Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp =================================================================== --- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp +++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp @@ -11125,6 +11125,22 @@ return SDValue(N, 0); // Return N so it doesn't get rechecked! } + // fold (sext_inreg (masked_load x)) -> (sext_masked_load x) + // ignore it if the masked load is already sign extended + if (MaskedLoadSDNode *Ld = dyn_cast(N0)) { + if (ExtVT == Ld->getMemoryVT() && !LegalOperations && + Ld->getExtensionType() != ISD::LoadExtType::SEXTLOAD && N0.hasOneUse() && TLI.isLoadExtLegal(ISD::SEXTLOAD, VT, ExtVT)) { + SDValue ExtMaskedLoad = + DAG.getMaskedLoad(VT, SDLoc(N), Ld->getChain(), Ld->getBasePtr(), + Ld->getOffset(), Ld->getMask(), Ld->getPassThru(), + ExtVT, Ld->getMemOperand(), Ld->getAddressingMode(), + ISD::SEXTLOAD, Ld->isExpandingLoad()); + CombineTo(N, ExtMaskedLoad); + CombineTo(N0.getNode(), ExtMaskedLoad, ExtMaskedLoad.getValue(1)); + return SDValue(N, 0); // Return N so it doesn't get rechecked! + } + } + // Form (sext_inreg (bswap >> 16)) or (sext_inreg (rotl (bswap) 16)) if (ExtVTBits <= 16 && N0.getOpcode() == ISD::OR) { if (SDValue BSwap = MatchBSwapHWordLow(N0.getNode(), N0.getOperand(0), Index: llvm/test/CodeGen/Thumb2/mve-masked-load.ll =================================================================== --- llvm/test/CodeGen/Thumb2/mve-masked-load.ll +++ llvm/test/CodeGen/Thumb2/mve-masked-load.ll @@ -2148,11 +2148,11 @@ ; CHECK-LE-NEXT: .vsave {d8, d9} ; CHECK-LE-NEXT: vpush {d8, d9} ; CHECK-LE-NEXT: vpt.s32 gt, q0, zr -; CHECK-LE-NEXT: vldrht.u32 q4, [r0] +; CHECK-LE-NEXT: vldrht.s32 q4, [r0] ; CHECK-LE-NEXT: vmov r0, r1, d8 ; CHECK-LE-NEXT: vmov r2, r3, d9 ; CHECK-LE-NEXT: bl foo -; CHECK-LE-NEXT: vmovlb.s16 q0, q4 +; CHECK-LE-NEXT: vmov q0, q4 ; CHECK-LE-NEXT: vpop {d8, d9} ; CHECK-LE-NEXT: pop {r7, pc} ; @@ -2164,13 +2164,12 @@ ; CHECK-BE-NEXT: vpush {d8, d9} ; CHECK-BE-NEXT: vrev64.32 q1, q0 ; CHECK-BE-NEXT: vpt.s32 gt, q1, zr -; CHECK-BE-NEXT: vldrht.u32 q4, [r0] -; CHECK-BE-NEXT: vrev64.32 q0, q4 -; CHECK-BE-NEXT: vmov r1, r0, d0 -; CHECK-BE-NEXT: vmov r3, r2, d1 +; CHECK-BE-NEXT: vldrht.s32 q0, [r0] +; CHECK-BE-NEXT: vrev64.32 q4, q0 +; CHECK-BE-NEXT: vmov r1, r0, d8 +; CHECK-BE-NEXT: vmov r3, r2, d9 ; CHECK-BE-NEXT: bl foo -; CHECK-BE-NEXT: vmovlb.s16 q1, q4 -; CHECK-BE-NEXT: vrev64.32 q0, q1 +; CHECK-BE-NEXT: vmov q0, q4 ; CHECK-BE-NEXT: vpop {d8, d9} ; CHECK-BE-NEXT: pop {r7, pc} entry: Index: llvm/test/CodeGen/Thumb2/sext-masked-load.ll =================================================================== --- /dev/null +++ llvm/test/CodeGen/Thumb2/sext-masked-load.ll @@ -0,0 +1,19 @@ +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py +; RUN: llc -mtriple=thumbv8.1m.main-none-none-eabi -mattr=+mve.fp -verify-machineinstrs -tail-predication=enabled -o - %s | FileCheck %s + +define arm_aapcs_vfpcc <4 x float> @foo_v4i16(<4 x i16>* nocapture readonly %pSrc, i32 %blockSize, <4 x i16> %a) { +; CHECK-LABEL: foo_v4i16: +; CHECK: @ %bb.0: @ %entry +; CHECK-NEXT: vmovlb.s16 q0, q0 +; CHECK-NEXT: vpt.s32 lt, q0, zr +; CHECK-NEXT: vldrht.s32 q0, [r0] +; CHECK-NEXT: vcvt.f32.s32 q0, q0 +; CHECK-NEXT: bx lr +entry: + %active.lane.mask = icmp slt <4 x i16> %a, zeroinitializer + %wide.masked.load = call <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>* %pSrc, i32 2, <4 x i1> %active.lane.mask, <4 x i16> undef) + %0 = sitofp <4 x i16> %wide.masked.load to <4 x float> + ret <4 x float> %0 +} + +declare <4 x i16> @llvm.masked.load.v4i16.p0v4i16(<4 x i16>*, i32 immarg, <4 x i1>, <4 x i16>)