[llvm-branch-commits] [llvm] [LoongArch] Lower splat of variable-index extract to [x]vreplve (PR #244542)
via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Fri Nov 20 01:21:47 PST 2026
https://github.com/heiher created https://github.com/llvm/llvm-project/pull/244542
Lower a splat of a variable-index extract_vector_elt to [x]vreplve. For 256-bit vectors, each 128-bit lane must splat the element at the same position within the lane, which is the per-lane semantics of xvreplve.
Also, [x]vreplve only uses the low bits of the index, so drop redundant masking of it.
Assisted-By: LLM
>From 65f9f9bfa5137a0d24d4316b517896b99f9117ff Mon Sep 17 00:00:00 2001
From: WANG Rui <wangrui at loongson.cn>
Date: Fri, 20 Nov 2026 16:15:10 +0800
Subject: [PATCH] [LoongArch] Lower splat of variable-index extract to
[x]vreplve
Lower a splat of a variable-index extract_vector_elt to [x]vreplve. For
256-bit vectors, each 128-bit lane must splat the element at the same
position within the lane, which is the per-lane semantics of xvreplve.
Also, [x]vreplve only uses the low bits of the index, so drop redundant
masking of it.
Assisted-By: LLM
---
.../LoongArch/LoongArchISelLowering.cpp | 110 +++++++++
.../LoongArch/lasx/splat-extract-var-idx.ll | 224 ++----------------
.../LoongArch/lsx/splat-extract-var-idx.ll | 47 +---
3 files changed, 139 insertions(+), 242 deletions(-)
diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
index 15e2e198abeae1..d53ee0151d4654 100644
--- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
+++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp
@@ -3652,6 +3652,89 @@ static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
return SDValue();
}
+/// Lower \p Elts to [X]VREPLVE Src, Idx if each 128-bit lane L is a splat of
+/// (extract_vector_elt Src, Idx + L * N), where N is the number of elements
+/// per lane. For 256-bit vectors, Idx must be known to be less than N.
+static SDValue
+lowerSplatOfExtractAsVREPLVE(const SDLoc &DL, MVT VT, ArrayRef<SDValue> Elts,
+ SelectionDAG &DAG,
+ const LoongArchSubtarget &Subtarget) {
+ unsigned NumElts = Elts.size();
+ unsigned NumLanes = VT.getSizeInBits() / 128;
+ unsigned NumLaneElts = NumElts / NumLanes;
+ SDValue Src, Idx;
+
+ for (unsigned L = 0; L < NumLanes; ++L) {
+ SDValue LaneElt;
+ for (SDValue Elt : Elts.slice(L * NumLaneElts, NumLaneElts)) {
+ if (!Elt || Elt.isUndef())
+ continue;
+ if (!LaneElt)
+ LaneElt = Elt;
+ else if (LaneElt != Elt)
+ return SDValue();
+ }
+
+ if (!LaneElt || LaneElt.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ return SDValue();
+ SDValue Vec = LaneElt.getOperand(0);
+ SDValue EltIdx = LaneElt.getOperand(1);
+ if (isa<ConstantSDNode>(EltIdx))
+ return SDValue();
+
+ if (L == 0) {
+ EVT SrcVT = Vec.getValueType();
+ if (SrcVT.getSizeInBits() != VT.getSizeInBits() ||
+ SrcVT.getVectorNumElements() != NumElts)
+ return SDValue();
+ Src = Vec;
+ Idx = EltIdx;
+ continue;
+ }
+
+ // Must select the same position within the lane.
+ if (Vec != Src || !DAG.isADDLike(EltIdx) || EltIdx.getOperand(0) != Idx)
+ return SDValue();
+ auto *Offset = dyn_cast<ConstantSDNode>(EltIdx.getOperand(1));
+ if (!Offset || Offset->getZExtValue() != L * NumLaneElts)
+ return SDValue();
+ }
+
+ if (NumLanes > 1 && DAG.computeKnownBits(Idx).getMaxValue().uge(NumLaneElts))
+ return SDValue();
+
+ MVT IntVT = VT.changeVectorElementTypeToInteger();
+ Idx = DAG.getZExtOrTrunc(Idx, DL, Subtarget.getGRLenVT());
+ SDValue Res = DAG.getNode(LoongArchISD::VREPLVE, DL, IntVT,
+ DAG.getBitcast(IntVT, Src), Idx);
+ return DAG.getBitcast(VT, Res);
+}
+
+/// Lower a VECTOR_SHUFFLE of BUILD_VECTORs to [X]VREPLVE.
+static SDValue
+lowerVECTOR_SHUFFLE_VREPLVE(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
+ SDValue V1, SDValue V2, SelectionDAG &DAG,
+ const LoongArchSubtarget &Subtarget) {
+ int NumElts = Mask.size();
+ SmallVector<SDValue, 32> Elts;
+ for (int M : Mask) {
+ if (M < 0) {
+ Elts.push_back(SDValue());
+ continue;
+ }
+ SDValue V = M < NumElts ? V1 : V2;
+ if (V.isUndef()) {
+ Elts.push_back(SDValue());
+ continue;
+ }
+ if (V.getOpcode() != ISD::BUILD_VECTOR)
+ return SDValue();
+ Elts.push_back(V.getOperand(M % NumElts));
+ }
+
+ return lowerSplatOfExtractAsVREPLVE(DL, VT, Elts, DAG, Subtarget);
+}
+
SDValue LoongArchTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
SelectionDAG &DAG) const {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
@@ -3692,6 +3775,10 @@ SDValue LoongArchTargetLowering::lowerVECTOR_SHUFFLE(SDValue Op,
[&](int M) { return -1 <= M && M < MaskUpperLimit; }) &&
"Out of bounds shuffle index");
+ if (SDValue Result =
+ lowerVECTOR_SHUFFLE_VREPLVE(DL, OrigMask, VT, V1, V2, DAG, Subtarget))
+ return Result;
+
// For each vector width, delegate to a specialized lowering routine.
if (VT.is128BitVector())
return lower128BitShuffle(DL, OrigMask, VT, V1, V2, DAG, Subtarget);
@@ -3924,6 +4011,11 @@ SDValue LoongArchTargetLowering::lowerBUILD_VECTOR(SDValue Op,
return Result;
}
+ // Lower per-lane splat of a variable-index extract to [X]VREPLVE.
+ if (SDValue Result = lowerSplatOfExtractAsVREPLVE(
+ DL, VT, SmallVector<SDValue>(Node->op_values()), DAG, Subtarget))
+ return Result;
+
if (DAG.isSplatValue(Op, /*AllowUndefs=*/false))
return Op;
@@ -8958,6 +9050,22 @@ performSPLIT_PAIR_F64Combine(SDNode *N, SelectionDAG &DAG,
return SDValue();
}
+/// Do target-specific dag combines on LoongArchISD::VREPLVE nodes.
+static SDValue performVREPLVECombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const LoongArchSubtarget &Subtarget) {
+ // [X]VREPLVE only uses the low log2(N) bits of the index, where N is the
+ // number of elements per 128-bit lane.
+ SDValue Idx = N->getOperand(1);
+ unsigned NumLaneElts = 128 / N->getSimpleValueType(0).getScalarSizeInBits();
+ APInt DemandedBits =
+ APInt::getLowBitsSet(Idx.getValueSizeInBits(), Log2_32(NumLaneElts));
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (TLI.SimplifyDemandedBits(Idx, DemandedBits, DCI))
+ return SDValue(N, 0);
+ return SDValue();
+}
+
/// Do target-specific dag combines on LoongArchISD::VANDN nodes.
static SDValue performVANDNCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
@@ -9862,6 +9970,8 @@ SDValue LoongArchTargetLowering::PerformDAGCombine(SDNode *N,
return performSPLIT_PAIR_F64Combine(N, DAG, DCI, Subtarget);
case LoongArchISD::VANDN:
return performVANDNCombine(N, DAG, DCI, Subtarget);
+ case LoongArchISD::VREPLVE:
+ return performVREPLVECombine(N, DAG, DCI, Subtarget);
case ISD::CONCAT_VECTORS:
return performCONCAT_VECTORSCombine(N, DAG, DCI, Subtarget);
case ISD::VSELECT:
diff --git a/llvm/test/CodeGen/LoongArch/lasx/splat-extract-var-idx.ll b/llvm/test/CodeGen/LoongArch/lasx/splat-extract-var-idx.ll
index 8592555717ed96..44c440240a7fe0 100644
--- a/llvm/test/CodeGen/LoongArch/lasx/splat-extract-var-idx.ll
+++ b/llvm/test/CodeGen/LoongArch/lasx/splat-extract-var-idx.ll
@@ -25,7 +25,6 @@ declare <32 x i8> @llvm.loongarch.lasx.xvreplve.b(<32 x i8>, i32)
define <32 x i8> @intrinsic_xvreplve_b_mask(<32 x i8> %a, i32 %b) nounwind {
; CHECK-LABEL: intrinsic_xvreplve_b_mask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 15
; CHECK-NEXT: xvreplve.b $xr0, $xr0, $a0
; CHECK-NEXT: ret
entry:
@@ -38,7 +37,6 @@ entry:
define <32 x i8> @intrinsic_xvreplve_b_mask31(<32 x i8> %a, i32 %b) nounwind {
; CHECK-LABEL: intrinsic_xvreplve_b_mask31:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 31
; CHECK-NEXT: xvreplve.b $xr0, $xr0, $a0
; CHECK-NEXT: ret
entry:
@@ -61,39 +59,10 @@ entry:
}
;; Splat of a[idx] in the low lane and a[idx + 16] in the high lane is XVREPLVE.
define <32 x i8> @lane_splat_extract_v32i8(<32 x i8> %a, i32 %b) nounwind {
-; LA32-LABEL: lane_splat_extract_v32i8:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: andi $a1, $a0, 15
-; LA32-NEXT: movgr2fr.w $fa1, $a1
-; LA32-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA32-NEXT: xvshuf.b $xr1, $xr2, $xr0, $xr1
-; LA32-NEXT: vpickve2gr.b $a1, $vr1, 0
-; LA32-NEXT: xvreplgr2vr.b $xr1, $a1
-; LA32-NEXT: ori $a1, $zero, 1
-; LA32-NEXT: bstrins.w $a0, $a1, 31, 4
-; LA32-NEXT: movgr2fr.w $fa3, $a0
-; LA32-NEXT: xvshuf.b $xr0, $xr2, $xr0, $xr3
-; LA32-NEXT: vpickve2gr.b $a0, $vr0, 0
-; LA32-NEXT: xvreplgr2vr.b $xr0, $a0
-; LA32-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA32-NEXT: ret
-;
-; LA64-LABEL: lane_splat_extract_v32i8:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 15
-; LA64-NEXT: movgr2fr.w $fa1, $a1
-; LA64-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA64-NEXT: xvshuf.b $xr1, $xr2, $xr0, $xr1
-; LA64-NEXT: vpickve2gr.b $a1, $vr1, 0
-; LA64-NEXT: xvreplgr2vr.b $xr1, $a1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 4
-; LA64-NEXT: movgr2fr.w $fa3, $a0
-; LA64-NEXT: xvshuf.b $xr0, $xr2, $xr0, $xr3
-; LA64-NEXT: vpickve2gr.b $a0, $vr0, 0
-; LA64-NEXT: xvreplgr2vr.b $xr0, $a0
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA64-NEXT: ret
+; CHECK-LABEL: lane_splat_extract_v32i8:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: xvreplve.b $xr0, $xr0, $a0
+; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 15
%e0 = extractelement <32 x i8> %a, i32 %idx
@@ -108,39 +77,10 @@ entry:
}
define <16 x i16> @lane_splat_extract_v16i16(<16 x i16> %a, i32 %b) nounwind {
-; LA32-LABEL: lane_splat_extract_v16i16:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: andi $a1, $a0, 7
-; LA32-NEXT: movgr2fr.w $fa1, $a1
-; LA32-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA32-NEXT: xvshuf.h $xr1, $xr2, $xr0
-; LA32-NEXT: vpickve2gr.h $a1, $vr1, 0
-; LA32-NEXT: xvreplgr2vr.h $xr1, $a1
-; LA32-NEXT: ori $a1, $zero, 1
-; LA32-NEXT: bstrins.w $a0, $a1, 31, 3
-; LA32-NEXT: movgr2fr.w $fa3, $a0
-; LA32-NEXT: xvshuf.h $xr3, $xr2, $xr0
-; LA32-NEXT: vpickve2gr.h $a0, $vr3, 0
-; LA32-NEXT: xvreplgr2vr.h $xr0, $a0
-; LA32-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA32-NEXT: ret
-;
-; LA64-LABEL: lane_splat_extract_v16i16:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 7
-; LA64-NEXT: movgr2fr.w $fa1, $a1
-; LA64-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA64-NEXT: xvshuf.h $xr1, $xr2, $xr0
-; LA64-NEXT: vpickve2gr.h $a1, $vr1, 0
-; LA64-NEXT: xvreplgr2vr.h $xr1, $a1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 3
-; LA64-NEXT: movgr2fr.w $fa3, $a0
-; LA64-NEXT: xvshuf.h $xr3, $xr2, $xr0
-; LA64-NEXT: vpickve2gr.h $a0, $vr3, 0
-; LA64-NEXT: xvreplgr2vr.h $xr0, $a0
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA64-NEXT: ret
+; CHECK-LABEL: lane_splat_extract_v16i16:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: xvreplve.h $xr0, $xr0, $a0
+; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 7
%e0 = extractelement <16 x i16> %a, i32 %idx
@@ -155,37 +95,10 @@ entry:
}
define <8 x i32> @lane_splat_extract_v8i32(<8 x i32> %a, i32 %b) nounwind {
-; LA32-LABEL: lane_splat_extract_v8i32:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: andi $a1, $a0, 3
-; LA32-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA32-NEXT: xvperm.w $xr1, $xr0, $xr1
-; LA32-NEXT: xvpickve2gr.w $a1, $xr1, 0
-; LA32-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA32-NEXT: ori $a1, $zero, 1
-; LA32-NEXT: bstrins.w $a0, $a1, 31, 2
-; LA32-NEXT: xvreplgr2vr.w $xr2, $a0
-; LA32-NEXT: xvperm.w $xr0, $xr0, $xr2
-; LA32-NEXT: xvpickve2gr.w $a0, $xr0, 0
-; LA32-NEXT: xvreplgr2vr.w $xr0, $a0
-; LA32-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA32-NEXT: ret
-;
-; LA64-LABEL: lane_splat_extract_v8i32:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 3
-; LA64-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA64-NEXT: xvperm.w $xr1, $xr0, $xr1
-; LA64-NEXT: xvpickve2gr.w $a1, $xr1, 0
-; LA64-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 2
-; LA64-NEXT: xvreplgr2vr.w $xr2, $a0
-; LA64-NEXT: xvperm.w $xr0, $xr0, $xr2
-; LA64-NEXT: xvpickve2gr.w $a0, $xr0, 0
-; LA64-NEXT: xvreplgr2vr.w $xr0, $a0
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA64-NEXT: ret
+; CHECK-LABEL: lane_splat_extract_v8i32:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: xvreplve.w $xr0, $xr0, $a0
+; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 3
%e0 = extractelement <8 x i32> %a, i32 %idx
@@ -236,19 +149,7 @@ define <4 x i64> @lane_splat_extract_v4i64(<4 x i64> %a, i32 %b) nounwind {
;
; LA64-LABEL: lane_splat_extract_v4i64:
; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 1
-; LA64-NEXT: movgr2fr.w $fa1, $a1
-; LA64-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA64-NEXT: xvshuf.d $xr1, $xr2, $xr0
-; LA64-NEXT: xvpickve2gr.d $a1, $xr1, 0
-; LA64-NEXT: xvreplgr2vr.d $xr1, $a1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 1
-; LA64-NEXT: movgr2fr.w $fa3, $a0
-; LA64-NEXT: xvshuf.d $xr3, $xr2, $xr0
-; LA64-NEXT: xvpickve2gr.d $a0, $xr3, 0
-; LA64-NEXT: xvreplgr2vr.d $xr0, $a0
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
+; LA64-NEXT: xvreplve.d $xr0, $xr0, $a0
; LA64-NEXT: ret
entry:
%idx = and i32 %b, 1
@@ -264,33 +165,10 @@ entry:
}
define <8 x float> @lane_splat_extract_v8f32(<8 x float> %a, i32 %b) nounwind {
-; LA32-LABEL: lane_splat_extract_v8f32:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: andi $a1, $a0, 3
-; LA32-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA32-NEXT: xvperm.w $xr1, $xr0, $xr1
-; LA32-NEXT: xvreplve0.w $xr1, $xr1
-; LA32-NEXT: ori $a1, $zero, 1
-; LA32-NEXT: bstrins.w $a0, $a1, 31, 2
-; LA32-NEXT: xvreplgr2vr.w $xr2, $a0
-; LA32-NEXT: xvperm.w $xr0, $xr0, $xr2
-; LA32-NEXT: xvreplve0.w $xr0, $xr0
-; LA32-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA32-NEXT: ret
-;
-; LA64-LABEL: lane_splat_extract_v8f32:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 3
-; LA64-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA64-NEXT: xvperm.w $xr1, $xr0, $xr1
-; LA64-NEXT: xvreplve0.w $xr1, $xr1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 2
-; LA64-NEXT: xvreplgr2vr.w $xr2, $a0
-; LA64-NEXT: xvperm.w $xr0, $xr0, $xr2
-; LA64-NEXT: xvreplve0.w $xr0, $xr0
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA64-NEXT: ret
+; CHECK-LABEL: lane_splat_extract_v8f32:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: xvreplve.w $xr0, $xr0, $a0
+; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 3
%e0 = extractelement <8 x float> %a, i32 %idx
@@ -305,35 +183,10 @@ entry:
}
define <4 x double> @lane_splat_extract_v4f64(<4 x double> %a, i32 %b) nounwind {
-; LA32-LABEL: lane_splat_extract_v4f64:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: andi $a1, $a0, 1
-; LA32-NEXT: movgr2fr.w $fa1, $a1
-; LA32-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA32-NEXT: xvshuf.d $xr1, $xr2, $xr0
-; LA32-NEXT: xvreplve0.d $xr1, $xr1
-; LA32-NEXT: ori $a1, $zero, 1
-; LA32-NEXT: bstrins.w $a0, $a1, 31, 1
-; LA32-NEXT: movgr2fr.w $fa3, $a0
-; LA32-NEXT: xvshuf.d $xr3, $xr2, $xr0
-; LA32-NEXT: xvreplve0.d $xr0, $xr3
-; LA32-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA32-NEXT: ret
-;
-; LA64-LABEL: lane_splat_extract_v4f64:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 1
-; LA64-NEXT: movgr2fr.w $fa1, $a1
-; LA64-NEXT: xvpermi.q $xr2, $xr0, 1
-; LA64-NEXT: xvshuf.d $xr1, $xr2, $xr0
-; LA64-NEXT: xvreplve0.d $xr1, $xr1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 1
-; LA64-NEXT: movgr2fr.w $fa3, $a0
-; LA64-NEXT: xvshuf.d $xr3, $xr2, $xr0
-; LA64-NEXT: xvreplve0.d $xr0, $xr3
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA64-NEXT: ret
+; CHECK-LABEL: lane_splat_extract_v4f64:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: xvreplve.d $xr0, $xr0, $a0
+; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 1
%e0 = extractelement <4 x double> %a, i32 %idx
@@ -349,37 +202,10 @@ entry:
;; The high lane index may also be formed by an add.
define <8 x i32> @lane_splat_extract_v8i32_add(<8 x i32> %a, i32 %b) nounwind {
-; LA32-LABEL: lane_splat_extract_v8i32_add:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: andi $a1, $a0, 3
-; LA32-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA32-NEXT: xvperm.w $xr1, $xr0, $xr1
-; LA32-NEXT: xvpickve2gr.w $a1, $xr1, 0
-; LA32-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA32-NEXT: ori $a1, $zero, 1
-; LA32-NEXT: bstrins.w $a0, $a1, 31, 2
-; LA32-NEXT: xvreplgr2vr.w $xr2, $a0
-; LA32-NEXT: xvperm.w $xr0, $xr0, $xr2
-; LA32-NEXT: xvpickve2gr.w $a0, $xr0, 0
-; LA32-NEXT: xvreplgr2vr.w $xr0, $a0
-; LA32-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA32-NEXT: ret
-;
-; LA64-LABEL: lane_splat_extract_v8i32_add:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a1, $a0, 3
-; LA64-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA64-NEXT: xvperm.w $xr1, $xr0, $xr1
-; LA64-NEXT: xvpickve2gr.w $a1, $xr1, 0
-; LA64-NEXT: xvreplgr2vr.w $xr1, $a1
-; LA64-NEXT: ori $a1, $zero, 1
-; LA64-NEXT: bstrins.d $a0, $a1, 63, 2
-; LA64-NEXT: xvreplgr2vr.w $xr2, $a0
-; LA64-NEXT: xvperm.w $xr0, $xr0, $xr2
-; LA64-NEXT: xvpickve2gr.w $a0, $xr0, 0
-; LA64-NEXT: xvreplgr2vr.w $xr0, $a0
-; LA64-NEXT: xvpermi.q $xr0, $xr1, 48
-; LA64-NEXT: ret
+; CHECK-LABEL: lane_splat_extract_v8i32_add:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: xvreplve.w $xr0, $xr0, $a0
+; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 3
%e0 = extractelement <8 x i32> %a, i32 %idx
diff --git a/llvm/test/CodeGen/LoongArch/lsx/splat-extract-var-idx.ll b/llvm/test/CodeGen/LoongArch/lsx/splat-extract-var-idx.ll
index e4299afa743363..131cbfcbb6cc3f 100644
--- a/llvm/test/CodeGen/LoongArch/lsx/splat-extract-var-idx.ll
+++ b/llvm/test/CodeGen/LoongArch/lsx/splat-extract-var-idx.ll
@@ -5,11 +5,7 @@
define <16 x i8> @splat_extract_v16i8(<16 x i8> %a, i32 %b) nounwind {
; CHECK-LABEL: splat_extract_v16i8:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 15
; CHECK-NEXT: vreplve.b $vr0, $vr0, $a0
-; CHECK-NEXT: movfr2gr.s $a0, $fa0
-; CHECK-NEXT: srai.w $a0, $a0, 24
-; CHECK-NEXT: vreplgr2vr.b $vr0, $a0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 15
@@ -22,11 +18,7 @@ entry:
define <8 x i16> @splat_extract_v8i16(<8 x i16> %a, i32 %b) nounwind {
; CHECK-LABEL: splat_extract_v8i16:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 7
; CHECK-NEXT: vreplve.h $vr0, $vr0, $a0
-; CHECK-NEXT: movfr2gr.s $a0, $fa0
-; CHECK-NEXT: srai.w $a0, $a0, 16
-; CHECK-NEXT: vreplgr2vr.h $vr0, $a0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 7
@@ -39,10 +31,7 @@ entry:
define <4 x i32> @splat_extract_v4i32(<4 x i32> %a, i32 %b) nounwind {
; CHECK-LABEL: splat_extract_v4i32:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 3
; CHECK-NEXT: vreplve.w $vr0, $vr0, $a0
-; CHECK-NEXT: movfr2gr.s $a0, $fa0
-; CHECK-NEXT: vreplgr2vr.w $vr0, $a0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 3
@@ -70,10 +59,7 @@ define <2 x i64> @splat_extract_v2i64(<2 x i64> %a, i32 %b) nounwind {
;
; LA64-LABEL: splat_extract_v2i64:
; LA64: # %bb.0: # %entry
-; LA64-NEXT: andi $a0, $a0, 1
; LA64-NEXT: vreplve.d $vr0, $vr0, $a0
-; LA64-NEXT: movfr2gr.d $a0, $fa0
-; LA64-NEXT: vreplgr2vr.d $vr0, $a0
; LA64-NEXT: ret
entry:
%idx = and i32 %b, 1
@@ -86,9 +72,7 @@ entry:
define <4 x float> @splat_extract_v4f32(<4 x float> %a, i32 %b) nounwind {
; CHECK-LABEL: splat_extract_v4f32:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 3
; CHECK-NEXT: vreplve.w $vr0, $vr0, $a0
-; CHECK-NEXT: vreplvei.w $vr0, $vr0, 0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 3
@@ -101,9 +85,7 @@ entry:
define <2 x double> @splat_extract_v2f64(<2 x double> %a, i32 %b) nounwind {
; CHECK-LABEL: splat_extract_v2f64:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 1
; CHECK-NEXT: vreplve.d $vr0, $vr0, $a0
-; CHECK-NEXT: vreplvei.d $vr0, $vr0, 0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 1
@@ -115,20 +97,10 @@ entry:
;; Unmasked index: out-of-range extract yields poison, so VREPLVE is still fine.
define <4 x i32> @splat_extract_v4i32_nomask(<4 x i32> %a, i32 %b) nounwind {
-; LA32-LABEL: splat_extract_v4i32_nomask:
-; LA32: # %bb.0: # %entry
-; LA32-NEXT: vreplve.w $vr0, $vr0, $a0
-; LA32-NEXT: movfr2gr.s $a0, $fa0
-; LA32-NEXT: vreplgr2vr.w $vr0, $a0
-; LA32-NEXT: ret
-;
-; LA64-LABEL: splat_extract_v4i32_nomask:
-; LA64: # %bb.0: # %entry
-; LA64-NEXT: bstrpick.d $a0, $a0, 31, 0
-; LA64-NEXT: vreplve.w $vr0, $vr0, $a0
-; LA64-NEXT: movfr2gr.s $a0, $fa0
-; LA64-NEXT: vreplgr2vr.w $vr0, $a0
-; LA64-NEXT: ret
+; CHECK-LABEL: splat_extract_v4i32_nomask:
+; CHECK: # %bb.0: # %entry
+; CHECK-NEXT: vreplve.w $vr0, $vr0, $a0
+; CHECK-NEXT: ret
entry:
%e = extractelement <4 x i32> %a, i32 %b
%ins = insertelement <4 x i32> poison, i32 %e, i64 0
@@ -140,11 +112,7 @@ entry:
define <16 x i8> @splat_extract_v16i8_mask255(<16 x i8> %a, i32 %b) nounwind {
; CHECK-LABEL: splat_extract_v16i8_mask255:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 255
; CHECK-NEXT: vreplve.b $vr0, $vr0, $a0
-; CHECK-NEXT: movfr2gr.s $a0, $fa0
-; CHECK-NEXT: srai.w $a0, $a0, 24
-; CHECK-NEXT: vreplgr2vr.b $vr0, $a0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 255
@@ -160,9 +128,6 @@ define <16 x i8> @splat_extract_v16i8_mask7(<16 x i8> %a, i32 %b) nounwind {
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: andi $a0, $a0, 7
; CHECK-NEXT: vreplve.b $vr0, $vr0, $a0
-; CHECK-NEXT: movfr2gr.s $a0, $fa0
-; CHECK-NEXT: srai.w $a0, $a0, 24
-; CHECK-NEXT: vreplgr2vr.b $vr0, $a0
; CHECK-NEXT: ret
entry:
%idx = and i32 %b, 7
@@ -178,8 +143,6 @@ define <4 x i32> @splat_extract_v4i32_idx_multiuse(<4 x i32> %a, i32 %b, ptr %p)
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: andi $a0, $a0, 3
; CHECK-NEXT: vreplve.w $vr0, $vr0, $a0
-; CHECK-NEXT: movfr2gr.s $a2, $fa0
-; CHECK-NEXT: vreplgr2vr.w $vr0, $a2
; CHECK-NEXT: st.w $a0, $a1, 0
; CHECK-NEXT: ret
entry:
@@ -198,7 +161,6 @@ define <4 x i32> @splat_extract_v4i32_elt_multiuse(<4 x i32> %a, i32 %b, ptr %p)
; CHECK-NEXT: andi $a0, $a0, 3
; CHECK-NEXT: vreplve.w $vr0, $vr0, $a0
; CHECK-NEXT: movfr2gr.s $a0, $fa0
-; CHECK-NEXT: vreplgr2vr.w $vr0, $a0
; CHECK-NEXT: st.w $a0, $a1, 0
; CHECK-NEXT: ret
entry:
@@ -249,7 +211,6 @@ declare <16 x i8> @llvm.loongarch.lsx.vreplve.b(<16 x i8>, i32)
define <16 x i8> @intrinsic_vreplve_b_mask(<16 x i8> %a, i32 %b) nounwind {
; CHECK-LABEL: intrinsic_vreplve_b_mask:
; CHECK: # %bb.0: # %entry
-; CHECK-NEXT: andi $a0, $a0, 15
; CHECK-NEXT: vreplve.b $vr0, $vr0, $a0
; CHECK-NEXT: ret
entry:
More information about the llvm-branch-commits
mailing list