+// Group template arguments that can be derived from the vector type (EltNum x
+// EltVT). These are things like the register class for the writemask, etc.
+// The idea is to pass one of these as the template argument rather than the
+// individual arguments.
+class X86VectorVTInfo<int NumElts, ValueType EltVT, RegisterClass rc,
+ string suffix = ""> {
+ RegisterClass RC = rc;
+
+ // Corresponding mask register class.
+ RegisterClass KRC = !cast<RegisterClass>("VK" # NumElts);
+
+ // Corresponding write-mask register class.
+ RegisterClass KRCWM = !cast<RegisterClass>("VK" # NumElts # "WM");
+
+ // The GPR register class that can hold the write mask. Use GR8 for fewer
+ // than 8 elements. Use shift-right and equal to work around the lack of
+ // !lt in tablegen.
+ RegisterClass MRC =
+ !cast<RegisterClass>("GR" #
+ !if (!eq (!srl(NumElts, 3), 0), 8, NumElts));
+
+ // Suffix used in the instruction mnemonic.
+ string Suffix = suffix;
+
+ string VTName = "v" # NumElts # EltVT;
+
+ // The vector VT.
+ ValueType VT = !cast<ValueType>(VTName);
+
+ string EltTypeName = !cast<string>(EltVT);
+ // Size of the element type in bits, e.g. 32 for v16i32.
+ string EltSizeName = !subst("i", "", !subst("f", "", EltTypeName));
+ int EltSize = EltVT.Size;
+
+ // "i" for integer types and "f" for floating-point types
+ string TypeVariantName = !subst(EltSizeName, "", EltTypeName);
+
+ // Size of RC in bits, e.g. 512 for VR512.
+ int Size = VT.Size;
+
+ // The corresponding memory operand, e.g. i512mem for VR512.
+ X86MemOperand MemOp = !cast<X86MemOperand>(TypeVariantName # Size # "mem");
+ X86MemOperand ScalarMemOp = !cast<X86MemOperand>(EltVT # "mem");
+
+ // Load patterns
+ // Note: For 128/256-bit integer VT we choose loadv2i64/loadv4i64
+ // due to load promotion during legalization
+ PatFrag LdFrag = !cast<PatFrag>("load" #
+ !if (!eq (TypeVariantName, "i"),
+ !if (!eq (Size, 128), "v2i64",
+ !if (!eq (Size, 256), "v4i64",
+ VTName)), VTName));
+ PatFrag ScalarLdFrag = !cast<PatFrag>("load" # EltVT);
+
+ // The corresponding float type, e.g. v16f32 for v16i32
+ // Note: For EltSize < 32, FloatVT is illegal and TableGen
+ // fails to compile, so we choose FloatVT = VT
+ ValueType FloatVT = !cast<ValueType>(
+ !if (!eq (!srl(EltSize,5),0),
+ VTName,
+ !if (!eq(TypeVariantName, "i"),
+ "v" # NumElts # "f" # EltSize,
+ VTName)));
+
+ // The string to specify embedded broadcast in assembly.
+ string BroadcastStr = "{1to" # NumElts # "}";
+}
+
+def v64i8_info : X86VectorVTInfo<64, i8, VR512, "b">;
+def v32i16_info : X86VectorVTInfo<32, i16, VR512, "w">;
+def v16i32_info : X86VectorVTInfo<16, i32, VR512, "d">;
+def v8i64_info : X86VectorVTInfo<8, i64, VR512, "q">;
+
+// "x" in v32i8x_info means RC = VR256X
+def v32i8x_info : X86VectorVTInfo<32, i8, VR256X, "b">;
+def v16i16x_info : X86VectorVTInfo<16, i16, VR256X, "w">;
+def v8i32x_info : X86VectorVTInfo<8, i32, VR256X, "d">;
+def v4i64x_info : X86VectorVTInfo<4, i64, VR256X, "q">;
+
+def v16i8x_info : X86VectorVTInfo<16, i8, VR128X, "b">;
+def v8i16x_info : X86VectorVTInfo<8, i16, VR128X, "w">;
+def v4i32x_info : X86VectorVTInfo<4, i32, VR128X, "d">;
+def v2i64x_info : X86VectorVTInfo<2, i64, VR128X, "q">;
+
+class AVX512VLVectorVTInfo<X86VectorVTInfo i512, X86VectorVTInfo i256,
+ X86VectorVTInfo i128> {
+ X86VectorVTInfo info512 = i512;
+ X86VectorVTInfo info256 = i256;
+ X86VectorVTInfo info128 = i128;
+}
+
+def avx512vl_i8_info : AVX512VLVectorVTInfo<v64i8_info, v32i8x_info,
+ v16i8x_info>;
+def avx512vl_i16_info : AVX512VLVectorVTInfo<v32i16_info, v16i16x_info,
+ v8i16x_info>;
+def avx512vl_i32_info : AVX512VLVectorVTInfo<v16i32_info, v8i32x_info,
+ v4i32x_info>;
+def avx512vl_i64_info : AVX512VLVectorVTInfo<v8i64_info, v4i64x_info,
+ v2i64x_info>;
+
+
+// Common base class of AVX512_masking and AVX512_masking_3src.
+multiclass AVX512_masking_common<bits<8> O, Format F, dag Outs, dag Ins,
+ dag MaskingIns, dag ZeroMaskingIns,
+ string OpcodeStr,
+ string AttSrcAsm, string IntelSrcAsm,
+ dag RHS, dag MaskingRHS, ValueType OpVT,
+ RegisterClass RC, RegisterClass KRC,
+ string MaskingConstraint = ""> {
+ def NAME: AVX512<O, F, Outs, Ins,
+ OpcodeStr#" \t{"#AttSrcAsm#", $dst|"#
+ "$dst, "#IntelSrcAsm#"}",
+ [(set RC:$dst, RHS)]>;
+
+ // Prefer over VMOV*rrk Pat<>
+ let AddedComplexity = 20 in
+ def NAME#k: AVX512<O, F, Outs, MaskingIns,
+ OpcodeStr#" \t{"#AttSrcAsm#", $dst {${mask}}|"#
+ "$dst {${mask}}, "#IntelSrcAsm#"}",
+ [(set RC:$dst, MaskingRHS)]>,
+ EVEX_K {
+ // In case of the 3src subclass this is overridden with a let.
+ string Constraints = MaskingConstraint;
+ }
+ let AddedComplexity = 30 in // Prefer over VMOV*rrkz Pat<>
+ def NAME#kz: AVX512<O, F, Outs, ZeroMaskingIns,
+ OpcodeStr#" \t{"#AttSrcAsm#", $dst {${mask}} {z}|"#
+ "$dst {${mask}} {z}, "#IntelSrcAsm#"}",
+ [(set RC:$dst,
+ (vselect KRC:$mask, RHS,
+ (OpVT (bitconvert
+ (v16i32 immAllZerosV)))))]>,
+ EVEX_KZ;
+}
+
+// This multiclass generates the unconditional/non-masking, the masking and
+// the zero-masking variant of the instruction. In the masking case, the
+// perserved vector elements come from a new dummy input operand tied to $dst.
+multiclass AVX512_masking<bits<8> O, Format F, dag Outs, dag Ins,
+ string OpcodeStr,
+ string AttSrcAsm, string IntelSrcAsm,
+ dag RHS, ValueType OpVT, RegisterClass RC,
+ RegisterClass KRC> :
+ AVX512_masking_common<O, F, Outs,
+ Ins,
+ !con((ins RC:$src0, KRC:$mask), Ins),
+ !con((ins KRC:$mask), Ins),
+ OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
+ (vselect KRC:$mask, RHS, RC:$src0), OpVT, RC, KRC,
+ "$src0 = $dst">;
+
+// Similar to AVX512_masking but in this case one of the source operands
+// ($src1) is already tied to $dst so we just use that for the preserved
+// vector elements. NOTE that the NonTiedIns (the ins dag) should exclude
+// $src1.
+multiclass AVX512_masking_3src<bits<8> O, Format F, dag Outs, dag NonTiedIns,
+ string OpcodeStr,
+ string AttSrcAsm, string IntelSrcAsm,
+ dag RHS, ValueType OpVT,
+ RegisterClass RC, RegisterClass KRC> :
+ AVX512_masking_common<O, F, Outs,
+ !con((ins RC:$src1), NonTiedIns),
+ !con((ins RC:$src1), !con((ins KRC:$mask),
+ NonTiedIns)),
+ !con((ins RC:$src1), !con((ins KRC:$mask),
+ NonTiedIns)),
+ OpcodeStr, AttSrcAsm, IntelSrcAsm, RHS,
+ (vselect KRC:$mask, RHS, RC:$src1), OpVT, RC, KRC>;
+
// Bitcasts between 512-bit vector types. Return the original type since
// no instruction is needed for the conversion
let Predicates = [HasAVX512] in {
def : Pat<(v8i64 (bitconvert (v8f64 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v16f32 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v16i32 (bitconvert (v8i64 VR512:$src))), (v16i32 VR512:$src)>;
+ def : Pat<(v16i32 (bitconvert (v16f32 VR512:$src))), (v16i32 VR512:$src)>;
def : Pat<(v16i32 (bitconvert (v32i16 VR512:$src))), (v16i32 VR512:$src)>;
def : Pat<(v16i32 (bitconvert (v64i8 VR512:$src))), (v16i32 VR512:$src)>;
def : Pat<(v16i32 (bitconvert (v8f64 VR512:$src))), (v16i32 VR512:$src)>;
// vinsertps - insert f32 to XMM
def VINSERTPSzrr : AVX512AIi8<0x21, MRMSrcReg, (outs VR128X:$dst),
- (ins VR128X:$src1, VR128X:$src2, u32u8imm:$src3),
+ (ins VR128X:$src1, VR128X:$src2, i8imm:$src3),
"vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
[(set VR128X:$dst, (X86insertps VR128X:$src1, VR128X:$src2, imm:$src3))]>,
EVEX_4V;
def VINSERTPSzrm: AVX512AIi8<0x21, MRMSrcMem, (outs VR128X:$dst),
- (ins VR128X:$src1, f32mem:$src2, u32u8imm:$src3),
+ (ins VR128X:$src1, f32mem:$src2, i8imm:$src3),
"vinsertps\t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}",
[(set VR128X:$dst, (X86insertps VR128X:$src1,
(v4f32 (scalar_to_vector (loadf32 addr:$src2))),
(EXTRACT_get_vextract128_imm VR128X:$ext)))>;
def : Pat<(vextract128_extract:$ext VR512:$src1, (iPTR imm)),
- (v4i32 (VEXTRACTF32x4rr VR512:$src1,
+ (v4i32 (VEXTRACTI32x4rr VR512:$src1,
(EXTRACT_get_vextract128_imm VR128X:$ext)))>;
def : Pat<(vextract128_extract:$ext (v8f64 VR512:$src1), (iPTR imm)),
// vextractps - extract 32 bits from XMM
def VEXTRACTPSzrr : AVX512AIi8<0x17, MRMDestReg, (outs GR32:$dst),
- (ins VR128X:$src1, u32u8imm:$src2),
+ (ins VR128X:$src1, i32i8imm:$src2),
"vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
[(set GR32:$dst, (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2))]>,
EVEX;
def VEXTRACTPSzmr : AVX512AIi8<0x17, MRMDestMem, (outs),
- (ins f32mem:$dst, VR128X:$src1, u32u8imm:$src2),
+ (ins f32mem:$dst, VR128X:$src1, i32i8imm:$src2),
"vextractps\t{$src2, $src1, $dst|$dst, $src1, $src2}",
[(store (extractelt (bc_v4i32 (v4f32 VR128X:$src1)), imm:$src2),
addr:$dst)]>, EVEX, EVEX_CD8<32, CD8VT1>;
XD, VEX_W;
}
-multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, RegisterClass KRC,
- RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag,
- SDNode OpNode, ValueType vt> {
+multiclass avx512_icmp_packed<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ X86VectorVTInfo _> {
def rr : AVX512BI<opc, MRMSrcReg,
- (outs KRC:$dst), (ins RC:$src1, RC:$src2),
- !strconcat(OpcodeStr, " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set KRC:$dst, (OpNode (vt RC:$src1), (vt RC:$src2)))],
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.RC:$src2),
+ !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+ [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2)))],
IIC_SSE_ALU_F32P_RR>, EVEX_4V;
+ let mayLoad = 1 in
def rm : AVX512BI<opc, MRMSrcMem,
- (outs KRC:$dst), (ins RC:$src1, x86memop:$src2),
- !strconcat(OpcodeStr, " \t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set KRC:$dst, (OpNode (vt RC:$src1), (memop_frag addr:$src2)))],
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2),
+ !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
+ [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
+ (_.VT (bitconvert (_.LdFrag addr:$src2)))))],
IIC_SSE_ALU_F32P_RM>, EVEX_4V;
+ def rrk : AVX512BI<opc, MRMSrcReg,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2),
+ !strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, $src2}"),
+ [(set _.KRC:$dst, (and _.KRCWM:$mask,
+ (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2))))],
+ IIC_SSE_ALU_F32P_RR>, EVEX_4V, EVEX_K;
+ let mayLoad = 1 in
+ def rmk : AVX512BI<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2),
+ !strconcat(OpcodeStr, "\t{$src2, $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, $src2}"),
+ [(set _.KRC:$dst, (and _.KRCWM:$mask,
+ (OpNode (_.VT _.RC:$src1),
+ (_.VT (bitconvert
+ (_.LdFrag addr:$src2))))))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K;
+}
+
+multiclass avx512_icmp_packed_rmb<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ X86VectorVTInfo _> :
+ avx512_icmp_packed<opc, OpcodeStr, OpNode, _> {
+ let mayLoad = 1 in {
+ def rmb : AVX512BI<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2),
+ !strconcat(OpcodeStr, "\t{${src2}", _.BroadcastStr, ", $src1, $dst",
+ "|$dst, $src1, ${src2}", _.BroadcastStr, "}"),
+ [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
+ (X86VBroadcast (_.ScalarLdFrag addr:$src2))))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_B;
+ def rmbk : AVX512BI<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1,
+ _.ScalarMemOp:$src2),
+ !strconcat(OpcodeStr,
+ "\t{${src2}", _.BroadcastStr, ", $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, ${src2}", _.BroadcastStr, "}"),
+ [(set _.KRC:$dst, (and _.KRCWM:$mask,
+ (OpNode (_.VT _.RC:$src1),
+ (X86VBroadcast
+ (_.ScalarLdFrag addr:$src2)))))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K, EVEX_B;
+ }
}
-defm VPCMPEQDZ : avx512_icmp_packed<0x76, "vpcmpeqd", VK16, VR512, i512mem,
- memopv16i32, X86pcmpeqm, v16i32>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
-defm VPCMPEQQZ : avx512_icmp_packed<0x29, "vpcmpeqq", VK8, VR512, i512mem,
- memopv8i64, X86pcmpeqm, v8i64>, T8PD, EVEX_V512,
- VEX_W, EVEX_CD8<64, CD8VF>;
+multiclass avx512_icmp_packed_vl<bits<8> opc, string OpcodeStr, SDNode OpNode,
+ AVX512VLVectorVTInfo VTInfo, Predicate prd> {
+ let Predicates = [prd] in
+ defm Z : avx512_icmp_packed<opc, OpcodeStr, OpNode, VTInfo.info512>,
+ EVEX_V512;
-defm VPCMPGTDZ : avx512_icmp_packed<0x66, "vpcmpgtd", VK16, VR512, i512mem,
- memopv16i32, X86pcmpgtm, v16i32>, EVEX_V512,
- EVEX_CD8<32, CD8VF>;
-defm VPCMPGTQZ : avx512_icmp_packed<0x37, "vpcmpgtq", VK8, VR512, i512mem,
- memopv8i64, X86pcmpgtm, v8i64>, T8PD, EVEX_V512,
- VEX_W, EVEX_CD8<64, CD8VF>;
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_icmp_packed<opc, OpcodeStr, OpNode, VTInfo.info256>,
+ EVEX_V256;
+ defm Z128 : avx512_icmp_packed<opc, OpcodeStr, OpNode, VTInfo.info128>,
+ EVEX_V128;
+ }
+}
+
+multiclass avx512_icmp_packed_rmb_vl<bits<8> opc, string OpcodeStr,
+ SDNode OpNode, AVX512VLVectorVTInfo VTInfo,
+ Predicate prd> {
+ let Predicates = [prd] in
+ defm Z : avx512_icmp_packed_rmb<opc, OpcodeStr, OpNode, VTInfo.info512>,
+ EVEX_V512;
+
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_icmp_packed_rmb<opc, OpcodeStr, OpNode, VTInfo.info256>,
+ EVEX_V256;
+ defm Z128 : avx512_icmp_packed_rmb<opc, OpcodeStr, OpNode, VTInfo.info128>,
+ EVEX_V128;
+ }
+}
+
+defm VPCMPEQB : avx512_icmp_packed_vl<0x74, "vpcmpeqb", X86pcmpeqm,
+ avx512vl_i8_info, HasBWI>,
+ EVEX_CD8<8, CD8VF>;
+
+defm VPCMPEQW : avx512_icmp_packed_vl<0x75, "vpcmpeqw", X86pcmpeqm,
+ avx512vl_i16_info, HasBWI>,
+ EVEX_CD8<16, CD8VF>;
+
+defm VPCMPEQD : avx512_icmp_packed_rmb_vl<0x76, "vpcmpeqd", X86pcmpeqm,
+ avx512vl_i32_info, HasAVX512>,
+ EVEX_CD8<32, CD8VF>;
+
+defm VPCMPEQQ : avx512_icmp_packed_rmb_vl<0x29, "vpcmpeqq", X86pcmpeqm,
+ avx512vl_i64_info, HasAVX512>,
+ T8PD, VEX_W, EVEX_CD8<64, CD8VF>;
+
+defm VPCMPGTB : avx512_icmp_packed_vl<0x64, "vpcmpgtb", X86pcmpgtm,
+ avx512vl_i8_info, HasBWI>,
+ EVEX_CD8<8, CD8VF>;
+
+defm VPCMPGTW : avx512_icmp_packed_vl<0x65, "vpcmpgtw", X86pcmpgtm,
+ avx512vl_i16_info, HasBWI>,
+ EVEX_CD8<16, CD8VF>;
+
+defm VPCMPGTD : avx512_icmp_packed_rmb_vl<0x66, "vpcmpgtd", X86pcmpgtm,
+ avx512vl_i32_info, HasAVX512>,
+ EVEX_CD8<32, CD8VF>;
+
+defm VPCMPGTQ : avx512_icmp_packed_rmb_vl<0x37, "vpcmpgtq", X86pcmpgtm,
+ avx512vl_i64_info, HasAVX512>,
+ T8PD, VEX_W, EVEX_CD8<64, CD8VF>;
def : Pat<(v8i1 (X86pcmpgtm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))),
- (COPY_TO_REGCLASS (VPCMPGTDZrr
+ (COPY_TO_REGCLASS (VPCMPGTDZrr
(v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
(v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm))), VK8)>;
def : Pat<(v8i1 (X86pcmpeqm (v8i32 VR256X:$src1), (v8i32 VR256X:$src2))),
- (COPY_TO_REGCLASS (VPCMPEQDZrr
+ (COPY_TO_REGCLASS (VPCMPEQDZrr
(v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src1, sub_ymm)),
(v16i32 (SUBREG_TO_REG (i32 0), VR256X:$src2, sub_ymm))), VK8)>;
-multiclass avx512_icmp_cc<bits<8> opc, RegisterClass WMRC, RegisterClass KRC,
- RegisterClass RC, X86MemOperand x86memop, PatFrag memop_frag,
- SDNode OpNode, ValueType vt, Operand CC, string Suffix> {
+multiclass avx512_icmp_cc<bits<8> opc, string Suffix, SDNode OpNode,
+ X86VectorVTInfo _> {
def rri : AVX512AIi8<opc, MRMSrcReg,
- (outs KRC:$dst), (ins RC:$src1, RC:$src2, CC:$cc),
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.RC:$src2, AVXCC:$cc),
!strconcat("vpcmp${cc}", Suffix,
"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set KRC:$dst, (OpNode (vt RC:$src1), (vt RC:$src2), imm:$cc))],
+ [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2),
+ imm:$cc))],
IIC_SSE_ALU_F32P_RR>, EVEX_4V;
+ let mayLoad = 1 in
def rmi : AVX512AIi8<opc, MRMSrcMem,
- (outs KRC:$dst), (ins RC:$src1, x86memop:$src2, CC:$cc),
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2, AVXCC:$cc),
!strconcat("vpcmp${cc}", Suffix,
"\t{$src2, $src1, $dst|$dst, $src1, $src2}"),
- [(set KRC:$dst, (OpNode (vt RC:$src1), (memop_frag addr:$src2),
- imm:$cc))], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
+ [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
+ (_.VT (bitconvert (_.LdFrag addr:$src2))),
+ imm:$cc))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V;
+ def rrik : AVX512AIi8<opc, MRMSrcReg,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2,
+ AVXCC:$cc),
+ !strconcat("vpcmp${cc}", Suffix,
+ "\t{$src2, $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, $src2}"),
+ [(set _.KRC:$dst, (and _.KRCWM:$mask,
+ (OpNode (_.VT _.RC:$src1), (_.VT _.RC:$src2),
+ imm:$cc)))],
+ IIC_SSE_ALU_F32P_RR>, EVEX_4V, EVEX_K;
+ let mayLoad = 1 in
+ def rmik : AVX512AIi8<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2,
+ AVXCC:$cc),
+ !strconcat("vpcmp${cc}", Suffix,
+ "\t{$src2, $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, $src2}"),
+ [(set _.KRC:$dst, (and _.KRCWM:$mask,
+ (OpNode (_.VT _.RC:$src1),
+ (_.VT (bitconvert (_.LdFrag addr:$src2))),
+ imm:$cc)))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K;
+
// Accept explicit immediate argument form instead of comparison code.
let isAsmParserOnly = 1, hasSideEffects = 0 in {
def rri_alt : AVX512AIi8<opc, MRMSrcReg,
- (outs KRC:$dst), (ins RC:$src1, RC:$src2, i8imm:$cc),
- !strconcat("vpcmp", Suffix,
- "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"),
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.RC:$src2, i8imm:$cc),
+ !strconcat("vpcmp", Suffix, "\t{$cc, $src2, $src1, $dst|",
+ "$dst, $src1, $src2, $cc}"),
[], IIC_SSE_ALU_F32P_RR>, EVEX_4V;
+ def rmi_alt : AVX512AIi8<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.MemOp:$src2, i8imm:$cc),
+ !strconcat("vpcmp", Suffix, "\t{$cc, $src2, $src1, $dst|",
+ "$dst, $src1, $src2, $cc}"),
+ [], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
def rrik_alt : AVX512AIi8<opc, MRMSrcReg,
- (outs KRC:$dst), (ins WMRC:$mask, RC:$src1, RC:$src2, i8imm:$cc),
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.RC:$src2,
+ i8imm:$cc),
!strconcat("vpcmp", Suffix,
- "\t{$cc, $src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2, $cc}"),
+ "\t{$cc, $src2, $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, $src2, $cc}"),
[], IIC_SSE_ALU_F32P_RR>, EVEX_4V, EVEX_K;
- def rmi_alt : AVX512AIi8<opc, MRMSrcMem,
- (outs KRC:$dst), (ins RC:$src1, x86memop:$src2, i8imm:$cc),
- !strconcat("vpcmp", Suffix,
- "\t{$cc, $src2, $src1, $dst|$dst, $src1, $src2, $cc}"),
- [], IIC_SSE_ALU_F32P_RM>, EVEX_4V;
def rmik_alt : AVX512AIi8<opc, MRMSrcMem,
- (outs KRC:$dst), (ins WMRC:$mask, RC:$src1, x86memop:$src2, i8imm:$cc),
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1, _.MemOp:$src2,
+ i8imm:$cc),
!strconcat("vpcmp", Suffix,
- "\t{$cc, $src2, $src1, $dst {${mask}}|$dst {${mask}}, $src1, $src2, $cc}"),
+ "\t{$cc, $src2, $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, $src2, $cc}"),
[], IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K;
}
}
-defm VPCMPDZ : avx512_icmp_cc<0x1F, VK16WM, VK16, VR512, i512mem, memopv16i32,
- X86cmpm, v16i32, AVXCC, "d">,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VPCMPUDZ : avx512_icmp_cc<0x1E, VK16WM, VK16, VR512, i512mem, memopv16i32,
- X86cmpmu, v16i32, AVXCC, "ud">,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
+multiclass avx512_icmp_cc_rmb<bits<8> opc, string Suffix, SDNode OpNode,
+ X86VectorVTInfo _> :
+ avx512_icmp_cc<opc, Suffix, OpNode, _> {
+ let mayLoad = 1 in {
+ def rmib : AVX512AIi8<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2,
+ AVXCC:$cc),
+ !strconcat("vpcmp${cc}", Suffix,
+ "\t{${src2}", _.BroadcastStr, ", $src1, $dst|",
+ "$dst, $src1, ${src2}", _.BroadcastStr, "}"),
+ [(set _.KRC:$dst, (OpNode (_.VT _.RC:$src1),
+ (X86VBroadcast (_.ScalarLdFrag addr:$src2)),
+ imm:$cc))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_B;
+ def rmibk : AVX512AIi8<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1,
+ _.ScalarMemOp:$src2, AVXCC:$cc),
+ !strconcat("vpcmp${cc}", Suffix,
+ "\t{${src2}", _.BroadcastStr, ", $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, ${src2}", _.BroadcastStr, "}"),
+ [(set _.KRC:$dst, (and _.KRCWM:$mask,
+ (OpNode (_.VT _.RC:$src1),
+ (X86VBroadcast (_.ScalarLdFrag addr:$src2)),
+ imm:$cc)))],
+ IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K, EVEX_B;
+ }
+
+ // Accept explicit immediate argument form instead of comparison code.
+ let isAsmParserOnly = 1, hasSideEffects = 0 in {
+ def rmib_alt : AVX512AIi8<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.RC:$src1, _.ScalarMemOp:$src2,
+ i8imm:$cc),
+ !strconcat("vpcmp", Suffix,
+ "\t{$cc, ${src2}", _.BroadcastStr, ", $src1, $dst|",
+ "$dst, $src1, ${src2}", _.BroadcastStr, ", $cc}"),
+ [], IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_B;
+ def rmibk_alt : AVX512AIi8<opc, MRMSrcMem,
+ (outs _.KRC:$dst), (ins _.KRCWM:$mask, _.RC:$src1,
+ _.ScalarMemOp:$src2, i8imm:$cc),
+ !strconcat("vpcmp", Suffix,
+ "\t{$cc, ${src2}", _.BroadcastStr, ", $src1, $dst {${mask}}|",
+ "$dst {${mask}}, $src1, ${src2}", _.BroadcastStr, ", $cc}"),
+ [], IIC_SSE_ALU_F32P_RM>, EVEX_4V, EVEX_K, EVEX_B;
+ }
+}
+
+multiclass avx512_icmp_cc_vl<bits<8> opc, string Suffix, SDNode OpNode,
+ AVX512VLVectorVTInfo VTInfo, Predicate prd> {
+ let Predicates = [prd] in
+ defm Z : avx512_icmp_cc<opc, Suffix, OpNode, VTInfo.info512>, EVEX_V512;
+
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_icmp_cc<opc, Suffix, OpNode, VTInfo.info256>, EVEX_V256;
+ defm Z128 : avx512_icmp_cc<opc, Suffix, OpNode, VTInfo.info128>, EVEX_V128;
+ }
+}
+
+multiclass avx512_icmp_cc_rmb_vl<bits<8> opc, string Suffix, SDNode OpNode,
+ AVX512VLVectorVTInfo VTInfo, Predicate prd> {
+ let Predicates = [prd] in
+ defm Z : avx512_icmp_cc_rmb<opc, Suffix, OpNode, VTInfo.info512>,
+ EVEX_V512;
-defm VPCMPQZ : avx512_icmp_cc<0x1F, VK8WM, VK8, VR512, i512mem, memopv8i64,
- X86cmpm, v8i64, AVXCC, "q">,
- VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
-defm VPCMPUQZ : avx512_icmp_cc<0x1E, VK8WM, VK8, VR512, i512mem, memopv8i64,
- X86cmpmu, v8i64, AVXCC, "uq">,
- VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_icmp_cc_rmb<opc, Suffix, OpNode, VTInfo.info256>,
+ EVEX_V256;
+ defm Z128 : avx512_icmp_cc_rmb<opc, Suffix, OpNode, VTInfo.info128>,
+ EVEX_V128;
+ }
+}
+
+defm VPCMPB : avx512_icmp_cc_vl<0x3F, "b", X86cmpm, avx512vl_i8_info,
+ HasBWI>, EVEX_CD8<8, CD8VF>;
+defm VPCMPUB : avx512_icmp_cc_vl<0x3E, "ub", X86cmpmu, avx512vl_i8_info,
+ HasBWI>, EVEX_CD8<8, CD8VF>;
+
+defm VPCMPW : avx512_icmp_cc_vl<0x3F, "w", X86cmpm, avx512vl_i16_info,
+ HasBWI>, VEX_W, EVEX_CD8<16, CD8VF>;
+defm VPCMPUW : avx512_icmp_cc_vl<0x3E, "uw", X86cmpmu, avx512vl_i16_info,
+ HasBWI>, VEX_W, EVEX_CD8<16, CD8VF>;
+
+defm VPCMPD : avx512_icmp_cc_rmb_vl<0x1F, "d", X86cmpm, avx512vl_i32_info,
+ HasAVX512>, EVEX_CD8<32, CD8VF>;
+defm VPCMPUD : avx512_icmp_cc_rmb_vl<0x1E, "ud", X86cmpmu, avx512vl_i32_info,
+ HasAVX512>, EVEX_CD8<32, CD8VF>;
+
+defm VPCMPQ : avx512_icmp_cc_rmb_vl<0x1F, "q", X86cmpm, avx512vl_i64_info,
+ HasAVX512>, VEX_W, EVEX_CD8<64, CD8VF>;
+defm VPCMPUQ : avx512_icmp_cc_rmb_vl<0x1E, "uq", X86cmpmu, avx512vl_i64_info,
+ HasAVX512>, VEX_W, EVEX_CD8<64, CD8VF>;
// avx512_cmp_packed - compare packed instructions
multiclass avx512_cmp_packed<RegisterClass KRC, RegisterClass RC,
}
let Predicates = [HasAVX512] in {
+ def : Pat<(i1 (trunc (i64 GR64:$src))),
+ (COPY_TO_REGCLASS (KMOVWkr (AND32ri (EXTRACT_SUBREG $src, sub_32bit),
+ (i32 1))), VK1)>;
+
def : Pat<(i1 (trunc (i32 GR32:$src))),
(COPY_TO_REGCLASS (KMOVWkr (AND32ri $src, (i32 1))), VK1)>;
!strconcat(asm, " \t{$src, $dst|$dst, $src}"),
[(set RC:$dst, (mem_pat addr:$src))], IIC_SSE_MOV_S_RM>,
EVEX, VEX_LIG;
+ let mayStore = 1 in {
def mr: SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, RC:$src),
!strconcat(asm, " \t{$src, $dst|$dst, $src}"),
[(store RC:$src, addr:$dst)], IIC_SSE_MOV_S_MR>,
EVEX, VEX_LIG;
+ def mrk: SI<0x11, MRMDestMem, (outs), (ins x86memop:$dst, VK1WM:$mask, RC:$src),
+ !strconcat(asm, " \t{$src, $dst {${mask}}|$dst {${mask}}, $src}"),
+ [], IIC_SSE_MOV_S_MR>,
+ EVEX, VEX_LIG, EVEX_K;
+ } // mayStore
} //hasSideEffects = 0
}
(COPY_TO_REGCLASS (VMOVSDZrrk (COPY_TO_REGCLASS FR64X:$src2, VR128X),
VK1WM:$mask, (f64 (IMPLICIT_DEF)), FR64X:$src1), FR64X)>;
+def : Pat<(int_x86_avx512_mask_store_ss addr:$dst, VR128X:$src, GR8:$mask),
+ (VMOVSSZmrk addr:$dst, (i1 (COPY_TO_REGCLASS GR8:$mask, VK1WM)),
+ (COPY_TO_REGCLASS VR128X:$src, FR32X))>;
+
// For the disassembler
let isCodeGenOnly = 1, ForceDisassemble = 1, hasSideEffects = 0 in {
def VMOVSSZrr_REV : SI<0x11, MRMDestReg, (outs VR128X:$dst),
//===----------------------------------------------------------------------===//
// AVX-512 - Non-temporals
//===----------------------------------------------------------------------===//
+let SchedRW = [WriteLoad] in {
+ def VMOVNTDQAZrm : AVX512PI<0x2A, MRMSrcMem, (outs VR512:$dst),
+ (ins i512mem:$src), "vmovntdqa\t{$src, $dst|$dst, $src}",
+ [(set VR512:$dst, (int_x86_avx512_movntdqa addr:$src))],
+ SSEPackedInt>, EVEX, T8PD, EVEX_V512,
+ EVEX_CD8<64, CD8VF>;
+
+ let Predicates = [HasAVX512, HasVLX] in {
+ def VMOVNTDQAZ256rm : AVX512PI<0x2A, MRMSrcMem, (outs VR256X:$dst),
+ (ins i256mem:$src),
+ "vmovntdqa\t{$src, $dst|$dst, $src}", [],
+ SSEPackedInt>, EVEX, T8PD, EVEX_V256,
+ EVEX_CD8<64, CD8VF>;
+
+ def VMOVNTDQAZ128rm : AVX512PI<0x2A, MRMSrcMem, (outs VR128X:$dst),
+ (ins i128mem:$src),
+ "vmovntdqa\t{$src, $dst|$dst, $src}", [],
+ SSEPackedInt>, EVEX, T8PD, EVEX_V128,
+ EVEX_CD8<64, CD8VF>;
+ }
+}
+
+multiclass avx512_movnt<bits<8> opc, string OpcodeStr, PatFrag st_frag,
+ ValueType OpVT, RegisterClass RC, X86MemOperand memop,
+ Domain d, InstrItinClass itin = IIC_SSE_MOVNT> {
+ let SchedRW = [WriteStore], mayStore = 1,
+ AddedComplexity = 400 in
+ def mr : AVX512PI<opc, MRMDestMem, (outs), (ins memop:$dst, RC:$src),
+ !strconcat(OpcodeStr, "\t{$src, $dst|$dst, $src}"),
+ [(st_frag (OpVT RC:$src), addr:$dst)], d, itin>, EVEX;
+}
-def VMOVNTDQAZrm : AVX5128I<0x2A, MRMSrcMem, (outs VR512:$dst),
- (ins i512mem:$src),
- "vmovntdqa\t{$src, $dst|$dst, $src}",
- [(set VR512:$dst,
- (int_x86_avx512_movntdqa addr:$src))]>,
- EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
-
-// Prefer non-temporal over temporal versions
-let AddedComplexity = 400, SchedRW = [WriteStore] in {
-
-def VMOVNTPSZmr : AVX512PSI<0x2B, MRMDestMem, (outs),
- (ins f512mem:$dst, VR512:$src),
- "vmovntps\t{$src, $dst|$dst, $src}",
- [(alignednontemporalstore (v16f32 VR512:$src),
- addr:$dst)],
- IIC_SSE_MOVNT>,
- EVEX, EVEX_V512, EVEX_CD8<32, CD8VF>;
-
-def VMOVNTPDZmr : AVX512PDI<0x2B, MRMDestMem, (outs),
- (ins f512mem:$dst, VR512:$src),
- "vmovntpd\t{$src, $dst|$dst, $src}",
- [(alignednontemporalstore (v8f64 VR512:$src),
- addr:$dst)],
- IIC_SSE_MOVNT>,
- EVEX, EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
-
-
-def VMOVNTDQZmr : AVX512BI<0xE7, MRMDestMem, (outs),
- (ins i512mem:$dst, VR512:$src),
- "vmovntdq\t{$src, $dst|$dst, $src}",
- [(alignednontemporalstore (v8i64 VR512:$src),
- addr:$dst)],
- IIC_SSE_MOVNT>,
- EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
+multiclass avx512_movnt_vl<bits<8> opc, string OpcodeStr, PatFrag st_frag,
+ string elty, string elsz, string vsz512,
+ string vsz256, string vsz128, Domain d,
+ Predicate prd, InstrItinClass itin = IIC_SSE_MOVNT> {
+ let Predicates = [prd] in
+ defm Z : avx512_movnt<opc, OpcodeStr, st_frag,
+ !cast<ValueType>("v"##vsz512##elty##elsz), VR512,
+ !cast<X86MemOperand>(elty##"512mem"), d, itin>,
+ EVEX_V512;
+
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_movnt<opc, OpcodeStr, st_frag,
+ !cast<ValueType>("v"##vsz256##elty##elsz), VR256X,
+ !cast<X86MemOperand>(elty##"256mem"), d, itin>,
+ EVEX_V256;
+
+ defm Z128 : avx512_movnt<opc, OpcodeStr, st_frag,
+ !cast<ValueType>("v"##vsz128##elty##elsz), VR128X,
+ !cast<X86MemOperand>(elty##"128mem"), d, itin>,
+ EVEX_V128;
+ }
}
+defm VMOVNTDQ : avx512_movnt_vl<0xE7, "vmovntdq", alignednontemporalstore,
+ "i", "64", "8", "4", "2", SSEPackedInt,
+ HasAVX512>, PD, EVEX_CD8<64, CD8VF>;
+
+defm VMOVNTPD : avx512_movnt_vl<0x2B, "vmovntpd", alignednontemporalstore,
+ "f", "64", "8", "4", "2", SSEPackedDouble,
+ HasAVX512>, PD, VEX_W, EVEX_CD8<64, CD8VF>;
+
+defm VMOVNTPS : avx512_movnt_vl<0x2B, "vmovntps", alignednontemporalstore,
+ "f", "32", "16", "8", "4", SSEPackedSingle,
+ HasAVX512>, PS, EVEX_CD8<32, CD8VF>;
+
//===----------------------------------------------------------------------===//
// AVX-512 - Integer arithmetic
//
i512mem, v16i32>, PD, EVEX_V512, EVEX_CD8<32, CD8VF>;
let ExeDomain = SSEPackedSingle in
-defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilp,
+defm VPERMILPSZ : avx512_pshuf_imm<0x04, "vpermilps", VR512, X86VPermilpi,
memopv16f32, i512mem, v16f32>, TAPD, EVEX_V512,
EVEX_CD8<32, CD8VF>;
let ExeDomain = SSEPackedDouble in
-defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilp,
+defm VPERMILPDZ : avx512_pshuf_imm<0x05, "vpermilpd", VR512, X86VPermilpi,
memopv8f64, i512mem, v8f64>, TAPD, EVEX_V512,
VEX_W, EVEX_CD8<32, CD8VF>;
-def : Pat<(v16i32 (X86VPermilp VR512:$src1, (i8 imm:$imm))),
+def : Pat<(v16i32 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
(VPERMILPSZri VR512:$src1, imm:$imm)>;
-def : Pat<(v8i64 (X86VPermilp VR512:$src1, (i8 imm:$imm))),
+def : Pat<(v8i64 (X86VPermilpi VR512:$src1, (i8 imm:$imm))),
(VPERMILPDZri VR512:$src1, imm:$imm)>;
//===----------------------------------------------------------------------===//
multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr,
RegisterClass RC, X86MemOperand x86memop,
PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
- string BrdcstStr, SDNode OpNode, ValueType OpVT> {
- def r: AVX512FMA3<opc, MRMSrcReg, (outs RC:$dst),
- (ins RC:$src1, RC:$src2, RC:$src3),
- !strconcat(OpcodeStr," \t{$src3, $src2, $dst|$dst, $src2, $src3}"),
- [(set RC:$dst, (OpVT(OpNode RC:$src1, RC:$src2, RC:$src3)))]>;
+ string BrdcstStr, SDNode OpNode, ValueType OpVT,
+ RegisterClass KRC> {
+ defm r: AVX512_masking_3src<opc, MRMSrcReg, (outs RC:$dst),
+ (ins RC:$src2, RC:$src3),
+ OpcodeStr, "$src3, $src2", "$src2, $src3",
+ (OpVT (OpNode RC:$src1, RC:$src2, RC:$src3)), OpVT, RC, KRC>,
+ AVX512FMA3Base;
let mayLoad = 1 in
def m: AVX512FMA3<opc, MRMSrcMem, (outs RC:$dst),
let ExeDomain = SSEPackedSingle in {
defm VFMADD213PSZ : avx512_fma3p_rm<0xA8, "vfmadd213ps", VR512, f512mem,
memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmadd, v16f32>, EVEX_V512,
+ X86Fmadd, v16f32, VK16WM>, EVEX_V512,
EVEX_CD8<32, CD8VF>;
defm VFMSUB213PSZ : avx512_fma3p_rm<0xAA, "vfmsub213ps", VR512, f512mem,
memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmsub, v16f32>, EVEX_V512,
+ X86Fmsub, v16f32, VK16WM>, EVEX_V512,
EVEX_CD8<32, CD8VF>;
defm VFMADDSUB213PSZ : avx512_fma3p_rm<0xA6, "vfmaddsub213ps", VR512, f512mem,
memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmaddsub, v16f32>,
+ X86Fmaddsub, v16f32, VK16WM>,
EVEX_V512, EVEX_CD8<32, CD8VF>;
defm VFMSUBADD213PSZ : avx512_fma3p_rm<0xA7, "vfmsubadd213ps", VR512, f512mem,
memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fmsubadd, v16f32>,
+ X86Fmsubadd, v16f32, VK16WM>,
EVEX_V512, EVEX_CD8<32, CD8VF>;
defm VFNMADD213PSZ : avx512_fma3p_rm<0xAC, "vfnmadd213ps", VR512, f512mem,
memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fnmadd, v16f32>, EVEX_V512,
+ X86Fnmadd, v16f32, VK16WM>, EVEX_V512,
EVEX_CD8<32, CD8VF>;
defm VFNMSUB213PSZ : avx512_fma3p_rm<0xAE, "vfnmsub213ps", VR512, f512mem,
memopv16f32, f32mem, loadf32, "{1to16}",
- X86Fnmsub, v16f32>, EVEX_V512,
+ X86Fnmsub, v16f32, VK16WM>, EVEX_V512,
EVEX_CD8<32, CD8VF>;
}
let ExeDomain = SSEPackedDouble in {
defm VFMADD213PDZ : avx512_fma3p_rm<0xA8, "vfmadd213pd", VR512, f512mem,
memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmadd, v8f64>, EVEX_V512,
+ X86Fmadd, v8f64, VK8WM>, EVEX_V512,
VEX_W, EVEX_CD8<64, CD8VF>;
defm VFMSUB213PDZ : avx512_fma3p_rm<0xAA, "vfmsub213pd", VR512, f512mem,
memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmsub, v8f64>, EVEX_V512, VEX_W,
+ X86Fmsub, v8f64, VK8WM>, EVEX_V512, VEX_W,
EVEX_CD8<64, CD8VF>;
defm VFMADDSUB213PDZ : avx512_fma3p_rm<0xA6, "vfmaddsub213pd", VR512, f512mem,
memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmaddsub, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
+ X86Fmaddsub, v8f64, VK8WM>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
defm VFMSUBADD213PDZ : avx512_fma3p_rm<0xA7, "vfmsubadd213pd", VR512, f512mem,
memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fmsubadd, v8f64>, EVEX_V512, VEX_W,
- EVEX_CD8<64, CD8VF>;
+ X86Fmsubadd, v8f64, VK8WM>,
+ EVEX_V512, VEX_W, EVEX_CD8<64, CD8VF>;
defm VFNMADD213PDZ : avx512_fma3p_rm<0xAC, "vfnmadd213pd", VR512, f512mem,
memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fnmadd, v8f64>, EVEX_V512, VEX_W,
+ X86Fnmadd, v8f64, VK8WM>, EVEX_V512, VEX_W,
EVEX_CD8<64, CD8VF>;
defm VFNMSUB213PDZ : avx512_fma3p_rm<0xAE, "vfnmsub213pd", VR512, f512mem,
memopv8f64, f64mem, loadf64, "{1to8}",
- X86Fnmsub, v8f64>, EVEX_V512, VEX_W,
+ X86Fnmsub, v8f64, VK8WM>, EVEX_V512, VEX_W,
EVEX_CD8<64, CD8VF>;
}
(memopv8i64 addr:$src2), (i8 imm:$imm))),
(VSHUFPDZrmi VR512:$src1, addr:$src2, imm:$imm)>;
-multiclass avx512_alignr<string Suffix, RegisterClass RC,
- X86MemOperand x86memop, ValueType IntVT,
- ValueType FloatVT> {
- def rri : AVX512AIi8<0x03, MRMSrcReg, (outs RC:$dst),
- (ins RC:$src1, RC:$src2, i8imm:$src3),
- !strconcat("valign"##Suffix,
- " \t{$src3, $src2, $src1, $dst|"
- "$dst, $src1, $src2, $src3}"),
- [(set RC:$dst,
- (IntVT (X86PAlignr RC:$src2, RC:$src1,
- (i8 imm:$src3))))]>, EVEX_4V;
+multiclass avx512_valign<X86VectorVTInfo _> {
+ defm rri : AVX512_masking<0x03, MRMSrcReg, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.RC:$src2, i8imm:$src3),
+ "valign"##_.Suffix,
+ "$src3, $src2, $src1", "$src1, $src2, $src3",
+ (_.VT (X86VAlign _.RC:$src2, _.RC:$src1,
+ (i8 imm:$src3))),
+ _.VT, _.RC, _.KRCWM>,
+ AVX512AIi8Base, EVEX_4V;
// Also match valign of packed floats.
- def : Pat<(FloatVT (X86PAlignr RC:$src1, RC:$src2, (i8 imm:$imm))),
- (!cast<Instruction>(NAME##rri) RC:$src2, RC:$src1, imm:$imm)>;
+ def : Pat<(_.FloatVT (X86VAlign _.RC:$src1, _.RC:$src2, (i8 imm:$imm))),
+ (!cast<Instruction>(NAME##rri) _.RC:$src2, _.RC:$src1, imm:$imm)>;
let mayLoad = 1 in
- def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs RC:$dst),
- (ins RC:$src1, x86memop:$src2, i8imm:$src3),
- !strconcat("valign"##Suffix,
+ def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.MemOp:$src2, i8imm:$src3),
+ !strconcat("valign"##_.Suffix,
" \t{$src3, $src2, $src1, $dst|"
"$dst, $src1, $src2, $src3}"),
[]>, EVEX_4V;
}
-defm VALIGND : avx512_alignr<"d", VR512, i512mem, v16i32, v16f32>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VALIGNQ : avx512_alignr<"q", VR512, i512mem, v8i64, v8f64>,
- VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
+defm VALIGND : avx512_valign<v16i32_info>, EVEX_V512, EVEX_CD8<32, CD8VF>;
+defm VALIGNQ : avx512_valign<v8i64_info>, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
// Helper fragments to match sext vXi1 to vXiY.
def v16i1sextv16i32 : PatLeaf<(v16i32 (X86vsrai VR512:$src, (i8 31)))>;