+(define_expand "aarch64_ld2_lane<VQ:mode>"
+ [(match_operand:OI 0 "register_operand" "=w")
+ (match_operand:DI 1 "register_operand" "w")
+ (match_operand:OI 2 "register_operand" "0")
+ (match_operand:SI 3 "immediate_operand" "i")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ "TARGET_SIMD"
+{
+ enum machine_mode mode = <V_TWO_ELEM>mode;
+ rtx mem = gen_rtx_MEM (mode, operands[1]);
+ operands[3] = GEN_INT (ENDIAN_LANE_N (<MODE>mode, INTVAL (operands[3])));
+
The endianess lane correction breaks this for BE.
You don't need the endianess lane correction here - we always call neon
intrinsics with the architectural lane number - irrespective of
endianness. Unless ofcourse you flip it somewhere to make it a part of
RTL vec_select lane patterns, which you don't here.
You could also do some lane-bounds checking here in the expander.
+ emit_insn (gen_vec_load_lanesoi_lane<VQ:mode> (operands[0],
+ mem,
+ operands[2],
+ operands[3]));
+ DONE;
+})
+
+(define_expand "aarch64_ld3_lane<VQ:mode>"
+ [(match_operand:CI 0 "register_operand" "=w")
+ (match_operand:DI 1 "register_operand" "w")
+ (match_operand:CI 2 "register_operand" "0")
+ (match_operand:SI 3 "immediate_operand" "i")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ "TARGET_SIMD"
+{
+ enum machine_mode mode = <V_THREE_ELEM>mode;
+ rtx mem = gen_rtx_MEM (mode, operands[1]);
+ operands[3] = GEN_INT (ENDIAN_LANE_N (<MODE>mode, INTVAL (operands[3])));
+
No endianness correction for lanes necessary.
+ emit_insn (gen_vec_load_lanesci_lane<VQ:mode> (operands[0],
+ mem,
+ operands[2],
+ operands[3]));
+ DONE;
+})
+
+(define_expand "aarch64_ld4_lane<VQ:mode>"
+ [(match_operand:XI 0 "register_operand" "=w")
+ (match_operand:DI 1 "register_operand" "w")
+ (match_operand:XI 2 "register_operand" "0")
+ (match_operand:SI 3 "immediate_operand" "i")
+ (unspec:VQ [(const_int 0)] UNSPEC_VSTRUCTDUMMY)]
+ "TARGET_SIMD"
+{
+ enum machine_mode mode = <V_FOUR_ELEM>mode;
+ rtx mem = gen_rtx_MEM (mode, operands[1]);
+ operands[3] = GEN_INT (ENDIAN_LANE_N (<MODE>mode, INTVAL (operands[3])));
+
Same.
+ emit_insn (gen_vec_load_lanesxi_lane<VQ:mode> (operands[0],
+ mem,
+ operands[2],
+ operands[3]));
+ DONE;
+})
+
+
+
;; Expanders for builtins to extract vector registers from large
;; opaque integer modes.
diff --git a/gcc/config/aarch64/aarch64.md b/gcc/config/aarch64/aarch64.md
index c60038a..ea924ab 100644
--- a/gcc/config/aarch64/aarch64.md
+++ b/gcc/config/aarch64/aarch64.md
@@ -92,6 +92,9 @@
UNSPEC_LD2
UNSPEC_LD3
UNSPEC_LD4
+ UNSPEC_LD2_LANE
+ UNSPEC_LD3_LANE
+ UNSPEC_LD4_LANE
UNSPEC_MB
UNSPEC_NOP
UNSPEC_PRLG_STK
Thanks,
Tejas.