[PATCH] D129966: [ARM] Remove VBICimm if no cleared bits are demanded

[Dave Green via Phabricator via llvm-commits]

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

If none of the bits of a VBICimm are demanded, we can remove the node entirely using the input operand instead.


https://reviews.llvm.org/D129966

Files:
  llvm/lib/Target/ARM/ARMISelLowering.cpp
  llvm/test/CodeGen/Thumb2/mve-satmul-loops.ll
  llvm/test/CodeGen/Thumb2/mve-vqmovn-combine.ll


Index: llvm/test/CodeGen/Thumb2/mve-vqmovn-combine.ll
===================================================================
--- llvm/test/CodeGen/Thumb2/mve-vqmovn-combine.ll
+++ llvm/test/CodeGen/Thumb2/mve-vqmovn-combine.ll
@@ -198,9 +198,7 @@
 define arm_aapcs_vfpcc <16 x i8> @vqmovni16_uminmax_t1(<8 x i16> %s0, <16 x i8> %src1) {
 ; CHECK-LABEL: vqmovni16_uminmax_t1:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vqmovnb.u16 q0, q0
-; CHECK-NEXT:    vmovlb.u8 q0, q0
-; CHECK-NEXT:    vmovnt.i16 q1, q0
+; CHECK-NEXT:    vqmovnt.u16 q1, q0
 ; CHECK-NEXT:    vmov q0, q1
 ; CHECK-NEXT:    bx lr
 entry:
@@ -215,7 +213,6 @@
 ; CHECK-LABEL: vqmovni16_uminmax_t2:
 ; CHECK:       @ %bb.0: @ %entry
 ; CHECK-NEXT:    vqmovnb.u16 q0, q0
-; CHECK-NEXT:    vmovlb.u8 q0, q0
 ; CHECK-NEXT:    vmovnt.i16 q0, q1
 ; CHECK-NEXT:    bx lr
 entry:
@@ -229,9 +226,7 @@
 define arm_aapcs_vfpcc <16 x i8> @vqmovni16_uminmax_b1(<8 x i16> %s0, <16 x i8> %src1) {
 ; CHECK-LABEL: vqmovni16_uminmax_b1:
 ; CHECK:       @ %bb.0: @ %entry
-; CHECK-NEXT:    vqmovnb.u16 q0, q0
-; CHECK-NEXT:    vmovlb.u8 q0, q0
-; CHECK-NEXT:    vmovnb.i16 q1, q0
+; CHECK-NEXT:    vqmovnb.u16 q1, q0
 ; CHECK-NEXT:    vmov q0, q1
 ; CHECK-NEXT:    bx lr
 entry:
Index: llvm/test/CodeGen/Thumb2/mve-satmul-loops.ll
===================================================================
--- llvm/test/CodeGen/Thumb2/mve-satmul-loops.ll
+++ llvm/test/CodeGen/Thumb2/mve-satmul-loops.ll
@@ -3099,11 +3099,8 @@
 ; CHECK-NEXT:    vldrb.u8 q1, [r1], #16
 ; CHECK-NEXT:    vmullt.u8 q2, q1, q0
 ; CHECK-NEXT:    vmullb.u8 q0, q1, q0
-; CHECK-NEXT:    vqshrnb.u16 q2, q2, #7
 ; CHECK-NEXT:    vqshrnb.u16 q0, q0, #7
-; CHECK-NEXT:    vmovlb.u8 q2, q2
-; CHECK-NEXT:    vmovlb.u8 q0, q0
-; CHECK-NEXT:    vmovnt.i16 q0, q2
+; CHECK-NEXT:    vqshrnt.u16 q0, q2, #7
 ; CHECK-NEXT:    vstrb.8 q0, [r2], #16
 ; CHECK-NEXT:    le lr, .LBB21_4
 ; CHECK-NEXT:  @ %bb.5: @ %middle.block
Index: llvm/lib/Target/ARM/ARMISelLowering.cpp
===================================================================
--- llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -19962,6 +19962,14 @@
     }
     break;
   }
+  case ARMISD::VBICIMM: {
+    SDValue Op0 = Op.getOperand(0);
+    unsigned ModImm = Op.getConstantOperandVal(1);
+    unsigned EltBits = 0;
+    uint64_t Mask = ARM_AM::decodeVMOVModImm(ModImm, EltBits);
+    if ((OriginalDemandedBits & Mask) == 0)
+      return TLO.CombineTo(Op, Op0);
+  }
   }
 
   return TargetLowering::SimplifyDemandedBitsForTargetNode(


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D129966.445324.patch
Type: text/x-patch
Size: 2553 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220717/1c3fd998/attachment.bin>