[PATCH] D95427: [ARM] Add neon FP16 scalar_to_vector patterns.

[Dave Green via Phabricator via llvm-commits]

dmgreen created this revision.
dmgreen added reviewers: SjoerdMeijer, miyuki, NickGuy, simon_tatham, ostannard, efriedma.
Herald added subscribers: danielkiss, hiraditya, kristof.beyls.
dmgreen requested review of this revision.
Herald added a project: LLVM.

This adds some simple fp16 scalar_to_vector patterns, preventing a selection failure if this came up.


https://reviews.llvm.org/D95427

Files:
  llvm/lib/Target/ARM/ARMInstrNEON.td
  llvm/test/CodeGen/ARM/fp16-insert-extract.ll


Index: llvm/test/CodeGen/ARM/fp16-insert-extract.ll
===================================================================
--- llvm/test/CodeGen/ARM/fp16-insert-extract.ll
+++ llvm/test/CodeGen/ARM/fp16-insert-extract.ll
@@ -74,6 +74,39 @@
   ret float %conv
 }
 
+define <4 x half> @insert_v4f16(half %a) {
+; CHECKHARD-LABEL: insert_v4f16:
+; CHECKHARD:       @ %bb.0: @ %entry
+; CHECKHARD-NEXT:    @ kill: def $s0 killed $s0 def $d0
+; CHECKHARD-NEXT:    bx lr
+;
+; CHECKSOFT-LABEL: insert_v4f16:
+; CHECKSOFT:       @ %bb.0: @ %entry
+; CHECKSOFT-NEXT:    vmov.f16 s0, r0
+; CHECKSOFT-NEXT:    vmov r0, r1, d0
+; CHECKSOFT-NEXT:    bx lr
+entry:
+  %res = insertelement <4 x half> undef, half %a, i32 0
+  ret <4 x half> %res
+}
+
+define <8 x half> @insert_v8f16(half %a) {
+; CHECKHARD-LABEL: insert_v8f16:
+; CHECKHARD:       @ %bb.0: @ %entry
+; CHECKHARD-NEXT:    @ kill: def $s0 killed $s0 def $q0
+; CHECKHARD-NEXT:    bx lr
+;
+; CHECKSOFT-LABEL: insert_v8f16:
+; CHECKSOFT:       @ %bb.0: @ %entry
+; CHECKSOFT-NEXT:    vmov.f16 s0, r0
+; CHECKSOFT-NEXT:    vmov r2, r3, d1
+; CHECKSOFT-NEXT:    vmov r0, r1, d0
+; CHECKSOFT-NEXT:    bx lr
+entry:
+  %res = insertelement <8 x half> undef, half %a, i32 0
+  ret <8 x half> %res
+}
+
 define <4 x half> @test_vset_lane_f16(<4 x half> %a, float %fb) nounwind {
 ; CHECKHARD-LABEL: test_vset_lane_f16:
 ; CHECKHARD:       @ %bb.0: @ %entry
Index: llvm/lib/Target/ARM/ARMInstrNEON.td
===================================================================
--- llvm/lib/Target/ARM/ARMInstrNEON.td
+++ llvm/lib/Target/ARM/ARMInstrNEON.td
@@ -6494,6 +6494,11 @@
 def : Pat<(v4f32 (scalar_to_vector SPR:$src)),
           (INSERT_SUBREG (v4f32 (IMPLICIT_DEF)), SPR:$src, ssub_0)>;
 
+def : Pat<(v4f16 (scalar_to_vector (f16 HPR:$src))),
+          (INSERT_SUBREG (v4f16 (IMPLICIT_DEF)), HPR:$src, ssub_0)>;
+def : Pat<(v8f16 (scalar_to_vector (f16 HPR:$src))),
+          (INSERT_SUBREG (v8f16 (IMPLICIT_DEF)), HPR:$src, ssub_0)>;
+
 def : Pat<(v8i8 (scalar_to_vector GPR:$src)),
           (VSETLNi8  (v8i8  (IMPLICIT_DEF)), GPR:$src, (i32 0))>;
 def : Pat<(v4i16 (scalar_to_vector GPR:$src)),


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D95427.319238.patch
Type: text/x-patch
Size: 2145 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210126/5ece5edb/attachment.bin>