https://github.com/heiher created https://github.com/llvm/llvm-project/pull/244485
The Rust portable implementations of the [x]vdiv/[x]vmod intrinsics guard the divisor against zero and signed overflow, e.g. ``` %bz = icmp eq %b, 0 %guard = or %bz, (and (icmp eq %a, MIN), (icmp eq %b, -1)) %d = select %guard, splat (1), %b %q = sdiv %a, %d %r = select %bz, splat (0), %q ``` The hardware already returns 0 for x / 0 and x % 0, MIN for MIN / -1 and 0 for MIN % -1, so the guards are redundant. Fold them into new VDIV/VDIV_U/VMOD/VMOD_U nodes to emit a single [x]vdiv/[x]vmod. Assisted-by: LLM >From 653600c5adc3dca24681fa1ab14b7ce465522f7f Mon Sep 17 00:00:00 2001 From: WANG Rui <[email protected]> Date: Thu, 19 Nov 2026 22:20:49 +0800 Subject: [PATCH] [LoongArch] Remove redundant divisor guards from vector div/rem The Rust portable implementations of the [x]vdiv/[x]vmod intrinsics guard the divisor against zero and signed overflow, e.g. ``` %bz = icmp eq %b, 0 %guard = or %bz, (and (icmp eq %a, MIN), (icmp eq %b, -1)) %d = select %guard, splat (1), %b %q = sdiv %a, %d %r = select %bz, splat (0), %q ``` The hardware already returns 0 for x / 0 and x % 0, MIN for MIN / -1 and 0 for MIN % -1, so the guards are redundant. Fold them into new VDIV/VDIV_U/VMOD/VMOD_U nodes to emit a single [x]vdiv/[x]vmod. Assisted-by: LLM --- .../LoongArch/LoongArchISelLowering.cpp | 123 +++++++++++++ .../LoongArch/LoongArchLASXInstrInfo.td | 4 + .../Target/LoongArch/LoongArchLSXInstrInfo.td | 11 ++ .../CodeGen/LoongArch/lasx/guarded-div-rem.ll | 162 ++---------------- .../CodeGen/LoongArch/lsx/guarded-div-rem.ll | 162 ++---------------- 5 files changed, 170 insertions(+), 292 deletions(-) diff --git a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp index c263ed4570bde0..91e76ea6b93381 100644 --- a/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp +++ b/llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp @@ -546,6 +546,7 @@ LoongArchTargetLowering::LoongArchTargetLowering(const TargetMachine &TM, setTargetDAGCombine(ISD::ZERO_EXTEND); setTargetDAGCombine(ISD::SIGN_EXTEND); setTargetDAGCombine(ISD::MUL); + setTargetDAGCombine({ISD::SREM, ISD::UREM}); } // Set DAG combine for 'LASX' feature. @@ -9380,9 +9381,128 @@ performCONCAT_VECTORSCombine(SDNode *N, SelectionDAG &DAG, return SDValue(); } +// Match (seteq X, splat(C)) where C satisfies Pred. +static bool isSetEqSplat(SDValue V, SDValue X, + function_ref<bool(const APInt &)> Pred) { + if (V.getOpcode() != ISD::SETCC || + cast<CondCodeSDNode>(V.getOperand(2))->get() != ISD::SETEQ) + return false; + + SDValue LHS = V.getOperand(0); + SDValue RHS = V.getOperand(1); + if (RHS == X) + std::swap(LHS, RHS); + if (LHS != X) + return false; + + ConstantSDNode *C = isConstOrConstSplat(RHS, /*AllowUndefs=*/false, + /*AllowTruncation=*/true); + return C && Pred(C->getAPIntValue().trunc(X.getScalarValueSizeInBits())); +} + +// The [x]vdiv/[x]vmod instructions return x / 0 = 0, x % 0 = 0, +// MIN / -1 = MIN and MIN % -1 = 0. Return true if, in the lanes where Cond is +// set, dividing A by 1 gives the same result as the instruction with B. For +// division, the caller must zero the lanes where B is zero. +static bool isRedundantDivisorGuard(SDValue Cond, SDValue A, SDValue B, + unsigned Opc, unsigned Depth = 0) { + if (Depth >= 6) + return false; + + switch (Cond.getOpcode()) { + case ISD::OR: + return isRedundantDivisorGuard(Cond.getOperand(0), A, B, Opc, Depth + 1) && + isRedundantDivisorGuard(Cond.getOperand(1), A, B, Opc, Depth + 1); + case ISD::AND: { + SDValue Op0 = Cond.getOperand(0); + SDValue Op1 = Cond.getOperand(1); + if (isRedundantDivisorGuard(Op0, A, B, Opc, Depth + 1) || + isRedundantDivisorGuard(Op1, A, B, Opc, Depth + 1)) + return true; + + // MIN / -1 = MIN / 1. + auto IsOverflow = [&](SDValue X, SDValue Y) { + return isSetEqSplat( + X, A, [](const APInt &C) { return C.isMinSignedValue(); }) && + isSetEqSplat(Y, B, [](const APInt &C) { return C.isAllOnes(); }); + }; + return Opc == ISD::SDIV && (IsOverflow(Op0, Op1) || IsOverflow(Op1, Op0)); + } + default: + break; + } + + // x / 1 = x and x % 1 = x % 0 = 0. + if (isSetEqSplat(Cond, B, [](const APInt &C) { return C.isOne(); }) || + isSetEqSplat(Cond, B, [](const APInt &C) { return C.isZero(); })) + return true; + + // x % -1 = x % 1 = 0. + return Opc == ISD::SREM && + isSetEqSplat(Cond, B, [](const APInt &C) { return C.isAllOnes(); }); +} + +// Match B guarded by a redundant select or umax, and return B. +static SDValue matchGuardedDivisor(SDValue D, SDValue A, unsigned Opc) { + // umax(b, 1) == select(b == 0, 1, b). + if (D.getOpcode() == ISD::UMAX && isOneOrOneSplat(D.getOperand(1))) + return D.getOperand(0); + + if (D.getOpcode() == ISD::VSELECT && isOneOrOneSplat(D.getOperand(1)) && + isRedundantDivisorGuard(D.getOperand(0), A, D.getOperand(2), Opc)) + return D.getOperand(2); + + return SDValue(); +} + +// (srem/urem a, (guarded b)) -> (VMOD/VMOD_U a, b) +static SDValue performREMCombine(SDNode *N, SelectionDAG &DAG, + TargetLowering::DAGCombinerInfo &DCI, + const LoongArchSubtarget &Subtarget) { + EVT VT = N->getValueType(0); + if (!VT.isVector() || !DAG.getTargetLoweringInfo().isTypeLegal(VT)) + return SDValue(); + + unsigned Opc = N->getOpcode(); + SDValue A = N->getOperand(0); + SDValue B = matchGuardedDivisor(N->getOperand(1), A, Opc); + if (!B) + return SDValue(); + + return DAG.getNode(Opc == ISD::SREM ? LoongArchISD::VMOD + : LoongArchISD::VMOD_U, + SDLoc(N), VT, A, B); +} + +// (vselect (seteq b, 0), 0, (sdiv/udiv a, (guarded b))) -> (VDIV/VDIV_U a, b) +static SDValue combineGuardedVectorDiv(SDNode *N, SelectionDAG &DAG) { + EVT VT = N->getValueType(0); + if (!VT.isVector() || !DAG.getTargetLoweringInfo().isTypeLegal(VT)) + return SDValue(); + + SDValue Div = N->getOperand(2); + unsigned Opc = Div.getOpcode(); + if ((Opc != ISD::SDIV && Opc != ISD::UDIV) || !Div.hasOneUse() || + !ISD::isConstantSplatVectorAllZeros(N->getOperand(1))) + return SDValue(); + + SDValue A = Div.getOperand(0); + SDValue B = matchGuardedDivisor(Div.getOperand(1), A, Opc); + if (!B || !isSetEqSplat(N->getOperand(0), B, + [](const APInt &C) { return C.isZero(); })) + return SDValue(); + + return DAG.getNode(Opc == ISD::SDIV ? LoongArchISD::VDIV + : LoongArchISD::VDIV_U, + SDLoc(N), VT, A, B); +} + static SDValue performVSELECTCombine(SDNode *N, SelectionDAG &DAG, TargetLowering::DAGCombinerInfo &DCI, const LoongArchSubtarget &Subtarget) { + if (SDValue V = combineGuardedVectorDiv(N, DAG)) + return V; + if (DCI.isBeforeLegalizeOps()) return SDValue(); @@ -9864,6 +9984,9 @@ SDValue LoongArchTargetLowering::PerformDAGCombine(SDNode *N, break; case ISD::MUL: return performMULCombine(N, DAG, DCI, Subtarget); + case ISD::SREM: + case ISD::UREM: + return performREMCombine(N, DAG, DCI, Subtarget); } return SDValue(); } diff --git a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td index 79d897646d3bad..2ae9e62d3e2ad6 100644 --- a/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td +++ b/llvm/lib/Target/LoongArch/LoongArchLASXInstrInfo.td @@ -2103,10 +2103,14 @@ defm : PatXrXrXr<mulsub, "XVMSUB">; // XVDIV_{B/H/W/D}[U] defm : PatXrXr<sdiv, "XVDIV">; defm : PatXrXrU<udiv, "XVDIV">; +defm : PatXrXr<loongarch_vdiv, "XVDIV">; +defm : PatXrXrU<loongarch_vdiv_u, "XVDIV">; // XVMOD_{B/H/W/D}[U] defm : PatXrXr<srem, "XVMOD">; defm : PatXrXrU<urem, "XVMOD">; +defm : PatXrXr<loongarch_vmod, "XVMOD">; +defm : PatXrXrU<loongarch_vmod_u, "XVMOD">; // XVAND_V foreach vt = [v32i8, v16i16, v8i32, v4i64] in diff --git a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td index 3f60a13e433f0f..d04963a217f606 100644 --- a/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td +++ b/llvm/lib/Target/LoongArch/LoongArchLSXInstrInfo.td @@ -75,6 +75,13 @@ def loongarch_vilvl: SDNode<"LoongArchISD::VILVL", SDT_LoongArchV2R>; def loongarch_vilvh: SDNode<"LoongArchISD::VILVH", SDT_LoongArchV2R>; def loongarch_vandn: SDNode<"LoongArchISD::VANDN", SDT_LoongArchV2R>; +// Vector div/rem with defined results: x / 0 = 0, x % 0 = 0, MIN / -1 = MIN +// and MIN % -1 = 0. +def loongarch_vdiv : SDNode<"LoongArchISD::VDIV", SDT_LoongArchV2R>; +def loongarch_vdiv_u : SDNode<"LoongArchISD::VDIV_U", SDT_LoongArchV2R>; +def loongarch_vmod : SDNode<"LoongArchISD::VMOD", SDT_LoongArchV2R>; +def loongarch_vmod_u : SDNode<"LoongArchISD::VMOD_U", SDT_LoongArchV2R>; + def loongarch_vshuf4i: SDNode<"LoongArchISD::VSHUF4I", SDT_LoongArchV1RUimm>; def loongarch_vshuf4i_d : SDNode<"LoongArchISD::VSHUF4I_D", SDT_LoongArchV2RUimm>; def loongarch_vreplvei: SDNode<"LoongArchISD::VREPLVEI", SDT_LoongArchV1RUimm>; @@ -2351,10 +2358,14 @@ defm : PatVrVrVr<mulsub, "VMSUB">; // VDIV_{B/H/W/D}[U] defm : PatVrVr<sdiv, "VDIV">; defm : PatVrVrU<udiv, "VDIV">; +defm : PatVrVr<loongarch_vdiv, "VDIV">; +defm : PatVrVrU<loongarch_vdiv_u, "VDIV">; // VMOD_{B/H/W/D}[U] defm : PatVrVr<srem, "VMOD">; defm : PatVrVrU<urem, "VMOD">; +defm : PatVrVr<loongarch_vmod, "VMOD">; +defm : PatVrVrU<loongarch_vmod_u, "VMOD">; // VAND_V foreach vt = [v16i8, v8i16, v4i32, v2i64] in diff --git a/llvm/test/CodeGen/LoongArch/lasx/guarded-div-rem.ll b/llvm/test/CodeGen/LoongArch/lasx/guarded-div-rem.ll index ded1fac16a9849..550826b8104c8c 100644 --- a/llvm/test/CodeGen/LoongArch/lasx/guarded-div-rem.ll +++ b/llvm/test/CodeGen/LoongArch/lasx/guarded-div-rem.ll @@ -1,6 +1,6 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 -; RUN: llc --mtriple=loongarch32 --mattr=+32s,+lasx < %s | FileCheck %s --check-prefixes=CHECK,LA32 -; RUN: llc --mtriple=loongarch64 --mattr=+lasx < %s | FileCheck %s --check-prefixes=CHECK,LA64 +; RUN: llc --mtriple=loongarch32 --mattr=+32s,+lasx < %s | FileCheck %s +; RUN: llc --mtriple=loongarch64 --mattr=+lasx < %s | FileCheck %s ; The divisor guards are redundant, as the hardware returns x / 0 = 0, ; x % 0 = 0, MIN / -1 = MIN and MIN % -1 = 0. @@ -8,15 +8,7 @@ define <32 x i8> @lasx_xvdiv_b(<32 x i8> %a, <32 x i8> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_b: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.b $xr2, $xr1, 0 -; CHECK-NEXT: xvrepli.b $xr3, -128 -; CHECK-NEXT: xvseq.b $xr3, $xr0, $xr3 -; CHECK-NEXT: xvseqi.b $xr4, $xr1, -1 -; CHECK-NEXT: xvand.v $xr3, $xr3, $xr4 -; CHECK-NEXT: xvor.v $xr3, $xr2, $xr3 -; CHECK-NEXT: xvbitseli.b $xr3, $xr1, 1 -; CHECK-NEXT: xvdiv.b $xr0, $xr0, $xr3 -; CHECK-NEXT: xvandn.v $xr0, $xr2, $xr0 +; CHECK-NEXT: xvdiv.b $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <32 x i8> %b, zeroinitializer @@ -33,10 +25,7 @@ entry: define <32 x i8> @lasx_xvdiv_bu(<32 x i8> %a, <32 x i8> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_bu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.bu $xr2, $xr1, 1 -; CHECK-NEXT: xvdiv.bu $xr0, $xr0, $xr2 -; CHECK-NEXT: xvseqi.b $xr1, $xr1, 0 -; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0 +; CHECK-NEXT: xvdiv.bu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <32 x i8> %b, zeroinitializer @@ -49,14 +38,7 @@ entry: define <32 x i8> @lasx_xvmod_b(<32 x i8> %a, <32 x i8> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_b: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.b $xr2, $xr1, 0 -; CHECK-NEXT: xvrepli.b $xr3, -128 -; CHECK-NEXT: xvseq.b $xr3, $xr0, $xr3 -; CHECK-NEXT: xvseqi.b $xr4, $xr1, -1 -; CHECK-NEXT: xvand.v $xr3, $xr3, $xr4 -; CHECK-NEXT: xvor.v $xr2, $xr2, $xr3 -; CHECK-NEXT: xvbitseli.b $xr2, $xr1, 1 -; CHECK-NEXT: xvmod.b $xr0, $xr0, $xr2 +; CHECK-NEXT: xvmod.b $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <32 x i8> %b, zeroinitializer @@ -72,7 +54,6 @@ entry: define <32 x i8> @lasx_xvmod_bu(<32 x i8> %a, <32 x i8> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_bu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.bu $xr1, $xr1, 1 ; CHECK-NEXT: xvmod.bu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -84,16 +65,7 @@ entry: define <16 x i16> @lasx_xvdiv_h(<16 x i16> %a, <16 x i16> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_h: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.h $xr2, $xr1, 0 -; CHECK-NEXT: xvldi $xr3, -2688 -; CHECK-NEXT: xvseq.h $xr3, $xr0, $xr3 -; CHECK-NEXT: xvseqi.h $xr4, $xr1, -1 -; CHECK-NEXT: xvand.v $xr3, $xr3, $xr4 -; CHECK-NEXT: xvor.v $xr3, $xr2, $xr3 -; CHECK-NEXT: xvrepli.h $xr4, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr4, $xr3 ; CHECK-NEXT: xvdiv.h $xr0, $xr0, $xr1 -; CHECK-NEXT: xvandn.v $xr0, $xr2, $xr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <16 x i16> %b, zeroinitializer @@ -110,10 +82,7 @@ entry: define <16 x i16> @lasx_xvdiv_hu(<16 x i16> %a, <16 x i16> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_hu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.hu $xr2, $xr1, 1 -; CHECK-NEXT: xvdiv.hu $xr0, $xr0, $xr2 -; CHECK-NEXT: xvseqi.h $xr1, $xr1, 0 -; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0 +; CHECK-NEXT: xvdiv.hu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <16 x i16> %b, zeroinitializer @@ -126,14 +95,6 @@ entry: define <16 x i16> @lasx_xvmod_h(<16 x i16> %a, <16 x i16> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_h: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.h $xr2, $xr1, 0 -; CHECK-NEXT: xvldi $xr3, -2688 -; CHECK-NEXT: xvseq.h $xr3, $xr0, $xr3 -; CHECK-NEXT: xvseqi.h $xr4, $xr1, -1 -; CHECK-NEXT: xvand.v $xr3, $xr3, $xr4 -; CHECK-NEXT: xvor.v $xr2, $xr2, $xr3 -; CHECK-NEXT: xvrepli.h $xr3, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 ; CHECK-NEXT: xvmod.h $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -150,7 +111,6 @@ entry: define <16 x i16> @lasx_xvmod_hu(<16 x i16> %a, <16 x i16> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_hu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.hu $xr1, $xr1, 1 ; CHECK-NEXT: xvmod.hu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -162,16 +122,7 @@ entry: define <8 x i32> @lasx_xvdiv_w(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_w: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.w $xr2, $xr1, 0 -; CHECK-NEXT: xvldi $xr3, -3200 -; CHECK-NEXT: xvseq.w $xr3, $xr0, $xr3 -; CHECK-NEXT: xvseqi.w $xr4, $xr1, -1 -; CHECK-NEXT: xvand.v $xr3, $xr3, $xr4 -; CHECK-NEXT: xvor.v $xr3, $xr2, $xr3 -; CHECK-NEXT: xvrepli.w $xr4, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr4, $xr3 ; CHECK-NEXT: xvdiv.w $xr0, $xr0, $xr1 -; CHECK-NEXT: xvandn.v $xr0, $xr2, $xr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <8 x i32> %b, zeroinitializer @@ -188,10 +139,7 @@ entry: define <8 x i32> @lasx_xvdiv_wu(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_wu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.wu $xr2, $xr1, 1 -; CHECK-NEXT: xvdiv.wu $xr0, $xr0, $xr2 -; CHECK-NEXT: xvseqi.w $xr1, $xr1, 0 -; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0 +; CHECK-NEXT: xvdiv.wu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <8 x i32> %b, zeroinitializer @@ -204,14 +152,6 @@ entry: define <8 x i32> @lasx_xvmod_w(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_w: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.w $xr2, $xr1, 0 -; CHECK-NEXT: xvldi $xr3, -3200 -; CHECK-NEXT: xvseq.w $xr3, $xr0, $xr3 -; CHECK-NEXT: xvseqi.w $xr4, $xr1, -1 -; CHECK-NEXT: xvand.v $xr3, $xr3, $xr4 -; CHECK-NEXT: xvor.v $xr2, $xr2, $xr3 -; CHECK-NEXT: xvrepli.w $xr3, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 ; CHECK-NEXT: xvmod.w $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -228,7 +168,6 @@ entry: define <8 x i32> @lasx_xvmod_wu(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_wu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.wu $xr1, $xr1, 1 ; CHECK-NEXT: xvmod.wu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -238,35 +177,10 @@ entry: } define <4 x i64> @lasx_xvdiv_d(<4 x i64> %a, <4 x i64> %b) nounwind { -; LA32-LABEL: lasx_xvdiv_d: -; LA32: # %bb.0: # %entry -; LA32-NEXT: pcalau12i $a0, %pc_hi20(.LCPI12_0) -; LA32-NEXT: xvld $xr2, $a0, %pc_lo12(.LCPI12_0) -; LA32-NEXT: xvseqi.d $xr3, $xr1, 0 -; LA32-NEXT: xvseq.d $xr2, $xr0, $xr2 -; LA32-NEXT: xvseqi.d $xr4, $xr1, -1 -; LA32-NEXT: xvand.v $xr2, $xr2, $xr4 -; LA32-NEXT: xvor.v $xr2, $xr3, $xr2 -; LA32-NEXT: xvrepli.d $xr4, 1 -; LA32-NEXT: xvbitsel.v $xr1, $xr1, $xr4, $xr2 -; LA32-NEXT: xvdiv.d $xr0, $xr0, $xr1 -; LA32-NEXT: xvandn.v $xr0, $xr3, $xr0 -; LA32-NEXT: ret -; -; LA64-LABEL: lasx_xvdiv_d: -; LA64: # %bb.0: # %entry -; LA64-NEXT: xvseqi.d $xr2, $xr1, 0 -; LA64-NEXT: lu52i.d $a0, $zero, -2048 -; LA64-NEXT: xvreplgr2vr.d $xr3, $a0 -; LA64-NEXT: xvseq.d $xr3, $xr0, $xr3 -; LA64-NEXT: xvseqi.d $xr4, $xr1, -1 -; LA64-NEXT: xvand.v $xr3, $xr3, $xr4 -; LA64-NEXT: xvor.v $xr3, $xr2, $xr3 -; LA64-NEXT: xvrepli.d $xr4, 1 -; LA64-NEXT: xvbitsel.v $xr1, $xr1, $xr4, $xr3 -; LA64-NEXT: xvdiv.d $xr0, $xr0, $xr1 -; LA64-NEXT: xvandn.v $xr0, $xr2, $xr0 -; LA64-NEXT: ret +; CHECK-LABEL: lasx_xvdiv_d: +; CHECK: # %bb.0: # %entry +; CHECK-NEXT: xvdiv.d $xr0, $xr0, $xr1 +; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i64> %b, zeroinitializer %amin = icmp eq <4 x i64> %a, splat (i64 -9223372036854775808) @@ -282,10 +196,7 @@ entry: define <4 x i64> @lasx_xvdiv_du(<4 x i64> %a, <4 x i64> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_du: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.du $xr2, $xr1, 1 -; CHECK-NEXT: xvdiv.du $xr0, $xr0, $xr2 -; CHECK-NEXT: xvseqi.d $xr1, $xr1, 0 -; CHECK-NEXT: xvandn.v $xr0, $xr1, $xr0 +; CHECK-NEXT: xvdiv.du $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i64> %b, zeroinitializer @@ -296,33 +207,10 @@ entry: } define <4 x i64> @lasx_xvmod_d(<4 x i64> %a, <4 x i64> %b) nounwind { -; LA32-LABEL: lasx_xvmod_d: -; LA32: # %bb.0: # %entry -; LA32-NEXT: pcalau12i $a0, %pc_hi20(.LCPI14_0) -; LA32-NEXT: xvld $xr2, $a0, %pc_lo12(.LCPI14_0) -; LA32-NEXT: xvseqi.d $xr3, $xr1, 0 -; LA32-NEXT: xvseq.d $xr2, $xr0, $xr2 -; LA32-NEXT: xvseqi.d $xr4, $xr1, -1 -; LA32-NEXT: xvand.v $xr2, $xr2, $xr4 -; LA32-NEXT: xvor.v $xr2, $xr3, $xr2 -; LA32-NEXT: xvrepli.d $xr3, 1 -; LA32-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 -; LA32-NEXT: xvmod.d $xr0, $xr0, $xr1 -; LA32-NEXT: ret -; -; LA64-LABEL: lasx_xvmod_d: -; LA64: # %bb.0: # %entry -; LA64-NEXT: xvseqi.d $xr2, $xr1, 0 -; LA64-NEXT: lu52i.d $a0, $zero, -2048 -; LA64-NEXT: xvreplgr2vr.d $xr3, $a0 -; LA64-NEXT: xvseq.d $xr3, $xr0, $xr3 -; LA64-NEXT: xvseqi.d $xr4, $xr1, -1 -; LA64-NEXT: xvand.v $xr3, $xr3, $xr4 -; LA64-NEXT: xvor.v $xr2, $xr2, $xr3 -; LA64-NEXT: xvrepli.d $xr3, 1 -; LA64-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 -; LA64-NEXT: xvmod.d $xr0, $xr0, $xr1 -; LA64-NEXT: ret +; CHECK-LABEL: lasx_xvmod_d: +; CHECK: # %bb.0: # %entry +; CHECK-NEXT: xvmod.d $xr0, $xr0, $xr1 +; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i64> %b, zeroinitializer %amin = icmp eq <4 x i64> %a, splat (i64 -9223372036854775808) @@ -337,7 +225,6 @@ entry: define <4 x i64> @lasx_xvmod_du(<4 x i64> %a, <4 x i64> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_du: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.du $xr1, $xr1, 1 ; CHECK-NEXT: xvmod.du $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -350,11 +237,7 @@ entry: define <8 x i32> @lasx_xvdiv_w_select_zero_guard(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_w_select_zero_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.w $xr2, $xr1, 0 -; CHECK-NEXT: xvrepli.w $xr3, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 ; CHECK-NEXT: xvdiv.w $xr0, $xr0, $xr1 -; CHECK-NEXT: xvandn.v $xr0, $xr2, $xr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <8 x i32> %b, zeroinitializer @@ -367,11 +250,7 @@ entry: define <8 x i32> @lasx_xvdiv_wu_select_guard(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvdiv_wu_select_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.w $xr2, $xr1, 0 -; CHECK-NEXT: xvrepli.w $xr3, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 ; CHECK-NEXT: xvdiv.wu $xr0, $xr0, $xr1 -; CHECK-NEXT: xvandn.v $xr0, $xr2, $xr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <8 x i32> %b, zeroinitializer @@ -384,7 +263,6 @@ entry: define <8 x i32> @lasx_xvmod_w_umax_guard(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_w_umax_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvmaxi.wu $xr1, $xr1, 1 ; CHECK-NEXT: xvmod.w $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -397,11 +275,6 @@ entry: define <8 x i32> @lasx_xvmod_w_minus_one_guard(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_w_minus_one_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.w $xr2, $xr1, 0 -; CHECK-NEXT: xvseqi.w $xr3, $xr1, -1 -; CHECK-NEXT: xvor.v $xr2, $xr2, $xr3 -; CHECK-NEXT: xvrepli.w $xr3, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 ; CHECK-NEXT: xvmod.w $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: @@ -416,9 +289,6 @@ entry: define <8 x i32> @lasx_xvmod_wu_select_guard(<8 x i32> %a, <8 x i32> %b) nounwind { ; CHECK-LABEL: lasx_xvmod_wu_select_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: xvseqi.w $xr2, $xr1, 0 -; CHECK-NEXT: xvrepli.w $xr3, 1 -; CHECK-NEXT: xvbitsel.v $xr1, $xr1, $xr3, $xr2 ; CHECK-NEXT: xvmod.wu $xr0, $xr0, $xr1 ; CHECK-NEXT: ret entry: diff --git a/llvm/test/CodeGen/LoongArch/lsx/guarded-div-rem.ll b/llvm/test/CodeGen/LoongArch/lsx/guarded-div-rem.ll index 7ca8714656ffd5..1b1301c5b84b6e 100644 --- a/llvm/test/CodeGen/LoongArch/lsx/guarded-div-rem.ll +++ b/llvm/test/CodeGen/LoongArch/lsx/guarded-div-rem.ll @@ -1,6 +1,6 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 -; RUN: llc --mtriple=loongarch32 --mattr=+32s,+lsx < %s | FileCheck %s --check-prefixes=CHECK,LA32 -; RUN: llc --mtriple=loongarch64 --mattr=+lsx < %s | FileCheck %s --check-prefixes=CHECK,LA64 +; RUN: llc --mtriple=loongarch32 --mattr=+32s,+lsx < %s | FileCheck %s +; RUN: llc --mtriple=loongarch64 --mattr=+lsx < %s | FileCheck %s ; The divisor guards are redundant, as the hardware returns x / 0 = 0, ; x % 0 = 0, MIN / -1 = MIN and MIN % -1 = 0. @@ -8,15 +8,7 @@ define <16 x i8> @lsx_vdiv_b(<16 x i8> %a, <16 x i8> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_b: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.b $vr2, $vr1, 0 -; CHECK-NEXT: vrepli.b $vr3, -128 -; CHECK-NEXT: vseq.b $vr3, $vr0, $vr3 -; CHECK-NEXT: vseqi.b $vr4, $vr1, -1 -; CHECK-NEXT: vand.v $vr3, $vr3, $vr4 -; CHECK-NEXT: vor.v $vr3, $vr2, $vr3 -; CHECK-NEXT: vbitseli.b $vr3, $vr1, 1 -; CHECK-NEXT: vdiv.b $vr0, $vr0, $vr3 -; CHECK-NEXT: vandn.v $vr0, $vr2, $vr0 +; CHECK-NEXT: vdiv.b $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <16 x i8> %b, zeroinitializer @@ -33,10 +25,7 @@ entry: define <16 x i8> @lsx_vdiv_bu(<16 x i8> %a, <16 x i8> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_bu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.bu $vr2, $vr1, 1 -; CHECK-NEXT: vdiv.bu $vr0, $vr0, $vr2 -; CHECK-NEXT: vseqi.b $vr1, $vr1, 0 -; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0 +; CHECK-NEXT: vdiv.bu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <16 x i8> %b, zeroinitializer @@ -49,14 +38,7 @@ entry: define <16 x i8> @lsx_vmod_b(<16 x i8> %a, <16 x i8> %b) nounwind { ; CHECK-LABEL: lsx_vmod_b: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.b $vr2, $vr1, 0 -; CHECK-NEXT: vrepli.b $vr3, -128 -; CHECK-NEXT: vseq.b $vr3, $vr0, $vr3 -; CHECK-NEXT: vseqi.b $vr4, $vr1, -1 -; CHECK-NEXT: vand.v $vr3, $vr3, $vr4 -; CHECK-NEXT: vor.v $vr2, $vr2, $vr3 -; CHECK-NEXT: vbitseli.b $vr2, $vr1, 1 -; CHECK-NEXT: vmod.b $vr0, $vr0, $vr2 +; CHECK-NEXT: vmod.b $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <16 x i8> %b, zeroinitializer @@ -72,7 +54,6 @@ entry: define <16 x i8> @lsx_vmod_bu(<16 x i8> %a, <16 x i8> %b) nounwind { ; CHECK-LABEL: lsx_vmod_bu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.bu $vr1, $vr1, 1 ; CHECK-NEXT: vmod.bu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -84,16 +65,7 @@ entry: define <8 x i16> @lsx_vdiv_h(<8 x i16> %a, <8 x i16> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_h: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.h $vr2, $vr1, 0 -; CHECK-NEXT: vldi $vr3, -2688 -; CHECK-NEXT: vseq.h $vr3, $vr0, $vr3 -; CHECK-NEXT: vseqi.h $vr4, $vr1, -1 -; CHECK-NEXT: vand.v $vr3, $vr3, $vr4 -; CHECK-NEXT: vor.v $vr3, $vr2, $vr3 -; CHECK-NEXT: vrepli.h $vr4, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr4, $vr3 ; CHECK-NEXT: vdiv.h $vr0, $vr0, $vr1 -; CHECK-NEXT: vandn.v $vr0, $vr2, $vr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <8 x i16> %b, zeroinitializer @@ -110,10 +82,7 @@ entry: define <8 x i16> @lsx_vdiv_hu(<8 x i16> %a, <8 x i16> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_hu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.hu $vr2, $vr1, 1 -; CHECK-NEXT: vdiv.hu $vr0, $vr0, $vr2 -; CHECK-NEXT: vseqi.h $vr1, $vr1, 0 -; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0 +; CHECK-NEXT: vdiv.hu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <8 x i16> %b, zeroinitializer @@ -126,14 +95,6 @@ entry: define <8 x i16> @lsx_vmod_h(<8 x i16> %a, <8 x i16> %b) nounwind { ; CHECK-LABEL: lsx_vmod_h: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.h $vr2, $vr1, 0 -; CHECK-NEXT: vldi $vr3, -2688 -; CHECK-NEXT: vseq.h $vr3, $vr0, $vr3 -; CHECK-NEXT: vseqi.h $vr4, $vr1, -1 -; CHECK-NEXT: vand.v $vr3, $vr3, $vr4 -; CHECK-NEXT: vor.v $vr2, $vr2, $vr3 -; CHECK-NEXT: vrepli.h $vr3, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 ; CHECK-NEXT: vmod.h $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -150,7 +111,6 @@ entry: define <8 x i16> @lsx_vmod_hu(<8 x i16> %a, <8 x i16> %b) nounwind { ; CHECK-LABEL: lsx_vmod_hu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.hu $vr1, $vr1, 1 ; CHECK-NEXT: vmod.hu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -162,16 +122,7 @@ entry: define <4 x i32> @lsx_vdiv_w(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_w: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.w $vr2, $vr1, 0 -; CHECK-NEXT: vldi $vr3, -3200 -; CHECK-NEXT: vseq.w $vr3, $vr0, $vr3 -; CHECK-NEXT: vseqi.w $vr4, $vr1, -1 -; CHECK-NEXT: vand.v $vr3, $vr3, $vr4 -; CHECK-NEXT: vor.v $vr3, $vr2, $vr3 -; CHECK-NEXT: vrepli.w $vr4, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr4, $vr3 ; CHECK-NEXT: vdiv.w $vr0, $vr0, $vr1 -; CHECK-NEXT: vandn.v $vr0, $vr2, $vr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i32> %b, zeroinitializer @@ -188,10 +139,7 @@ entry: define <4 x i32> @lsx_vdiv_wu(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_wu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.wu $vr2, $vr1, 1 -; CHECK-NEXT: vdiv.wu $vr0, $vr0, $vr2 -; CHECK-NEXT: vseqi.w $vr1, $vr1, 0 -; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0 +; CHECK-NEXT: vdiv.wu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i32> %b, zeroinitializer @@ -204,14 +152,6 @@ entry: define <4 x i32> @lsx_vmod_w(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vmod_w: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.w $vr2, $vr1, 0 -; CHECK-NEXT: vldi $vr3, -3200 -; CHECK-NEXT: vseq.w $vr3, $vr0, $vr3 -; CHECK-NEXT: vseqi.w $vr4, $vr1, -1 -; CHECK-NEXT: vand.v $vr3, $vr3, $vr4 -; CHECK-NEXT: vor.v $vr2, $vr2, $vr3 -; CHECK-NEXT: vrepli.w $vr3, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 ; CHECK-NEXT: vmod.w $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -228,7 +168,6 @@ entry: define <4 x i32> @lsx_vmod_wu(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vmod_wu: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.wu $vr1, $vr1, 1 ; CHECK-NEXT: vmod.wu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -238,35 +177,10 @@ entry: } define <2 x i64> @lsx_vdiv_d(<2 x i64> %a, <2 x i64> %b) nounwind { -; LA32-LABEL: lsx_vdiv_d: -; LA32: # %bb.0: # %entry -; LA32-NEXT: pcalau12i $a0, %pc_hi20(.LCPI12_0) -; LA32-NEXT: vld $vr2, $a0, %pc_lo12(.LCPI12_0) -; LA32-NEXT: vseqi.d $vr3, $vr1, 0 -; LA32-NEXT: vseq.d $vr2, $vr0, $vr2 -; LA32-NEXT: vseqi.d $vr4, $vr1, -1 -; LA32-NEXT: vand.v $vr2, $vr2, $vr4 -; LA32-NEXT: vor.v $vr2, $vr3, $vr2 -; LA32-NEXT: vrepli.d $vr4, 1 -; LA32-NEXT: vbitsel.v $vr1, $vr1, $vr4, $vr2 -; LA32-NEXT: vdiv.d $vr0, $vr0, $vr1 -; LA32-NEXT: vandn.v $vr0, $vr3, $vr0 -; LA32-NEXT: ret -; -; LA64-LABEL: lsx_vdiv_d: -; LA64: # %bb.0: # %entry -; LA64-NEXT: vseqi.d $vr2, $vr1, 0 -; LA64-NEXT: lu52i.d $a0, $zero, -2048 -; LA64-NEXT: vreplgr2vr.d $vr3, $a0 -; LA64-NEXT: vseq.d $vr3, $vr0, $vr3 -; LA64-NEXT: vseqi.d $vr4, $vr1, -1 -; LA64-NEXT: vand.v $vr3, $vr3, $vr4 -; LA64-NEXT: vor.v $vr3, $vr2, $vr3 -; LA64-NEXT: vrepli.d $vr4, 1 -; LA64-NEXT: vbitsel.v $vr1, $vr1, $vr4, $vr3 -; LA64-NEXT: vdiv.d $vr0, $vr0, $vr1 -; LA64-NEXT: vandn.v $vr0, $vr2, $vr0 -; LA64-NEXT: ret +; CHECK-LABEL: lsx_vdiv_d: +; CHECK: # %bb.0: # %entry +; CHECK-NEXT: vdiv.d $vr0, $vr0, $vr1 +; CHECK-NEXT: ret entry: %bz = icmp eq <2 x i64> %b, zeroinitializer %amin = icmp eq <2 x i64> %a, splat (i64 -9223372036854775808) @@ -282,10 +196,7 @@ entry: define <2 x i64> @lsx_vdiv_du(<2 x i64> %a, <2 x i64> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_du: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.du $vr2, $vr1, 1 -; CHECK-NEXT: vdiv.du $vr0, $vr0, $vr2 -; CHECK-NEXT: vseqi.d $vr1, $vr1, 0 -; CHECK-NEXT: vandn.v $vr0, $vr1, $vr0 +; CHECK-NEXT: vdiv.du $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: %bz = icmp eq <2 x i64> %b, zeroinitializer @@ -296,33 +207,10 @@ entry: } define <2 x i64> @lsx_vmod_d(<2 x i64> %a, <2 x i64> %b) nounwind { -; LA32-LABEL: lsx_vmod_d: -; LA32: # %bb.0: # %entry -; LA32-NEXT: pcalau12i $a0, %pc_hi20(.LCPI14_0) -; LA32-NEXT: vld $vr2, $a0, %pc_lo12(.LCPI14_0) -; LA32-NEXT: vseqi.d $vr3, $vr1, 0 -; LA32-NEXT: vseq.d $vr2, $vr0, $vr2 -; LA32-NEXT: vseqi.d $vr4, $vr1, -1 -; LA32-NEXT: vand.v $vr2, $vr2, $vr4 -; LA32-NEXT: vor.v $vr2, $vr3, $vr2 -; LA32-NEXT: vrepli.d $vr3, 1 -; LA32-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 -; LA32-NEXT: vmod.d $vr0, $vr0, $vr1 -; LA32-NEXT: ret -; -; LA64-LABEL: lsx_vmod_d: -; LA64: # %bb.0: # %entry -; LA64-NEXT: vseqi.d $vr2, $vr1, 0 -; LA64-NEXT: lu52i.d $a0, $zero, -2048 -; LA64-NEXT: vreplgr2vr.d $vr3, $a0 -; LA64-NEXT: vseq.d $vr3, $vr0, $vr3 -; LA64-NEXT: vseqi.d $vr4, $vr1, -1 -; LA64-NEXT: vand.v $vr3, $vr3, $vr4 -; LA64-NEXT: vor.v $vr2, $vr2, $vr3 -; LA64-NEXT: vrepli.d $vr3, 1 -; LA64-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 -; LA64-NEXT: vmod.d $vr0, $vr0, $vr1 -; LA64-NEXT: ret +; CHECK-LABEL: lsx_vmod_d: +; CHECK: # %bb.0: # %entry +; CHECK-NEXT: vmod.d $vr0, $vr0, $vr1 +; CHECK-NEXT: ret entry: %bz = icmp eq <2 x i64> %b, zeroinitializer %amin = icmp eq <2 x i64> %a, splat (i64 -9223372036854775808) @@ -337,7 +225,6 @@ entry: define <2 x i64> @lsx_vmod_du(<2 x i64> %a, <2 x i64> %b) nounwind { ; CHECK-LABEL: lsx_vmod_du: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.du $vr1, $vr1, 1 ; CHECK-NEXT: vmod.du $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -350,11 +237,7 @@ entry: define <4 x i32> @lsx_vdiv_w_select_zero_guard(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_w_select_zero_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.w $vr2, $vr1, 0 -; CHECK-NEXT: vrepli.w $vr3, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 ; CHECK-NEXT: vdiv.w $vr0, $vr0, $vr1 -; CHECK-NEXT: vandn.v $vr0, $vr2, $vr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i32> %b, zeroinitializer @@ -367,11 +250,7 @@ entry: define <4 x i32> @lsx_vdiv_wu_select_guard(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vdiv_wu_select_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.w $vr2, $vr1, 0 -; CHECK-NEXT: vrepli.w $vr3, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 ; CHECK-NEXT: vdiv.wu $vr0, $vr0, $vr1 -; CHECK-NEXT: vandn.v $vr0, $vr2, $vr0 ; CHECK-NEXT: ret entry: %bz = icmp eq <4 x i32> %b, zeroinitializer @@ -384,7 +263,6 @@ entry: define <4 x i32> @lsx_vmod_w_umax_guard(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vmod_w_umax_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vmaxi.wu $vr1, $vr1, 1 ; CHECK-NEXT: vmod.w $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -397,11 +275,6 @@ entry: define <4 x i32> @lsx_vmod_w_minus_one_guard(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vmod_w_minus_one_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.w $vr2, $vr1, 0 -; CHECK-NEXT: vseqi.w $vr3, $vr1, -1 -; CHECK-NEXT: vor.v $vr2, $vr2, $vr3 -; CHECK-NEXT: vrepli.w $vr3, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 ; CHECK-NEXT: vmod.w $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: @@ -416,9 +289,6 @@ entry: define <4 x i32> @lsx_vmod_wu_select_guard(<4 x i32> %a, <4 x i32> %b) nounwind { ; CHECK-LABEL: lsx_vmod_wu_select_guard: ; CHECK: # %bb.0: # %entry -; CHECK-NEXT: vseqi.w $vr2, $vr1, 0 -; CHECK-NEXT: vrepli.w $vr3, 1 -; CHECK-NEXT: vbitsel.v $vr1, $vr1, $vr3, $vr2 ; CHECK-NEXT: vmod.wu $vr0, $vr0, $vr1 ; CHECK-NEXT: ret entry: _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
