+// 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;
+
+ // The vector VT.
+ ValueType VT = !cast<ValueType>("v" # NumElts # EltVT);
+
+ string EltTypeName = !cast<string>(EltVT);
+ // Size of the element type in bits, e.g. 32 for v16i32.
+ string EltSize = !subst("i", "", !subst("f", "", EltTypeName));
+
+ // "i" for integer types and "f" for floating-point types
+ string TypeVariantName = !subst(EltSize, "", 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");
+
+ // The corresponding float type, e.g. v16f32 for v16i32
+ ValueType FloatVT = !if (!eq(TypeVariantName, "i"),
+ !cast<ValueType>("v" # NumElts # "f" # EltSize),
+ VT);
+
+ // The string to specify embedded broadcast in assembly.
+ string BroadcastStr = "{1to" # NumElts # "}";
+}
+
+def v16i32_info : X86VectorVTInfo<16, i32, VR512, "d">;
+def v8i64_info : X86VectorVTInfo<8, i64, VR512, "q">;
+
+// 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)>;
}
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 $src, (i32 1))), VK1)>;
//===----------------------------------------------------------------------===//
// AVX-512 - Non-temporals
//===----------------------------------------------------------------------===//
+let SchedRW = [WriteLoad] in {
+ def VMOVNTDQAZrm : AVX512PI<0x2A, MRMSrcMem, (outs VR512:$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:$dst)],
- IIC_SSE_MOVNT>,
- EVEX, EVEX_V512, EVEX_CD8<64, CD8VF>;
+multiclass avx512_movnt_vl<bits<8> opc, string OpcodeStr, PatFrag st_frag,
+ string elty, string elsz, string vsz512,
+ string vsz256, string vsz128, Domain d,
+ Predicate prd, InstrItinClass itin = IIC_SSE_MOVNT> {
+ let Predicates = [prd] in
+ defm Z : avx512_movnt<opc, OpcodeStr, st_frag,
+ !cast<ValueType>("v"##vsz512##elty##elsz), VR512,
+ !cast<X86MemOperand>(elty##"512mem"), d, itin>,
+ EVEX_V512;
+
+ let Predicates = [prd, HasVLX] in {
+ defm Z256 : avx512_movnt<opc, OpcodeStr, st_frag,
+ !cast<ValueType>("v"##vsz256##elty##elsz), VR256X,
+ !cast<X86MemOperand>(elty##"256mem"), d, itin>,
+ EVEX_V256;
+
+ defm Z128 : avx512_movnt<opc, OpcodeStr, st_frag,
+ !cast<ValueType>("v"##vsz128##elty##elsz), VR128X,
+ !cast<X86MemOperand>(elty##"128mem"), d, itin>,
+ EVEX_V128;
+ }
}
+defm VMOVNTDQ : avx512_movnt_vl<0xE7, "vmovntdq", alignednontemporalstore,
+ "i", "64", "8", "4", "2", SSEPackedInt,
+ HasAVX512>, PD, EVEX_CD8<64, CD8VF>;
+
+defm VMOVNTPD : avx512_movnt_vl<0x2B, "vmovntpd", alignednontemporalstore,
+ "f", "64", "8", "4", "2", SSEPackedDouble,
+ HasAVX512>, PD, VEX_W, EVEX_CD8<64, CD8VF>;
+
+defm VMOVNTPS : avx512_movnt_vl<0x2B, "vmovntps", alignednontemporalstore,
+ "f", "32", "16", "8", "4", SSEPackedSingle,
+ HasAVX512>, PS, EVEX_CD8<32, CD8VF>;
+
//===----------------------------------------------------------------------===//
// AVX-512 - Integer arithmetic
//
multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr,
RegisterClass RC, X86MemOperand x86memop,
PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
- string BrdcstStr, SDNode OpNode, ValueType OpVT> {
- def r: AVX512FMA3<opc, MRMSrcReg, (outs RC:$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 OpcodeStr, RegisterClass RC,
- X86MemOperand x86memop> {
- def rri : AVX512AIi8<0x03, MRMSrcReg, (outs RC:$dst),
- (ins RC:$src1, RC:$src2, i8imm:$src3),
- !strconcat(OpcodeStr,
- " \t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
- []>, EVEX_4V;
+multiclass avx512_valign<X86VectorVTInfo _> {
+ defm rri : AVX512_masking<0x03, MRMSrcReg, (outs _.RC:$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 (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(OpcodeStr,
- " \t{$src3, $src2, $src1, $dst|$dst, $src1, $src2, $src3}"),
+ def rmi : AVX512AIi8<0x03, MRMSrcMem, (outs _.RC:$dst),
+ (ins _.RC:$src1, _.MemOp:$src2, i8imm:$src3),
+ !strconcat("valign"##_.Suffix,
+ " \t{$src3, $src2, $src1, $dst|"
+ "$dst, $src1, $src2, $src3}"),
[]>, EVEX_4V;
}
-defm VALIGND : avx512_alignr<"valignd", VR512, i512mem>,
- EVEX_V512, EVEX_CD8<32, CD8VF>;
-defm VALIGNQ : avx512_alignr<"valignq", VR512, i512mem>,
- VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
-
-def : Pat<(v16f32 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
- (VALIGNDrri VR512:$src2, VR512:$src1, imm:$imm)>;
-def : Pat<(v8f64 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
- (VALIGNQrri VR512:$src2, VR512:$src1, imm:$imm)>;
-def : Pat<(v16i32 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
- (VALIGNDrri VR512:$src2, VR512:$src1, imm:$imm)>;
-def : Pat<(v8i64 (X86PAlignr VR512:$src1, VR512:$src2, (i8 imm:$imm))),
- (VALIGNQrri VR512:$src2, VR512:$src1, imm:$imm)>;
+defm VALIGND : avx512_valign<v16i32_info>, EVEX_V512, EVEX_CD8<32, CD8VF>;
+defm VALIGNQ : avx512_valign<v8i64_info>, VEX_W, EVEX_V512, EVEX_CD8<64, CD8VF>;
// Helper fragments to match sext vXi1 to vXiY.
def v16i1sextv16i32 : PatLeaf<(v16i32 (X86vsrai VR512:$src, (i8 31)))>;