PR 57431 is an ICE when compiling for Neon intrinsics. For completeness
vld1_dup_u64 is defined to fill a single element vector from a single
element of data, but this is in reality degenerate.
Unfortunately, by expanding this into a vec_duplicate operation we
confuse the mid-end of the compiler, since it expects a vector mode to
result from the duplicate operation.
The fix is to special-case the expand into the appropriate direct load
operation and to then simplify the logic for the remaining cases (which
now really are all duplicate operations).
PR target/57431
* arm/neon.md (neon_vld1_dupdi): New expand pattern.
(neon_vld1_dup<mode> VD iterator): Iterate over VD not VDX.
R.--- neon.md (revision 201427)
+++ neon.md (local)
@@ -4593,19 +4593,20 @@ (define_insn "neon_vld1_lane<mode>"
)
(define_insn "neon_vld1_dup<mode>"
- [(set (match_operand:VDX 0 "s_register_operand" "=w")
- (vec_duplicate:VDX (match_operand:<V_elem> 1 "neon_struct_operand"
"Um")))]
+ [(set (match_operand:VD 0 "s_register_operand" "=w")
+ (vec_duplicate:VD (match_operand:<V_elem> 1 "neon_struct_operand"
"Um")))]
"TARGET_NEON"
-{
- if (GET_MODE_NUNITS (<MODE>mode) > 1)
- return "vld1.<V_sz_elem>\t{%P0[]}, %A1";
- else
- return "vld1.<V_sz_elem>\t%h0, %A1";
-}
- [(set (attr "neon_type")
- (if_then_else (gt (const_string "<V_mode_nunits>") (const_string "1"))
- (const_string "neon_vld2_2_regs_vld1_vld2_all_lanes")
- (const_string "neon_vld1_1_2_regs")))]
+ "vld1.<V_sz_elem>\t{%P0[]}, %A1"
+ [(set_attr "neon_type" "neon_vld2_2_regs_vld1_vld2_all_lanes")]
+)
+
+;; Special case for DImode. Treat it exactly like a simple load.
+(define_expand "neon_vld1_dupdi"
+ [(set (match_operand:DI 0 "s_register_operand" "")
+ (unspec:DI [(match_operand:DI 1 "neon_struct_operand" "")]
+ UNSPEC_VLD1))]
+ "TARGET_NEON"
+ ""
)
(define_insn "neon_vld1_dup<mode>"
