-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> {
+// 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:$dst, RHS)]>;
// Prefer over VMOV*rrk Pat<>
- let Constraints = "$src0 = $dst", AddedComplexity = 20 in
- def NAME#k: AVX512<O, F, Outs,
- !con((ins RC:$src0, KRC:$mask), Ins),
+ let AddedComplexity = 20 in
+ def NAME#k: AVX512<O, F, Outs, MaskingIns,
OpcodeStr#" \t{"#AttSrcAsm#", $dst {${mask}}|"#
"$dst {${mask}}, "#IntelSrcAsm#"}",
- [(set RC:$dst,
- (vselect KRC:$mask, RHS, RC:$src0))]>,
- EVEX_K;
+ [(set RC:$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,
- !con((ins KRC:$mask), Ins),
+ def NAME#kz: AVX512<O, F, Outs, ZeroMaskingIns,
OpcodeStr#" \t{"#AttSrcAsm#", $dst {${mask}} {z}|"#
"$dst {${mask}} {z}, "#IntelSrcAsm#"}",
[(set RC:$dst,
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 {
}
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 $src, (i32 1))), VK1)>;
multiclass avx512_fma3p_rm<bits<8> opc, string OpcodeStr,
RegisterClass RC, X86MemOperand x86memop,
PatFrag mem_frag, X86MemOperand x86scalar_mop, PatFrag scalar_mfrag,
- string BrdcstStr, SDNode OpNode, ValueType OpVT> {
- def r: AVX512FMA3<opc, MRMSrcReg, (outs RC:$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>;
}
projects / oota-llvm.git / blobdiff