[AArch64][SVE] Lower unpredicated loads/stores as fixed LDR/STR with -msve-vector-bits=128.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -23654,6 +23654,29 @@ static SDValue combineV3I8LoadExt(LoadSDNode *LD, SelectionDAG &DAG) {
   return DAG.getMergeValues({Extract, TokenFactor}, DL);
 }
 
+// Replace packed scalable loads with fixed loads when vscale_range(1, 1).
+// This enables further optimisations such as LDP folds.
+static SDValue combineVScale1Load(LoadSDNode *LD, SelectionDAG &DAG,
+                                  TargetLowering::DAGCombinerInfo &DCI,
+                                  const AArch64Subtarget *Subtarget) {
+  EVT MemVT = LD->getMemoryVT();
+  if (!DCI.isBeforeLegalize() || !Subtarget->isLittleEndian() ||
+      !Subtarget->hasNEON() || !MemVT.isScalableVector() ||
+      LD->getExtensionType() != ISD::NON_EXTLOAD ||
+      MemVT.getSizeInBits().getKnownMinValue() != 128 ||
+      Subtarget->getMaxSVEVectorSizeInBits() != 128)
+    return SDValue();
+
+  SDLoc DL(LD);
+  EVT NewVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getVectorElementType(),
+                               MemVT.getVectorMinNumElements());
+  SDValue NewLoad = DAG.getLoad(
+      NewVT, DL, LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
+      LD->getOriginalAlign(), LD->getMemOperand()->getFlags(), LD->getAAInfo());
+  SDValue Insert = convertToScalableVector(DAG, MemVT, NewLoad);
+  return DAG.getMergeValues({Insert, NewLoad.getValue(1)}, DL);
+}
+
 // Perform TBI simplification if supported by the target and try to break up
 // nontemporal loads larger than 256-bits loads for odd types so LDNPQ 256-bit
 // load instructions can be selected.
@@ -23691,6 +23714,9 @@ static SDValue performLOADCombine(SDNode *N,
   if (SDValue Res = combineV3I8LoadExt(LD, DAG))
     return Res;
 
+  if (SDValue Res = combineVScale1Load(LD, DAG, DCI, Subtarget))
+    return Res;
+
   if (!LD->isNonTemporal())
     return SDValue(N, 0);
 
@@ -23949,6 +23975,29 @@ static SDValue combineI8TruncStore(StoreSDNode *ST, SelectionDAG &DAG,
   return Chain;
 }
 
+// Replace packed scalable stores with fixed stores when vscale_range(1, 1).
+static SDValue combineVScale1Store(StoreSDNode *ST, SelectionDAG &DAG,
+                                   TargetLowering::DAGCombinerInfo &DCI,
+                                   const AArch64Subtarget *Subtarget) {
+  SDValue Value = ST->getValue();
+  EVT ValueVT = Value.getValueType();
+  if (!DCI.isBeforeLegalize() || !Subtarget->isLittleEndian() ||
+      !Subtarget->hasNEON() || !ValueVT.isScalableVector() ||
+      ST->isTruncatingStore() ||
+      ValueVT.getSizeInBits().getKnownMinValue() != 128 ||
+      Subtarget->getMaxSVEVectorSizeInBits() != 128)
+    return SDValue();
+
+  SDLoc DL(ST);
+  EVT NewVT =
+      EVT::getVectorVT(*DAG.getContext(), ValueVT.getVectorElementType(),
+                       ValueVT.getVectorMinNumElements());
+  SDValue NewValue = convertFromScalableVector(DAG, NewVT, Value);
+  return DAG.getStore(ST->getChain(), DL, NewValue, ST->getBasePtr(),
+                      ST->getPointerInfo(), ST->getOriginalAlign(),
+                      ST->getMemOperand()->getFlags(), ST->getAAInfo());
+}
+
 static unsigned getFPSubregForVT(EVT VT) {
   assert(VT.isSimple() && "Expected simple VT");
   switch (VT.getSimpleVT().SimpleTy) {
@@ -23997,6 +24046,9 @@ static SDValue performSTORECombine(SDNode *N,
   if (SDValue Res = combineI8TruncStore(ST, DAG, Subtarget))
     return Res;
 
+  if (SDValue Res = combineVScale1Store(ST, DAG, DCI, 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/sve-fixed-length-offsets.ll

@@ -17,8 +17,8 @@ define void @nxv16i8(ptr %ldptr, ptr %stptr) {
 ;
 ; CHECK-128-LABEL: nxv16i8:
 ; CHECK-128:       // %bb.0:
-; CHECK-128-NEXT:    ldr z0, [x0, #16, mul vl]
-; CHECK-128-NEXT:    str z0, [x1, #16, mul vl]
+; CHECK-128-NEXT:    ldr q0, [x0, #256]
+; CHECK-128-NEXT:    str q0, [x1, #256]
 ; CHECK-128-NEXT:    ret
 ;
 ; CHECK-256-LABEL: nxv16i8:
@@ -62,8 +62,8 @@ define void @nxv8i16(ptr %ldptr, ptr %stptr) {
 ;
 ; CHECK-128-LABEL: nxv8i16:
 ; CHECK-128:       // %bb.0:
-; CHECK-128-NEXT:    ldr z0, [x0, #16, mul vl]
-; CHECK-128-NEXT:    str z0, [x1, #16, mul vl]
+; CHECK-128-NEXT:    ldr q0, [x0, #256]
+; CHECK-128-NEXT:    str q0, [x1, #256]
 ; CHECK-128-NEXT:    ret
 ;
 ; CHECK-256-LABEL: nxv8i16:
@@ -107,8 +107,8 @@ define void @nxv4i32(ptr %ldptr, ptr %stptr) {
 ;
 ; CHECK-128-LABEL: nxv4i32:
 ; CHECK-128:       // %bb.0:
-; CHECK-128-NEXT:    ldr z0, [x0, #16, mul vl]
-; CHECK-128-NEXT:    str z0, [x1, #16, mul vl]
+; CHECK-128-NEXT:    ldr q0, [x0, #256]
+; CHECK-128-NEXT:    str q0, [x1, #256]
 ; CHECK-128-NEXT:    ret
 ;
 ; CHECK-256-LABEL: nxv4i32:
@@ -152,8 +152,8 @@ define void @nxv2i64(ptr %ldptr, ptr %stptr) {
 ;
 ; CHECK-128-LABEL: nxv2i64:
 ; CHECK-128:       // %bb.0:
-; CHECK-128-NEXT:    ldr z0, [x0, #16, mul vl]
-; CHECK-128-NEXT:    str z0, [x1, #16, mul vl]
+; CHECK-128-NEXT:    ldr q0, [x0, #256]
+; CHECK-128-NEXT:    str q0, [x1, #256]
 ; CHECK-128-NEXT:    ret
 ;
 ; CHECK-256-LABEL: nxv2i64:

llvm/test/CodeGen/AArch64/sve-unpred-loads-stores.ll

@@ -0,0 +1,239 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve -aarch64-sve-vector-bits-max=128 < %s | FileCheck %s --check-prefix=CHECK-128
+; RUN: llc -mtriple=aarch64_be-linux-gnu -mattr=+sve -aarch64-sve-vector-bits-max=128 < %s | FileCheck %s --check-prefix=CHECK-BE-128
+
+define <vscale x 16 x i8> @ld_nxv16i8(ptr %0) {
+; CHECK-128-LABEL: ld_nxv16i8:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv16i8:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.b
+; CHECK-BE-128-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 16 x i8>, ptr %0, align 1
+  ret <vscale x 16 x i8> %2
+}
+
+define void @st_nxv16i8(ptr %0, <vscale x 16 x i8> %1) {
+; CHECK-128-LABEL: st_nxv16i8:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv16i8:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.b
+; CHECK-BE-128-NEXT:    st1b { z0.b }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 16 x i8> %1, ptr %0, align 1
+  ret void
+}
+
+define <vscale x 8 x i16> @ld_nxv8i16(ptr %0) {
+; CHECK-128-LABEL: ld_nxv8i16:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv8i16:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.h
+; CHECK-BE-128-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 8 x i16>, ptr %0, align 2
+  ret <vscale x 8 x i16> %2
+}
+
+define void @st_nxv8i16(ptr %0, <vscale x 8 x i16> %1) {
+; CHECK-128-LABEL: st_nxv8i16:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv8i16:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.h
+; CHECK-BE-128-NEXT:    st1h { z0.h }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 8 x i16> %1, ptr %0, align 2
+  ret void
+}
+
+define <vscale x 4 x i32> @ld_nxv4i32(ptr %0) {
+; CHECK-128-LABEL: ld_nxv4i32:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv4i32:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.s
+; CHECK-BE-128-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 4 x i32>, ptr %0, align 4
+  ret <vscale x 4 x i32> %2
+}
+
+define void @st_nxv4i32(ptr %0, <vscale x 4 x i32> %1) {
+; CHECK-128-LABEL: st_nxv4i32:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv4i32:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.s
+; CHECK-BE-128-NEXT:    st1w { z0.s }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 4 x i32> %1, ptr %0, align 4
+  ret void
+}
+
+define <vscale x 2 x i64> @ld_nxv2i64(ptr %0) {
+; CHECK-128-LABEL: ld_nxv2i64:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv2i64:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.d
+; CHECK-BE-128-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 2 x i64>, ptr %0, align 8
+  ret <vscale x 2 x i64> %2
+}
+
+define void @st_nxv2i64(ptr %0, <vscale x 2 x i64> %1) {
+; CHECK-128-LABEL: st_nxv2i64:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv2i64:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.d
+; CHECK-BE-128-NEXT:    st1d { z0.d }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 2 x i64> %1, ptr %0, align 8
+  ret void
+}
+
+define <vscale x 8 x half> @ld_nxv8f16(ptr %0) {
+; CHECK-128-LABEL: ld_nxv8f16:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv8f16:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.h
+; CHECK-BE-128-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 8 x half>, ptr %0, align 2
+  ret <vscale x 8 x half> %2
+}
+
+define void @st_nxv8f16(ptr %0, <vscale x 8 x half> %1) {
+; CHECK-128-LABEL: st_nxv8f16:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv8f16:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.h
+; CHECK-BE-128-NEXT:    st1h { z0.h }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 8 x half> %1, ptr %0, align 2
+  ret void
+}
+
+define <vscale x 4 x float> @ld_nxv4f32(ptr %0) {
+; CHECK-128-LABEL: ld_nxv4f32:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv4f32:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.s
+; CHECK-BE-128-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 4 x float>, ptr %0, align 4
+  ret <vscale x 4 x float> %2
+}
+
+define void @st_nxv4f32(ptr %0, <vscale x 4 x float> %1) {
+; CHECK-128-LABEL: st_nxv4f32:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv4f32:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.s
+; CHECK-BE-128-NEXT:    st1w { z0.s }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 4 x float> %1, ptr %0, align 4
+  ret void
+}
+
+define <vscale x 2 x double> @ld_nxv2f64(ptr %0) {
+; CHECK-128-LABEL: ld_nxv2f64:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldr q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ld_nxv2f64:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.d
+; CHECK-BE-128-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    ret
+  %2 = load <vscale x 2 x double>, ptr %0, align 8
+  ret <vscale x 2 x double> %2
+}
+
+define void @st_nxv2f64(ptr %0, <vscale x 2 x double> %1) {
+; CHECK-128-LABEL: st_nxv2f64:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    str q0, [x0]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: st_nxv2f64:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.d
+; CHECK-BE-128-NEXT:    st1d { z0.d }, p0, [x0]
+; CHECK-BE-128-NEXT:    ret
+  store <vscale x 2 x double> %1, ptr %0, align 8
+  ret void
+}
+
+; Test LDP/STP fold.
+define void @ldp_stp_nxv16i8_offset(ptr %ldptr, ptr %stptr) {
+; CHECK-128-LABEL: ldp_stp_nxv16i8_offset:
+; CHECK-128:       // %bb.0:
+; CHECK-128-NEXT:    ldp q0, q1, [x0, #-16]
+; CHECK-128-NEXT:    stp q0, q1, [x1, #-16]
+; CHECK-128-NEXT:    ret
+;
+; CHECK-BE-128-LABEL: ldp_stp_nxv16i8_offset:
+; CHECK-BE-128:       // %bb.0:
+; CHECK-BE-128-NEXT:    ptrue p0.b
+; CHECK-BE-128-NEXT:    mov x8, #-16 // =0xfffffffffffffff0
+; CHECK-BE-128-NEXT:    ld1b { z0.b }, p0/z, [x0, x8]
+; CHECK-BE-128-NEXT:    ld1b { z1.b }, p0/z, [x0]
+; CHECK-BE-128-NEXT:    st1b { z0.b }, p0, [x1, x8]
+; CHECK-BE-128-NEXT:    st1b { z1.b }, p0, [x1]
+; CHECK-BE-128-NEXT:    ret
+  %ldptr.1 = getelementptr inbounds i8, ptr %ldptr, i64 -16
+  %ld1 = load <vscale x 16 x i8>, ptr %ldptr.1, align 1
+  %ld2 = load <vscale x 16 x i8>, ptr %ldptr, align 1
+  %stptr.1 = getelementptr inbounds i8, ptr %stptr, i64 -16
+  store <vscale x 16 x i8> %ld1, ptr %stptr.1, align 1
+  store <vscale x 16 x i8> %ld2, ptr %stptr, align 1
+  ret void
+}