insn "aarch64_simd_ld3<mode>"
[(set (match_operand:CI 0 "register_operand" "=w")
(unspec:CI [(match_operand:CI 1 "aarch64_simd_struct_operand" "Utv")
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
@@ -4120,7 +4158,26 @@
[(set_attr "type" "neon_load3_one_lane")]
)
-(define_insn "vec_store_lanesci<mode>"
+(define_expand "vec_load_lanesci<mode>"
+ [(set (match_operand:CI 0 "register_operand" "=w")
+ (unspec:CI [(match_operand:CI 1 "aarch64_simd_struct_operand" "Utv")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ UNSPEC_LD3))]
+ "TARGET_SIMD"
+{
+ if (BYTES_BIG_ENDIAN)
+ {
+ rtx tmp = gen_reg_rtx (CImode);
+ rtx mask = aarch64_reverse_mask (<MODE>mode);
+ emit_insn (gen_aarch64_simd_ld3<mode> (tmp, operands[1]));
+ emit_insn (gen_aarch64_rev_reglistci (operands[0], tmp, mask));
+ }
+ else
+ emit_insn (gen_aarch64_simd_ld3<mode> (operands[0], operands[1]));
+ DONE;
+})
+
+(define_insn "aarch64_simd_st3<mode>"
[(set (match_operand:CI 0 "aarch64_simd_struct_operand" "=Utv")
(unspec:CI [(match_operand:CI 1 "register_operand" "w")
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
@@ -4141,7 +4198,26 @@
[(set_attr "type" "neon_store3_one_lane<q>")]
)
-(define_insn "vec_load_lanesxi<mode>"
+(define_expand "vec_store_lanesci<mode>"
+ [(set (match_operand:CI 0 "aarch64_simd_struct_operand" "=Utv")
+ (unspec:CI [(match_operand:CI 1 "register_operand" "w")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ UNSPEC_ST3))]
+ "TARGET_SIMD"
+{
+ if (BYTES_BIG_ENDIAN)
+ {
+ rtx tmp = gen_reg_rtx (CImode);
+ rtx mask = aarch64_reverse_mask (<MODE>mode);
+ emit_insn (gen_aarch64_rev_reglistci (tmp, operands[1], mask));
+ emit_insn (gen_aarch64_simd_st3<mode> (operands[0], tmp));
+ }
+ else
+ emit_insn (gen_aarch64_simd_st3<mode> (operands[0], operands[1]));
+ DONE;
+})
+
+(define_insn "aarch64_simd_ld4<mode>"
[(set (match_operand:XI 0 "register_operand" "=w")
(unspec:XI [(match_operand:XI 1 "aarch64_simd_struct_operand" "Utv")
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
@@ -4173,7 +4249,26 @@
[(set_attr "type" "neon_load4_one_lane")]
)
-(define_insn "vec_store_lanesxi<mode>"
+(define_expand "vec_load_lanesxi<mode>"
+ [(set (match_operand:XI 0 "register_operand" "=w")
+ (unspec:XI [(match_operand:XI 1 "aarch64_simd_struct_operand" "Utv")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ UNSPEC_LD4))]
+ "TARGET_SIMD"
+{
+ if (BYTES_BIG_ENDIAN)
+ {
+ rtx tmp = gen_reg_rtx (XImode);
+ rtx mask = aarch64_reverse_mask (<MODE>mode);
+ emit_insn (gen_aarch64_simd_ld4<mode> (tmp, operands[1]));
+ emit_insn (gen_aarch64_rev_reglistxi (operands[0], tmp, mask));
+ }
+ else
+ emit_insn (gen_aarch64_simd_ld4<mode> (operands[0], operands[1]));
+ DONE;
+})
+
+(define_insn "aarch64_simd_st4<mode>"
[(set (match_operand:XI 0 "aarch64_simd_struct_operand" "=Utv")
(unspec:XI [(match_operand:XI 1 "register_operand" "w")
(unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
@@ -4194,6 +4289,50 @@
[(set_attr "type" "neon_store4_one_lane<q>")]
)
+(define_expand "vec_store_lanesxi<mode>"
+ [(set (match_operand:XI 0 "aarch64_simd_struct_operand" "=Utv")
+ (unspec:XI [(match_operand:XI 1 "register_operand" "w")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ UNSPEC_ST4))]
+ "TARGET_SIMD"
+{
+ if (BYTES_BIG_ENDIAN)
+ {
+ rtx tmp = gen_reg_rtx (XImode);
+ rtx mask = aarch64_reverse_mask (<MODE>mode);
+ emit_insn (gen_aarch64_rev_reglistxi (tmp, operands[1], mask));
+ emit_insn (gen_aarch64_simd_st4<mode> (operands[0], tmp));
+ }
+ else
+ emit_insn (gen_aarch64_simd_st4<mode> (operands[0], operands[1]));
+ DONE;
+})
+
+(define_insn_and_split "aarch64_rev_reglist<mode>"
+[(set (match_operand:VSTRUCT 0 "register_operand" "=&w")
+ (unspec:VSTRUCT
+ [(match_operand:VSTRUCT 1 "register_operand" "w")
+ (match_operand:V16QI 2 "register_operand" "w")]
+ UNSPEC_REV_REGLIST))]
+ "TARGET_SIMD"
+ "#"
+ "&& reload_completed"
+ [(const_int 0)]
+{
+ int i;
+ int nregs = GET_MODE_SIZE (<MODE>mode) / UNITS_PER_VREG;
+ for (i = 0; i < nregs; i++)
+ {
+ rtx op0 = gen_rtx_REG (V16QImode, REGNO (operands[0]) + i);
+ rtx op1 = gen_rtx_REG (V16QImode, REGNO (operands[1]) + i);
+ emit_insn (gen_aarch64_tbl1v16qi (op0, op1, operands[2]));
+ }
+ DONE;
+}
+ [(set_attr "type" "neon_tbl1_q")
+ (set_attr "length" "<insn_count>")]
+)
+
;; Reload patterns for AdvSIMD register list operands.
(define_expand "mov<mode>"
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 0429d96..dde9690 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -8239,6 +8239,14 @@ aarch64_simd_attr_length_move (rtx_insn *insn)
return 4;
}
+/* Compute and return the length of aarch64_simd_reglist<mode>, where <mode> is
+ one of VSTRUCT modes: OI, CI, EI, or XI. */
+int
+aarch64_simd_attr_length_rglist (enum machine_mode mode)
+{
+ return (GET_MODE_SIZE (mode) / UNITS_PER_VREG) * 4;
+}
+
/* Implement target hook TARGET_VECTOR_ALIGNMENT. The AAPCS64 sets the maximum
alignment of a vector to 128 bits. */
static HOST_WIDE_INT
@@ -9767,6 +9775,27 @@ aarch64_cannot_change_mode_class (machine_mode from,
return true;
}
+rtx
+aarch64_reverse_mask (enum machine_mode mode)
+{
+ /* We have to reverse each vector because we dont have
+ a permuted load that can reverse-load according to ABI rules. */
+ rtx mask;
+ rtvec v = rtvec_alloc (16);
+ int i, j;
+ int nunits = GET_MODE_NUNITS (mode);
+ int usize = GET_MODE_UNIT_SIZE (mode);
+
+ gcc_assert (BYTES_BIG_ENDIAN);
+ gcc_assert (AARCH64_VALID_SIMD_QREG_MODE (mode));
+
+ for (i = 0; i < nunits; i++)
+ for (j = 0; j < usize; j++)
+ RTVEC_ELT (v, i * usize + j) = GEN_INT ((i + 1) * usize - 1 - j);
+ mask = gen_rtx_CONST_VECTOR (V16QImode, v);
+ return force_reg (V16QImode, mask);
+}
+
/* Implement MODES_TIEABLE_P. */
bool
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 9935167..1772157 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -298,6 +298,7 @@
UNSPEC_SHA256SU1 ; Used in aarch64-simd.md.
UNSPEC_PMULL ; Used in aarch64-simd.md.
UNSPEC_PMULL2 ; Used in aarch64-simd.md.
+ UNSPEC_REV_REGLIST ; Used in aarch64-simd.md.
])
;; -------------------------------------------------------------------
@@ -670,6 +671,8 @@
(V2DI "p") (V2DF "p")
(V2SF "p") (V4SF "v")])
+(define_mode_attr insn_count [(OI "8") (CI "12") (XI "16")])
+
;; -------------------------------------------------------------------
;; Code Iterators
;; -------------------------------------------------------------------
next reply other threads:[~2014-11-13 10:09 UTC|newest]
Thread overview: 17+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-11-13 10:11 David Sherwood [this message]
2014-11-13 14:23 ` Christophe Lyon
2014-11-13 14:33 ` David Sherwood
2014-11-17 21:18 ` Christophe Lyon
2014-11-18 9:17 ` David Sherwood
2014-11-18 10:33 ` Christophe Lyon
2014-11-27 15:03 ` David Sherwood
2014-12-11 10:16 ` David Sherwood
2014-12-11 13:46 ` Christophe Lyon
2014-12-15 9:58 ` David Sherwood
2014-12-15 13:48 ` Christophe Lyon
2014-11-20 18:41 ` Marcus Shawcroft
2014-12-17 15:23 ` Tejas Belagod
2014-12-17 16:54 ` Marcus Shawcroft
2014-12-17 17:05 ` Tejas Belagod
2015-01-25 23:07 ` Christophe Lyon
2015-08-14 15:06 ` Christophe Lyon
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to='000701cfff29$ea4fbcf0$beef36d0$@arm.com' \
--to=david.sherwood@arm.com \
--cc=gcc-patches@gcc.gnu.org \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
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).