[PATCH] D97291: [ARM] Handle Splats in MVE lane interleaving

[Dave Green via Phabricator via llvm-commits]

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

As another addition to MVE lane interleaving, this handles Splat shuffle vectors, as the shuffle of a splat is a splat.


https://reviews.llvm.org/D97291

Files:
  llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
  llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll


Index: llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll
===================================================================
--- llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll
+++ llvm/test/CodeGen/Thumb2/mve-laneinterleaving-cost.ll
@@ -422,17 +422,15 @@
 define arm_aapcs_vfpcc void @mul_i16(<8 x i16> *%A, <8 x i16> *%B, i32 %C, <8 x i16> *%D) {
 ; CHECK-LABEL: mul_i16:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vldrh.s32 q0, [r1]
-; CHECK-NEXT:    vldrh.s32 q1, [r0]
-; CHECK-NEXT:    vldrh.s32 q2, [r0, #8]
+; CHECK-NEXT:    vldrw.u32 q0, [r1]
+; CHECK-NEXT:    vldrw.u32 q1, [r0]
 ; CHECK-NEXT:    rsbs r2, r2, #0
-; CHECK-NEXT:    vmul.i32 q0, q1, q0
-; CHECK-NEXT:    vldrh.s32 q1, [r1, #8]
+; CHECK-NEXT:    vmullt.s16 q2, q1, q0
+; CHECK-NEXT:    vmullb.s16 q0, q1, q0
+; CHECK-NEXT:    vshl.s32 q2, r2
 ; CHECK-NEXT:    vshl.s32 q0, r2
-; CHECK-NEXT:    vmul.i32 q1, q2, q1
-; CHECK-NEXT:    vstrh.32 q0, [r3]
-; CHECK-NEXT:    vshl.s32 q1, r2
-; CHECK-NEXT:    vstrh.32 q1, [r3, #8]
+; CHECK-NEXT:    vmovnt.i32 q0, q2
+; CHECK-NEXT:    vstrw.32 q0, [r3]
 ; CHECK-NEXT:    bx lr
 entry:
   %a = load <8 x i16>, <8 x i16> *%A, align 4
@@ -451,17 +449,15 @@
 define arm_aapcs_vfpcc void @mul_i8(<16 x i8> *%A, <16 x i8> *%B, i16 %C, <16 x i8> *%D) {
 ; CHECK-LABEL: mul_i8:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vldrb.s16 q0, [r1]
-; CHECK-NEXT:    vldrb.s16 q1, [r0]
-; CHECK-NEXT:    vldrb.s16 q2, [r0, #8]
+; CHECK-NEXT:    vldrw.u32 q0, [r1]
+; CHECK-NEXT:    vldrw.u32 q1, [r0]
 ; CHECK-NEXT:    rsbs r2, r2, #0
-; CHECK-NEXT:    vmul.i16 q0, q1, q0
-; CHECK-NEXT:    vldrb.s16 q1, [r1, #8]
+; CHECK-NEXT:    vmullt.s8 q2, q1, q0
+; CHECK-NEXT:    vmullb.s8 q0, q1, q0
+; CHECK-NEXT:    vshl.s16 q2, r2
 ; CHECK-NEXT:    vshl.s16 q0, r2
-; CHECK-NEXT:    vmul.i16 q1, q2, q1
-; CHECK-NEXT:    vstrb.16 q0, [r3]
-; CHECK-NEXT:    vshl.s16 q1, r2
-; CHECK-NEXT:    vstrb.16 q1, [r3, #8]
+; CHECK-NEXT:    vmovnt.i16 q0, q2
+; CHECK-NEXT:    vstrw.32 q0, [r3]
 ; CHECK-NEXT:    bx lr
 entry:
   %a = load <16 x i8>, <16 x i8> *%A, align 4
Index: llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
===================================================================
--- llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
+++ llvm/lib/Target/ARM/MVELaneInterleavingPass.cpp
@@ -214,6 +214,12 @@
         Worklist.push_back(cast<Instruction>(Use));
       break;
 
+    case Instruction::ShuffleVector:
+      // A shuffle of a splat is a splat.
+      if (cast<ShuffleVectorInst>(I)->isZeroEltSplat())
+        continue;
+      LLVM_FALLTHROUGH;
+
     default:
       LLVM_DEBUG(dbgs() << "  Unhandled instruction: " << *I << "\n");
       return false;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D97291.325784.patch
Type: text/x-patch
Size: 2695 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210223/5d7c0197/attachment.bin>