Skip to content

[RISCV] Improve vmsge and vmsgeu selection #115435

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Nov 13, 2024
Merged

[RISCV] Improve vmsge and vmsgeu selection #115435

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
88f3ddd
[RISCV] Improve vmsge and vmsgeu selection
jacquesguan Nov 8, 2024
ec4ce89
Address comment
jacquesguan Nov 12, 2024
c1eb410
fix vmsge
jacquesguan Nov 13, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
65 changes: 51 additions & 14 deletions llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1616,28 +1616,37 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
SDValue Src1 = Node->getOperand(1);
SDValue Src2 = Node->getOperand(2);
bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu;
bool IsCmpUnsignedZero = false;
bool IsCmpConstant = false;
bool IsCmpMinimum = false;
// Only custom select scalar second operand.
if (Src2.getValueType() != XLenVT)
break;
// Small constants are handled with patterns.
int64_t CVal = 0;
MVT Src1VT = Src1.getSimpleValueType();
if (auto *C = dyn_cast<ConstantSDNode>(Src2)) {
int64_t CVal = C->getSExtValue();
IsCmpConstant = true;
CVal = C->getSExtValue();
if (CVal >= -15 && CVal <= 16) {
if (!IsUnsigned || CVal != 0)
break;
IsCmpUnsignedZero = true;
IsCmpMinimum = true;
} else if (!IsUnsigned && CVal == APInt::getSignedMinValue(
Src1VT.getScalarSizeInBits())
.getSExtValue()) {
IsCmpMinimum = true;
}
}
MVT Src1VT = Src1.getSimpleValueType();
unsigned VMSLTOpcode, VMNANDOpcode, VMSetOpcode;
unsigned VMSLTOpcode, VMNANDOpcode, VMSetOpcode, VMSGTOpcode;
switch (RISCVTargetLowering::getLMUL(Src1VT)) {
default:
llvm_unreachable("Unexpected LMUL!");
#define CASE_VMSLT_VMNAND_VMSET_OPCODES(lmulenum, suffix, suffix_b) \
case RISCVII::VLMUL::lmulenum: \
VMSLTOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix \
: RISCV::PseudoVMSLT_VX_##suffix; \
VMSGTOpcode = IsUnsigned ? RISCV::PseudoVMSGTU_VX_##suffix \
: RISCV::PseudoVMSGT_VX_##suffix; \
VMNANDOpcode = RISCV::PseudoVMNAND_MM_##suffix; \
VMSetOpcode = RISCV::PseudoVMSET_M_##suffix_b; \
break;
Expand All @@ -1655,12 +1664,21 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
SDValue VL;
selectVLOp(Node->getOperand(3), VL);

// If vmsgeu with 0 immediate, expand it to vmset.
if (IsCmpUnsignedZero) {
// If vmsge(u) with minimum value, expand it to vmset.
if (IsCmpMinimum) {
ReplaceNode(Node, CurDAG->getMachineNode(VMSetOpcode, DL, VT, VL, SEW));
return;
}

if (IsCmpConstant) {
SDValue Imm =
selectImm(CurDAG, SDLoc(Src2), XLenVT, CVal - 1, *Subtarget);

ReplaceNode(Node, CurDAG->getMachineNode(VMSGTOpcode, DL, VT,
{Src1, Imm, VL, SEW}));
return;
}

// Expand to
// vmslt{u}.vx vd, va, x; vmnand.mm vd, vd, vd
SDValue Cmp = SDValue(
Expand All @@ -1675,22 +1693,29 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
SDValue Src1 = Node->getOperand(2);
SDValue Src2 = Node->getOperand(3);
bool IsUnsigned = IntNo == Intrinsic::riscv_vmsgeu_mask;
bool IsCmpUnsignedZero = false;
bool IsCmpConstant = false;
bool IsCmpMinimum = false;
// Only custom select scalar second operand.
if (Src2.getValueType() != XLenVT)
break;
// Small constants are handled with patterns.
MVT Src1VT = Src1.getSimpleValueType();
int64_t CVal = 0;
if (auto *C = dyn_cast<ConstantSDNode>(Src2)) {
int64_t CVal = C->getSExtValue();
IsCmpConstant = true;
CVal = C->getSExtValue();
if (CVal >= -15 && CVal <= 16) {
if (!IsUnsigned || CVal != 0)
break;
IsCmpUnsignedZero = true;
IsCmpMinimum = true;
} else if (!IsUnsigned && CVal == APInt::getSignedMinValue(
Src1VT.getScalarSizeInBits())
.getSExtValue()) {
IsCmpMinimum = true;
}
}
MVT Src1VT = Src1.getSimpleValueType();
unsigned VMSLTOpcode, VMSLTMaskOpcode, VMXOROpcode, VMANDNOpcode,
VMOROpcode;
VMOROpcode, VMSGTMaskOpcode;
switch (RISCVTargetLowering::getLMUL(Src1VT)) {
default:
llvm_unreachable("Unexpected LMUL!");
Expand All @@ -1700,6 +1725,8 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
: RISCV::PseudoVMSLT_VX_##suffix; \
VMSLTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSLTU_VX_##suffix##_MASK \
: RISCV::PseudoVMSLT_VX_##suffix##_MASK; \
VMSGTMaskOpcode = IsUnsigned ? RISCV::PseudoVMSGTU_VX_##suffix##_MASK \
: RISCV::PseudoVMSGT_VX_##suffix##_MASK; \
break;
CASE_VMSLT_OPCODES(LMUL_F8, MF8, B1)
CASE_VMSLT_OPCODES(LMUL_F4, MF4, B2)
Expand Down Expand Up @@ -1737,8 +1764,8 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
SDValue MaskedOff = Node->getOperand(1);
SDValue Mask = Node->getOperand(4);

// If vmsgeu_mask with 0 immediate, expand it to vmor mask, maskedoff.
if (IsCmpUnsignedZero) {
// If vmsge(u) with minimum value, expand it to vmor mask, maskedoff.
if (IsCmpMinimum) {
// We don't need vmor if the MaskedOff and the Mask are the same
// value.
if (Mask == MaskedOff) {
Expand Down Expand Up @@ -1769,6 +1796,16 @@ void RISCVDAGToDAGISel::Select(SDNode *Node) {
SDValue Glue = Chain.getValue(1);
SDValue V0 = CurDAG->getRegister(RISCV::V0, VT);

if (IsCmpConstant) {
SDValue Imm =
selectImm(CurDAG, SDLoc(Src2), XLenVT, CVal - 1, *Subtarget);

ReplaceNode(Node, CurDAG->getMachineNode(
VMSGTMaskOpcode, DL, VT,
{MaskedOff, Src1, Imm, V0, VL, SEW, Glue}));
return;
}

// Otherwise use
// vmslt{u}.vx vd, va, x, v0.t; vmxor.mm vd, vd, v0
// The result is mask undisturbed.
Expand Down
70 changes: 70 additions & 0 deletions llvm/test/CodeGen/RISCV/rvv/vmsge.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1084,6 +1084,45 @@ entry:
ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsge_mask_vx_nxv4i8_i8_1(<vscale x 4 x i1> %0, <vscale x 4 x i8> %1, <vscale x 4 x i1> %2, iXLen %3) nounwind {
; CHECK-LABEL: intrinsic_vmsge_mask_vx_nxv4i8_i8_1:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: vmv1r.v v10, v0
; CHECK-NEXT: li a1, 99
; CHECK-NEXT: vmv1r.v v0, v9
; CHECK-NEXT: vsetvli zero, a0, e8, mf2, ta, mu
; CHECK-NEXT: vmsgt.vx v10, v8, a1, v0.t
; CHECK-NEXT: vmv1r.v v0, v10
; CHECK-NEXT: ret
entry:
%a = call <vscale x 4 x i1> @llvm.riscv.vmsge.mask.nxv4i8.i8(
<vscale x 4 x i1> %0,
<vscale x 4 x i8> %1,
i8 100,
<vscale x 4 x i1> %2,
iXLen %3)

ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsge_mask_vx_nxv4i8_i8_2(<vscale x 4 x i1> %0, <vscale x 4 x i8> %1, <vscale x 4 x i1> %2, iXLen %3) nounwind {
; CHECK-LABEL: intrinsic_vmsge_mask_vx_nxv4i8_i8_2:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: vsetvli zero, a0, e8, mf2, ta, ma
; CHECK-NEXT: vmor.mm v0, v9, v0
; CHECK-NEXT: ret
entry:
%a = call <vscale x 4 x i1> @llvm.riscv.vmsge.mask.nxv4i8.i8(
<vscale x 4 x i1> %0,
<vscale x 4 x i8> %1,
i8 -128,
<vscale x 4 x i1> %2,
iXLen %3)

ret <vscale x 4 x i1> %a
}


declare <vscale x 8 x i1> @llvm.riscv.vmsge.nxv8i8.i8(
<vscale x 8 x i8>,
i8,
Expand Down Expand Up @@ -1970,6 +2009,37 @@ entry:
ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsge_vi_nxv4i8_i8_1(<vscale x 4 x i8> %0, iXLen %1) nounwind {
; CHECK-LABEL: intrinsic_vmsge_vi_nxv4i8_i8_1:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: li a1, 99
; CHECK-NEXT: vsetvli zero, a0, e8, mf2, ta, ma
; CHECK-NEXT: vmsgt.vx v0, v8, a1
; CHECK-NEXT: ret
entry:
%a = call <vscale x 4 x i1> @llvm.riscv.vmsge.nxv4i8.i8(
<vscale x 4 x i8> %0,
i8 100,
iXLen %1)

ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsge_vi_nxv4i8_i8_2(<vscale x 4 x i8> %0, iXLen %1) nounwind {
; CHECK-LABEL: intrinsic_vmsge_vi_nxv4i8_i8_2:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: vsetvli zero, a0, e8, mf2, ta, ma
; CHECK-NEXT: vmset.m v0
; CHECK-NEXT: ret
entry:
%a = call <vscale x 4 x i1> @llvm.riscv.vmsge.nxv4i8.i8(
<vscale x 4 x i8> %0,
i8 -128,
iXLen %1)

ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsge_mask_vi_nxv4i8_i8(<vscale x 4 x i1> %0, <vscale x 4 x i8> %1, <vscale x 4 x i1> %2, iXLen %3) nounwind {
; CHECK-LABEL: intrinsic_vmsge_mask_vi_nxv4i8_i8:
; CHECK: # %bb.0: # %entry
Expand Down
37 changes: 37 additions & 0 deletions llvm/test/CodeGen/RISCV/rvv/vmsgeu.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1990,6 +1990,27 @@ entry:
ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsgeu_mask_vi_nxv4i8_i8_1(<vscale x 4 x i1> %0, <vscale x 4 x i8> %1, <vscale x 4 x i1> %2, iXLen %3) nounwind {
; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv4i8_i8_1:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: vmv1r.v v10, v0
; CHECK-NEXT: li a1, 99
; CHECK-NEXT: vmv1r.v v0, v9
; CHECK-NEXT: vsetvli zero, a0, e8, mf2, ta, mu
; CHECK-NEXT: vmsgtu.vx v10, v8, a1, v0.t
; CHECK-NEXT: vmv1r.v v0, v10
; CHECK-NEXT: ret
entry:
%a = call <vscale x 4 x i1> @llvm.riscv.vmsgeu.mask.nxv4i8.i8(
<vscale x 4 x i1> %0,
<vscale x 4 x i8> %1,
i8 100,
<vscale x 4 x i1> %2,
iXLen %3)

ret <vscale x 4 x i1> %a
}

define <vscale x 8 x i1> @intrinsic_vmsgeu_vi_nxv8i8_i8(<vscale x 8 x i8> %0, iXLen %1) nounwind {
; CHECK-LABEL: intrinsic_vmsgeu_vi_nxv8i8_i8:
; CHECK: # %bb.0: # %entry
Expand Down Expand Up @@ -2192,6 +2213,22 @@ entry:
ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsgeu_vi_nxv4i16_i16_1(<vscale x 4 x i16> %0, iXLen %1) nounwind {
; CHECK-LABEL: intrinsic_vmsgeu_vi_nxv4i16_i16_1:
; CHECK: # %bb.0: # %entry
; CHECK-NEXT: li a1, 99
; CHECK-NEXT: vsetvli zero, a0, e16, m1, ta, ma
; CHECK-NEXT: vmsgtu.vx v0, v8, a1
; CHECK-NEXT: ret
entry:
%a = call <vscale x 4 x i1> @llvm.riscv.vmsgeu.nxv4i16.i16(
<vscale x 4 x i16> %0,
i16 100,
iXLen %1)

ret <vscale x 4 x i1> %a
}

define <vscale x 4 x i1> @intrinsic_vmsgeu_mask_vi_nxv4i16_i16(<vscale x 4 x i1> %0, <vscale x 4 x i16> %1, <vscale x 4 x i1> %2, iXLen %3) nounwind {
; CHECK-LABEL: intrinsic_vmsgeu_mask_vi_nxv4i16_i16:
; CHECK: # %bb.0: # %entry
Expand Down
Loading