Hi all,
This fixes a bug in the vector load functions in which they load the
vector in the wrong order for big endian systems. This patch flips the
order conditionally in the vec_concats.
No testcase given because plenty of existing tests for vld functions.
Ran regression tests on aarch64_be-none-elf and aarch64-none-elf.
Vldx tests now pass on aarch64_be-none-elf and no regressions on both.
Ok for trunk?
I do not have commit rights so if ok can someone apply it for me?
Thanks,
Tamar
gcc/
2016-08-16 Tamar Christina <tamar.christ...@arm.com>
* gcc/config/aarch64/aarch64-simd.md
(aarch64_ld2<mode>_dreg_le): New.
(aarch64_ld2<mode>_dreg_be): New.
(aarch64_ld2<mode>_dreg): Removed.
(aarch64_ld3<mode>_dreg_le): New.
(aarch64_ld3<mode>_dreg_be): New.
(aarch64_ld3<mode>_dreg): Removed.
(aarch64_ld4<mode>_dreg_le): New.
(aarch64_ld4<mode>_dreg_be): New.
(aarch64_ld4<mode>_dreg): Removed.
(aarch64_ld<VSTRUCT:nregs><VDC:mode>): Wrapper around _le, _be.
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index feb5e96..7d3dfe8 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -4861,7 +4861,7 @@
DONE;
})
-(define_insn "aarch64_ld2<mode>_dreg"
+(define_insn "aarch64_ld2<mode>_dreg_le"
[(set (match_operand:OI 0 "register_operand" "=w")
(subreg:OI
(vec_concat:<VRL2>
@@ -4874,12 +4874,30 @@
(unspec:VD [(match_dup 1)]
UNSPEC_LD2)
(vec_duplicate:VD (const_int 0)))) 0))]
- "TARGET_SIMD"
+ "TARGET_SIMD && !BYTES_BIG_ENDIAN"
"ld2\\t{%S0.<Vtype> - %T0.<Vtype>}, %1"
[(set_attr "type" "neon_load2_2reg<q>")]
)
-(define_insn "aarch64_ld2<mode>_dreg"
+(define_insn "aarch64_ld2<mode>_dreg_be"
+ [(set (match_operand:OI 0 "register_operand" "=w")
+ (subreg:OI
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD
+ [(match_operand:BLK 1 "aarch64_simd_struct_operand" "Utv")]
+ UNSPEC_LD2))
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD [(match_dup 1)]
+ UNSPEC_LD2))) 0))]
+ "TARGET_SIMD && BYTES_BIG_ENDIAN"
+ "ld2\\t{%S0.<Vtype> - %T0.<Vtype>}, %1"
+ [(set_attr "type" "neon_load2_2reg<q>")]
+)
+
+(define_insn "aarch64_ld2<mode>_dreg_le"
[(set (match_operand:OI 0 "register_operand" "=w")
(subreg:OI
(vec_concat:<VRL2>
@@ -4892,12 +4910,30 @@
(unspec:DX [(match_dup 1)]
UNSPEC_LD2)
(const_int 0))) 0))]
- "TARGET_SIMD"
+ "TARGET_SIMD && !BYTES_BIG_ENDIAN"
+ "ld1\\t{%S0.1d - %T0.1d}, %1"
+ [(set_attr "type" "neon_load1_2reg<q>")]
+)
+
+(define_insn "aarch64_ld2<mode>_dreg_be"
+ [(set (match_operand:OI 0 "register_operand" "=w")
+ (subreg:OI
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX
+ [(match_operand:BLK 1 "aarch64_simd_struct_operand" "Utv")]
+ UNSPEC_LD2))
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX [(match_dup 1)]
+ UNSPEC_LD2))) 0))]
+ "TARGET_SIMD && BYTES_BIG_ENDIAN"
"ld1\\t{%S0.1d - %T0.1d}, %1"
[(set_attr "type" "neon_load1_2reg<q>")]
)
-(define_insn "aarch64_ld3<mode>_dreg"
+(define_insn "aarch64_ld3<mode>_dreg_le"
[(set (match_operand:CI 0 "register_operand" "=w")
(subreg:CI
(vec_concat:<VRL3>
@@ -4915,12 +4951,35 @@
(unspec:VD [(match_dup 1)]
UNSPEC_LD3)
(vec_duplicate:VD (const_int 0)))) 0))]
- "TARGET_SIMD"
+ "TARGET_SIMD && !BYTES_BIG_ENDIAN"
"ld3\\t{%S0.<Vtype> - %U0.<Vtype>}, %1"
[(set_attr "type" "neon_load3_3reg<q>")]
)
-(define_insn "aarch64_ld3<mode>_dreg"
+(define_insn "aarch64_ld3<mode>_dreg_be"
+ [(set (match_operand:CI 0 "register_operand" "=w")
+ (subreg:CI
+ (vec_concat:<VRL3>
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD
+ [(match_operand:BLK 1 "aarch64_simd_struct_operand" "Utv")]
+ UNSPEC_LD3))
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD [(match_dup 1)]
+ UNSPEC_LD3)))
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD [(match_dup 1)]
+ UNSPEC_LD3))) 0))]
+ "TARGET_SIMD && BYTES_BIG_ENDIAN"
+ "ld3\\t{%S0.<Vtype> - %U0.<Vtype>}, %1"
+ [(set_attr "type" "neon_load3_3reg<q>")]
+)
+
+(define_insn "aarch64_ld3<mode>_dreg_le"
[(set (match_operand:CI 0 "register_operand" "=w")
(subreg:CI
(vec_concat:<VRL3>
@@ -4938,12 +4997,35 @@
(unspec:DX [(match_dup 1)]
UNSPEC_LD3)
(const_int 0))) 0))]
- "TARGET_SIMD"
+ "TARGET_SIMD && !BYTES_BIG_ENDIAN"
+ "ld1\\t{%S0.1d - %U0.1d}, %1"
+ [(set_attr "type" "neon_load1_3reg<q>")]
+)
+
+(define_insn "aarch64_ld3<mode>_dreg_be"
+ [(set (match_operand:CI 0 "register_operand" "=w")
+ (subreg:CI
+ (vec_concat:<VRL3>
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX
+ [(match_operand:BLK 1 "aarch64_simd_struct_operand" "Utv")]
+ UNSPEC_LD3))
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX [(match_dup 1)]
+ UNSPEC_LD3)))
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX [(match_dup 1)]
+ UNSPEC_LD3))) 0))]
+ "TARGET_SIMD && BYTES_BIG_ENDIAN"
"ld1\\t{%S0.1d - %U0.1d}, %1"
[(set_attr "type" "neon_load1_3reg<q>")]
)
-(define_insn "aarch64_ld4<mode>_dreg"
+(define_insn "aarch64_ld4<mode>_dreg_le"
[(set (match_operand:XI 0 "register_operand" "=w")
(subreg:XI
(vec_concat:<VRL4>
@@ -4954,9 +5036,9 @@
UNSPEC_LD4)
(vec_duplicate:VD (const_int 0)))
(vec_concat:<VDBL>
- (unspec:VD [(match_dup 1)]
+ (unspec:VD [(match_dup 1)]
UNSPEC_LD4)
- (vec_duplicate:VD (const_int 0))))
+ (vec_duplicate:VD (const_int 0))))
(vec_concat:<VRL2>
(vec_concat:<VDBL>
(unspec:VD [(match_dup 1)]
@@ -4966,12 +5048,40 @@
(unspec:VD [(match_dup 1)]
UNSPEC_LD4)
(vec_duplicate:VD (const_int 0))))) 0))]
- "TARGET_SIMD"
+ "TARGET_SIMD && !BYTES_BIG_ENDIAN"
+ "ld4\\t{%S0.<Vtype> - %V0.<Vtype>}, %1"
+ [(set_attr "type" "neon_load4_4reg<q>")]
+)
+
+(define_insn "aarch64_ld4<mode>_dreg_be"
+ [(set (match_operand:XI 0 "register_operand" "=w")
+ (subreg:XI
+ (vec_concat:<VRL4>
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD
+ [(match_operand:BLK 1 "aarch64_simd_struct_operand" "Utv")]
+ UNSPEC_LD4))
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD [(match_dup 1)]
+ UNSPEC_LD4)))
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD [(match_dup 1)]
+ UNSPEC_LD4))
+ (vec_concat:<VDBL>
+ (vec_duplicate:VD (const_int 0))
+ (unspec:VD [(match_dup 1)]
+ UNSPEC_LD4)))) 0))]
+ "TARGET_SIMD && BYTES_BIG_ENDIAN"
"ld4\\t{%S0.<Vtype> - %V0.<Vtype>}, %1"
[(set_attr "type" "neon_load4_4reg<q>")]
)
-(define_insn "aarch64_ld4<mode>_dreg"
+(define_insn "aarch64_ld4<mode>_dreg_le"
[(set (match_operand:XI 0 "register_operand" "=w")
(subreg:XI
(vec_concat:<VRL4>
@@ -4984,7 +5094,7 @@
(vec_concat:<VDBL>
(unspec:DX [(match_dup 1)]
UNSPEC_LD4)
- (const_int 0)))
+ (const_int 0)))
(vec_concat:<VRL2>
(vec_concat:<VDBL>
(unspec:DX [(match_dup 1)]
@@ -4994,7 +5104,35 @@
(unspec:DX [(match_dup 1)]
UNSPEC_LD4)
(const_int 0)))) 0))]
- "TARGET_SIMD"
+ "TARGET_SIMD && !BYTES_BIG_ENDIAN"
+ "ld1\\t{%S0.1d - %V0.1d}, %1"
+ [(set_attr "type" "neon_load1_4reg<q>")]
+)
+
+(define_insn "aarch64_ld4<mode>_dreg_be"
+ [(set (match_operand:XI 0 "register_operand" "=w")
+ (subreg:XI
+ (vec_concat:<VRL4>
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX
+ [(match_operand:BLK 1 "aarch64_simd_struct_operand" "Utv")]
+ UNSPEC_LD4))
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX [(match_dup 1)]
+ UNSPEC_LD4)))
+ (vec_concat:<VRL2>
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX [(match_dup 1)]
+ UNSPEC_LD4))
+ (vec_concat:<VDBL>
+ (const_int 0)
+ (unspec:DX [(match_dup 1)]
+ UNSPEC_LD4)))) 0))]
+ "TARGET_SIMD && BYTES_BIG_ENDIAN"
"ld1\\t{%S0.1d - %V0.1d}, %1"
[(set_attr "type" "neon_load1_4reg<q>")]
)
@@ -5008,7 +5146,10 @@
rtx mem = gen_rtx_MEM (BLKmode, operands[1]);
set_mem_size (mem, <VSTRUCT:nregs> * 8);
- emit_insn (gen_aarch64_ld<VSTRUCT:nregs><VDC:mode>_dreg (operands[0], mem));
+ if (BYTES_BIG_ENDIAN)
+ emit_insn (gen_aarch64_ld<VSTRUCT:nregs><VDC:mode>_dreg_be (operands[0], mem));
+ else
+ emit_insn (gen_aarch64_ld<VSTRUCT:nregs><VDC:mode>_dreg_le (operands[0], mem));
DONE;
})