[AArch64] Combine store (trunc X to <3 x i8>) to sequence of ST1.b. (#78637)

Improve codegen for (trunc X to <3 x i8>) by converting it to a sequence
of 3 ST1.b, but first converting the truncate operand to either v8i8 or
v16i8, extracting the lanes for the truncate results and storing them.

At the moment, there are almost no cases in which such vector operations
will be generated automatically. The motivating case is non-power-of-2
SLP vectorization: #77790

PR: #78637

Author: fhahn

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -21471,6 +21471,53 @@ bool isHalvingTruncateOfLegalScalableType(EVT SrcVT, EVT DstVT) {
          (SrcVT == MVT::nxv2i64 && DstVT == MVT::nxv2i32);
 }
 
+// Combine store (trunc X to <3 x i8>) to sequence of ST1.b.
+static SDValue combineI8TruncStore(StoreSDNode *ST, SelectionDAG &DAG,
+                                   const AArch64Subtarget *Subtarget) {
+  SDValue Value = ST->getValue();
+  EVT ValueVT = Value.getValueType();
+
+  if (ST->isVolatile() || !Subtarget->isLittleEndian() ||
+      Value.getOpcode() != ISD::TRUNCATE ||
+      ValueVT != EVT::getVectorVT(*DAG.getContext(), MVT::i8, 3))
+    return SDValue();
+
+  assert(ST->getOffset().isUndef() && "undef offset expected");
+  SDLoc DL(ST);
+  auto WideVT = EVT::getVectorVT(
+      *DAG.getContext(),
+      Value->getOperand(0).getValueType().getVectorElementType(), 4);
+  SDValue UndefVector = DAG.getUNDEF(WideVT);
+  SDValue WideTrunc = DAG.getNode(
+      ISD::INSERT_SUBVECTOR, DL, WideVT,
+      {UndefVector, Value->getOperand(0), DAG.getVectorIdxConstant(0, DL)});
+  SDValue Cast = DAG.getNode(
+      ISD::BITCAST, DL, WideVT.getSizeInBits() == 64 ? MVT::v8i8 : MVT::v16i8,
+      WideTrunc);
+
+  MachineFunction &MF = DAG.getMachineFunction();
+  SDValue Chain = ST->getChain();
+  MachineMemOperand *MMO = ST->getMemOperand();
+  unsigned IdxScale = WideVT.getScalarSizeInBits() / 8;
+  SDValue E2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8, Cast,
+                           DAG.getConstant(2 * IdxScale, DL, MVT::i64));
+  TypeSize Offset2 = TypeSize::getFixed(2);
+  SDValue Ptr2 = DAG.getMemBasePlusOffset(ST->getBasePtr(), Offset2, DL);
+  Chain = DAG.getStore(Chain, DL, E2, Ptr2, MF.getMachineMemOperand(MMO, 2, 1));
+
+  SDValue E1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8, Cast,
+                           DAG.getConstant(1 * IdxScale, DL, MVT::i64));
+  TypeSize Offset1 = TypeSize::getFixed(1);
+  SDValue Ptr1 = DAG.getMemBasePlusOffset(ST->getBasePtr(), Offset1, DL);
+  Chain = DAG.getStore(Chain, DL, E1, Ptr1, MF.getMachineMemOperand(MMO, 1, 1));
+
+  SDValue E0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8, Cast,
+                           DAG.getConstant(0, DL, MVT::i64));
+  Chain = DAG.getStore(Chain, DL, E0, ST->getBasePtr(),
+                       MF.getMachineMemOperand(MMO, 0, 1));
+  return Chain;
+}
+
 static SDValue performSTORECombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI,
                                    SelectionDAG &DAG,
@@ -21486,6 +21533,9 @@ static SDValue performSTORECombine(SDNode *N,
     return EltVT == MVT::f32 || EltVT == MVT::f64;
   };
 
+  if (SDValue Res = combineI8TruncStore(ST, DAG, Subtarget))
+    return Res;
+
   // If this is an FP_ROUND followed by a store, fold this into a truncating
   // store. We can do this even if this is already a truncstore.
   // We purposefully don't care about legality of the nodes here as we know

llvm/test/CodeGen/AArch64/vec3-loads-ext-trunc-stores.ll

@@ -348,17 +348,15 @@ define <3 x i32> @load_v3i32(ptr %src) {
 define void @store_trunc_from_64bits(ptr %src, ptr %dst) {
 ; CHECK-LABEL: store_trunc_from_64bits:
 ; CHECK:       ; %bb.0: ; %entry
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
-; CHECK-NEXT:    ldr s0, [x0]
-; CHECK-NEXT:    ldrh w8, [x0, #4]
-; CHECK-NEXT:    mov.h v0[2], w8
-; CHECK-NEXT:    xtn.8b v0, v0
-; CHECK-NEXT:    str s0, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #2]
-; CHECK-NEXT:    strh w9, [x1]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    ldr w8, [x0]
+; CHECK-NEXT:    add x9, x0, #4
+; CHECK-NEXT:    ld1r.4h { v0 }, [x9]
+; CHECK-NEXT:    fmov s1, w8
+; CHECK-NEXT:    strb w8, [x1]
+; CHECK-NEXT:    add x8, x1, #1
+; CHECK-NEXT:    st1.b { v1 }[2], [x8]
+; CHECK-NEXT:    add x8, x1, #2
+; CHECK-NEXT:    st1.b { v0 }[4], [x8]
 ; CHECK-NEXT:    ret
 ;
 ; BE-LABEL: store_trunc_from_64bits:
@@ -387,23 +385,19 @@ entry:
 define void @store_trunc_add_from_64bits(ptr %src, ptr %dst) {
 ; CHECK-LABEL: store_trunc_add_from_64bits:
 ; CHECK:       ; %bb.0: ; %entry
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    ldr s0, [x0]
 ; CHECK-NEXT:    add x9, x0, #4
 ; CHECK-NEXT:  Lloh0:
 ; CHECK-NEXT:    adrp x8, lCPI9_0@PAGE
 ; CHECK-NEXT:  Lloh1:
 ; CHECK-NEXT:    ldr d1, [x8, lCPI9_0@PAGEOFF]
+; CHECK-NEXT:    add x8, x1, #1
 ; CHECK-NEXT:    ld1.h { v0 }[2], [x9]
+; CHECK-NEXT:    add x9, x1, #2
 ; CHECK-NEXT:    add.4h v0, v0, v1
-; CHECK-NEXT:    xtn.8b v1, v0
-; CHECK-NEXT:    umov.h w8, v0[2]
-; CHECK-NEXT:    str s1, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #2]
-; CHECK-NEXT:    strh w9, [x1]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    st1.b { v0 }[2], [x8]
+; CHECK-NEXT:    st1.b { v0 }[4], [x9]
+; CHECK-NEXT:    st1.b { v0 }[0], [x1]
 ; CHECK-NEXT:    ret
 ; CHECK-NEXT:    .loh AdrpLdr Lloh0, Lloh1
 ;
@@ -594,17 +588,13 @@ entry:
 define void @shift_trunc_store(ptr %src, ptr %dst) {
 ; CHECK-LABEL: shift_trunc_store:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    ldr q0, [x0]
-; CHECK-NEXT:    shrn.4h v0, v0, #16
-; CHECK-NEXT:    xtn.8b v1, v0
-; CHECK-NEXT:    umov.h w8, v0[2]
-; CHECK-NEXT:    str s1, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #2]
-; CHECK-NEXT:    strh w9, [x1]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    add x8, x1, #1
+; CHECK-NEXT:    add x9, x1, #2
+; CHECK-NEXT:    ushr.4s v0, v0, #16
+; CHECK-NEXT:    st1.b { v0 }[4], [x8]
+; CHECK-NEXT:    st1.b { v0 }[8], [x9]
+; CHECK-NEXT:    st1.b { v0 }[0], [x1]
 ; CHECK-NEXT:    ret
 ;
 ; BE-LABEL: shift_trunc_store:
@@ -632,17 +622,13 @@ define void @shift_trunc_store(ptr %src, ptr %dst) {
 define void @shift_trunc_store_default_align(ptr %src, ptr %dst) {
 ; CHECK-LABEL: shift_trunc_store_default_align:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    ldr q0, [x0]
-; CHECK-NEXT:    shrn.4h v0, v0, #16
-; CHECK-NEXT:    xtn.8b v1, v0
-; CHECK-NEXT:    umov.h w8, v0[2]
-; CHECK-NEXT:    str s1, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #2]
-; CHECK-NEXT:    strh w9, [x1]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    add x8, x1, #1
+; CHECK-NEXT:    add x9, x1, #2
+; CHECK-NEXT:    ushr.4s v0, v0, #16
+; CHECK-NEXT:    st1.b { v0 }[4], [x8]
+; CHECK-NEXT:    st1.b { v0 }[8], [x9]
+; CHECK-NEXT:    st1.b { v0 }[0], [x1]
 ; CHECK-NEXT:    ret
 ;
 ; BE-LABEL: shift_trunc_store_default_align:
@@ -670,17 +656,13 @@ define void @shift_trunc_store_default_align(ptr %src, ptr %dst) {
 define void @shift_trunc_store_align_4(ptr %src, ptr %dst) {
 ; CHECK-LABEL: shift_trunc_store_align_4:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    ldr q0, [x0]
-; CHECK-NEXT:    shrn.4h v0, v0, #16
-; CHECK-NEXT:    xtn.8b v1, v0
-; CHECK-NEXT:    umov.h w8, v0[2]
-; CHECK-NEXT:    str s1, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #2]
-; CHECK-NEXT:    strh w9, [x1]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    add x8, x1, #1
+; CHECK-NEXT:    add x9, x1, #2
+; CHECK-NEXT:    ushr.4s v0, v0, #16
+; CHECK-NEXT:    st1.b { v0 }[4], [x8]
+; CHECK-NEXT:    st1.b { v0 }[8], [x9]
+; CHECK-NEXT:    st1.b { v0 }[0], [x1]
 ; CHECK-NEXT:    ret
 ;
 ; BE-LABEL: shift_trunc_store_align_4:
@@ -708,17 +690,14 @@ define void @shift_trunc_store_align_4(ptr %src, ptr %dst) {
 define void @shift_trunc_store_const_offset_1(ptr %src, ptr %dst) {
 ; CHECK-LABEL: shift_trunc_store_const_offset_1:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    ldr q0, [x0]
-; CHECK-NEXT:    shrn.4h v0, v0, #16
-; CHECK-NEXT:    xtn.8b v1, v0
-; CHECK-NEXT:    umov.h w8, v0[2]
-; CHECK-NEXT:    str s1, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #3]
-; CHECK-NEXT:    sturh w9, [x1, #1]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    add x8, x1, #2
+; CHECK-NEXT:    add x9, x1, #3
+; CHECK-NEXT:    ushr.4s v0, v0, #16
+; CHECK-NEXT:    st1.b { v0 }[4], [x8]
+; CHECK-NEXT:    add x8, x1, #1
+; CHECK-NEXT:    st1.b { v0 }[8], [x9]
+; CHECK-NEXT:    st1.b { v0 }[0], [x8]
 ; CHECK-NEXT:    ret
 ;
 ; BE-LABEL: shift_trunc_store_const_offset_1:
@@ -747,17 +726,14 @@ define void @shift_trunc_store_const_offset_1(ptr %src, ptr %dst) {
 define void @shift_trunc_store_const_offset_3(ptr %src, ptr %dst) {
 ; CHECK-LABEL: shift_trunc_store_const_offset_3:
 ; CHECK:       ; %bb.0:
-; CHECK-NEXT:    sub sp, sp, #16
-; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    ldr q0, [x0]
-; CHECK-NEXT:    shrn.4h v0, v0, #16
-; CHECK-NEXT:    xtn.8b v1, v0
-; CHECK-NEXT:    umov.h w8, v0[2]
-; CHECK-NEXT:    str s1, [sp, #12]
-; CHECK-NEXT:    ldrh w9, [sp, #12]
-; CHECK-NEXT:    strb w8, [x1, #5]
-; CHECK-NEXT:    sturh w9, [x1, #3]
-; CHECK-NEXT:    add sp, sp, #16
+; CHECK-NEXT:    add x8, x1, #4
+; CHECK-NEXT:    add x9, x1, #5
+; CHECK-NEXT:    ushr.4s v0, v0, #16
+; CHECK-NEXT:    st1.b { v0 }[4], [x8]
+; CHECK-NEXT:    add x8, x1, #3
+; CHECK-NEXT:    st1.b { v0 }[8], [x9]
+; CHECK-NEXT:    st1.b { v0 }[0], [x8]
 ; CHECK-NEXT:    ret
 ;
 ; BE-LABEL: shift_trunc_store_const_offset_3: