Index: llvm/trunk/lib/CodeGen/GlobalISel/LegalizerHelper.cpp =================================================================== --- llvm/trunk/lib/CodeGen/GlobalISel/LegalizerHelper.cpp +++ llvm/trunk/lib/CodeGen/GlobalISel/LegalizerHelper.cpp @@ -567,6 +567,26 @@ MI.eraseFromParent(); return Legalized; } + case TargetOpcode::G_SEXT: { + if (TypeIdx != 0) + return UnableToLegalize; + + if (NarrowTy.getSizeInBits() != SizeOp0 / 2) { + LLVM_DEBUG(dbgs() << "Can't narrow sext to type " << NarrowTy << "\n"); + return UnableToLegalize; + } + + Register SrcReg = MI.getOperand(1).getReg(); + + // Shift the sign bit of the low register through the high register. + auto ShiftAmt = + MIRBuilder.buildConstant(LLT::scalar(64), NarrowTy.getSizeInBits() - 1); + auto Shift = MIRBuilder.buildAShr(NarrowTy, SrcReg, ShiftAmt); + MIRBuilder.buildMerge(MI.getOperand(0).getReg(), {SrcReg, Shift.getReg(0)}); + MI.eraseFromParent(); + return Legalized; + } + case TargetOpcode::G_ADD: { // FIXME: add support for when SizeOp0 isn't an exact multiple of // NarrowSize. Index: llvm/trunk/lib/Target/AArch64/AArch64LegalizerInfo.cpp =================================================================== --- llvm/trunk/lib/Target/AArch64/AArch64LegalizerInfo.cpp +++ llvm/trunk/lib/Target/AArch64/AArch64LegalizerInfo.cpp @@ -340,30 +340,36 @@ .widenScalarToNextPow2(1); // Extensions - getActionDefinitionsBuilder({G_ZEXT, G_SEXT, G_ANYEXT}) - .legalIf([=](const LegalityQuery &Query) { - unsigned DstSize = Query.Types[0].getSizeInBits(); - - // Make sure that we have something that will fit in a register, and - // make sure it's a power of 2. - if (DstSize < 8 || DstSize > 128 || !isPowerOf2_32(DstSize)) - return false; + auto ExtLegalFunc = [=](const LegalityQuery &Query) { + unsigned DstSize = Query.Types[0].getSizeInBits(); - const LLT &SrcTy = Query.Types[1]; + if (DstSize == 128 && !Query.Types[0].isVector()) + return false; // Extending to a scalar s128 is not legal. + + // Make sure that we have something that will fit in a register, and + // make sure it's a power of 2. + if (DstSize < 8 || DstSize > 128 || !isPowerOf2_32(DstSize)) + return false; - // Special case for s1. - if (SrcTy == s1) - return true; + const LLT &SrcTy = Query.Types[1]; - // Make sure we fit in a register otherwise. Don't bother checking that - // the source type is below 128 bits. We shouldn't be allowing anything - // through which is wider than the destination in the first place. - unsigned SrcSize = SrcTy.getSizeInBits(); - if (SrcSize < 8 || !isPowerOf2_32(SrcSize)) - return false; + // Special case for s1. + if (SrcTy == s1) + return true; + + // Make sure we fit in a register otherwise. Don't bother checking that + // the source type is below 128 bits. We shouldn't be allowing anything + // through which is wider than the destination in the first place. + unsigned SrcSize = SrcTy.getSizeInBits(); + if (SrcSize < 8 || !isPowerOf2_32(SrcSize)) + return false; - return true; - }); + return true; + }; + getActionDefinitionsBuilder({G_ZEXT, G_ANYEXT}).legalIf(ExtLegalFunc); + getActionDefinitionsBuilder(G_SEXT) + .legalIf(ExtLegalFunc) + .clampScalar(0, s64, s64); // Just for s128, others are handled above. getActionDefinitionsBuilder(G_TRUNC).alwaysLegal(); Index: llvm/trunk/test/CodeGen/AArch64/GlobalISel/legalize-sext-128.mir =================================================================== --- llvm/trunk/test/CodeGen/AArch64/GlobalISel/legalize-sext-128.mir +++ llvm/trunk/test/CodeGen/AArch64/GlobalISel/legalize-sext-128.mir @@ -0,0 +1,25 @@ +# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py +# RUN: llc -march=aarch64 -run-pass=legalizer -verify-machineinstrs %s -o - | FileCheck %s +--- +name: narrow_s128 +tracksRegLiveness: true +body: | + bb.1: + liveins: $x0, $x1 + + ; CHECK-LABEL: name: narrow_s128 + ; CHECK: liveins: $x0, $x1 + ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0 + ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1 + ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 63 + ; CHECK: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[COPY]], [[C]](s64) + ; CHECK: [[MV:%[0-9]+]]:_(s128) = G_MERGE_VALUES [[COPY]](s64), [[ASHR]](s64) + ; CHECK: G_STORE [[MV]](s128), [[COPY1]](p0) :: (store 16) + ; CHECK: RET_ReallyLR + %0:_(s64) = COPY $x0 + %1:_(p0) = COPY $x1 + %2:_(s128) = G_SEXT %0(s64) + G_STORE %2(s128), %1(p0) :: (store 16) + RET_ReallyLR + +...