[AArch64][GlobalISel] Legalize bswap <2 x i16>

Differential revision: https://reviews.llvm.org/D105935

Author: jroelofs

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

@@ -103,7 +103,8 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
   getActionDefinitionsBuilder(G_BSWAP)
       .legalFor({s32, s64, v4s32, v2s32, v2s64})
       .clampScalar(0, s32, s64)
-      .widenScalarToNextPow2(0);
+      .widenScalarToNextPow2(0)
+      .customIf(typeIs(0, v2s16)); // custom lower as G_REV32 + G_LSHR
 
   getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR})
       .legalFor({s32, s64, v2s32, v4s32, v4s16, v8s16, v16s8, v8s8})
@@ -791,6 +792,8 @@ bool AArch64LegalizerInfo::legalizeCustom(LegalizerHelper &Helper,
   case TargetOpcode::G_LOAD:
   case TargetOpcode::G_STORE:
     return legalizeLoadStore(MI, MRI, MIRBuilder, Observer);
+  case TargetOpcode::G_BSWAP:
+    return legalizeBSwap(MI, MRI, MIRBuilder);
   case TargetOpcode::G_SHL:
   case TargetOpcode::G_ASHR:
   case TargetOpcode::G_LSHR:
@@ -1001,6 +1004,46 @@ bool AArch64LegalizerInfo::legalizeLoadStore(
   return true;
 }
 
+bool AArch64LegalizerInfo::legalizeBSwap(MachineInstr &MI,
+                                         MachineRegisterInfo &MRI,
+                                         MachineIRBuilder &MIRBuilder) const {
+  assert(MI.getOpcode() == TargetOpcode::G_BSWAP);
+
+  // The <2 x half> case needs special lowering because there isn't an
+  // instruction that does that directly. Instead, we widen to <8 x i8>
+  // and emit a G_REV32 followed by a G_LSHR knowing that instruction selection
+  // will later match them as:
+  //
+  //   rev32.8b v0, v0
+  //   ushr.2s v0, v0, #16
+  //
+  // We could emit those here directly, but it seems better to keep things as
+  // generic as possible through legalization, and avoid committing layering
+  // violations by legalizing & selecting here at the same time.
+
+  Register ValReg = MI.getOperand(1).getReg();
+  assert(LLT::fixed_vector(2, 16) == MRI.getType(ValReg));
+  const LLT v2s32 = LLT::fixed_vector(2, 32);
+  const LLT v8s8 = LLT::fixed_vector(8, 8);
+  const LLT s32 = LLT::scalar(32);
+
+  auto Undef = MIRBuilder.buildUndef(v8s8);
+  auto Insert =
+      MIRBuilder
+          .buildInstr(TargetOpcode::INSERT_SUBREG, {v8s8}, {Undef, ValReg})
+          .addImm(AArch64::ssub);
+  auto Rev32 = MIRBuilder.buildInstr(AArch64::G_REV32, {v8s8}, {Insert});
+  auto Bitcast = MIRBuilder.buildBitcast(v2s32, Rev32);
+  auto Amt = MIRBuilder.buildConstant(v2s32, 16);
+  auto UShr =
+      MIRBuilder.buildInstr(TargetOpcode::G_LSHR, {v2s32}, {Bitcast, Amt});
+  auto Zero = MIRBuilder.buildConstant(s32, 0);
+  auto Extract = MIRBuilder.buildExtractVectorElement(s32, UShr, Zero);
+  MIRBuilder.buildBitcast({MI.getOperand(0).getReg()}, Extract);
+  MI.eraseFromParent();
+  return true;
+}
+
 bool AArch64LegalizerInfo::legalizeVaArg(MachineInstr &MI,
                                          MachineRegisterInfo &MRI,
                                          MachineIRBuilder &MIRBuilder) const {

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.h

@@ -35,6 +35,8 @@ class AArch64LegalizerInfo : public LegalizerInfo {
                          MachineInstr &MI) const override;
 
 private:
+  bool legalizeBSwap(MachineInstr &MI, MachineRegisterInfo &MRI,
+                     MachineIRBuilder &MIRBuilder) const;
   bool legalizeVaArg(MachineInstr &MI, MachineRegisterInfo &MRI,
                      MachineIRBuilder &MIRBuilder) const;
   bool legalizeLoadStore(MachineInstr &MI, MachineRegisterInfo &MRI,

llvm/test/CodeGen/AArch64/GlobalISel/legalize-bswap.mir

@@ -8,6 +8,8 @@
 
   define i16 @bswap_s16(i16 %a) { ret i16 0 }
 
+  define <2 x i16> @bswap_2xi16(<2 x i16> %a) { ret <2 x i16> <i16 0, i16 0> }
+
   attributes #0 = { nounwind readnone speculatable willreturn }
 
 ...
@@ -42,3 +44,40 @@ body:             |
     RET_ReallyLR implicit $w0
 
 ...
+---
+name:            bswap_2xi16
+alignment:       4
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: _ }
+  - { id: 1, class: _ }
+liveins:
+  - { reg: '$s0' }
+frameInfo:
+  maxAlignment:    1
+machineFunctionInfo: {}
+body:             |
+  bb.1:
+    liveins: $s0
+
+    ; CHECK-LABEL: name: bswap_2xi16
+    ; CHECK: liveins: $s0
+    ; CHECK: [[COPY:%[0-9]+]]:_(<2 x s16>) = COPY $s0
+    ; CHECK: [[DEF:%[0-9]+]]:_(<8 x s8>) = G_IMPLICIT_DEF
+    ; CHECK: [[INSERT_SUBREG:%[0-9]+]]:_(<8 x s8>) = INSERT_SUBREG [[DEF]](<8 x s8>), [[COPY]](<2 x s16>), %subreg.ssub
+    ; CHECK: [[REV32_:%[0-9]+]]:_(<8 x s8>) = G_REV32 [[INSERT_SUBREG]]
+    ; CHECK: [[BITCAST:%[0-9]+]]:_(<2 x s32>) = G_BITCAST [[REV32_]](<8 x s8>)
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[C]](s32), [[C]](s32)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(<2 x s32>) = G_LSHR [[BITCAST]], [[BUILD_VECTOR]](<2 x s32>)
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 0
+    ; CHECK: [[EVEC:%[0-9]+]]:_(s32) = G_EXTRACT_VECTOR_ELT [[LSHR]](<2 x s32>), [[C1]](s64)
+    ; CHECK: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[EVEC]](s32)
+    ; CHECK: $s0 = COPY [[BITCAST1]](<2 x s16>)
+    ; CHECK: RET_ReallyLR
+    %0:_(<2 x s16>) = COPY $s0
+    %1:_(<2 x s16>) = G_BSWAP %0
+    $s0 = COPY %1(<2 x s16>)
+    RET_ReallyLR
+
+...

llvm/test/CodeGen/AArch64/GlobalISel/legalizer-info-validation.mir

@@ -555,8 +555,8 @@
 # DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_BSWAP (opcode {{[0-9]+}}): 1 type index, 0 imm indices
-# DEBUG-NEXT: .. the first uncovered type index: 1, OK
-# DEBUG-NEXT: .. the first uncovered imm index: 0, OK
+# DEBUG-NEXT: .. type index coverage check SKIPPED: user-defined predicate detected
+# DEBUG-NEXT: .. imm index coverage check SKIPPED: user-defined predicate detected
 # DEBUG-NEXT: G_BITREVERSE (opcode {{[0-9]+}}): 1 type index, 0 imm indices
 # DEBUG-NEXT: .. the first uncovered type index: 1, OK
 # DEBUG-NEXT: .. the first uncovered imm index: 0, OK