diff options
Diffstat (limited to 'lib/Target/WebAssembly/WebAssemblyInstrSIMD.td')
| -rw-r--r-- | lib/Target/WebAssembly/WebAssemblyInstrSIMD.td | 221 |
1 files changed, 187 insertions, 34 deletions
diff --git a/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td b/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td index dd8930f079b0..fc5d73dac52e 100644 --- a/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td +++ b/lib/Target/WebAssembly/WebAssemblyInstrSIMD.td @@ -40,47 +40,124 @@ def LaneIdx#SIZE : ImmLeaf<i32, "return 0 <= Imm && Imm < "#SIZE#";">; //===----------------------------------------------------------------------===// // Load: v128.load -multiclass SIMDLoad<ValueType vec_t> { - let mayLoad = 1, UseNamedOperandTable = 1 in - defm LOAD_#vec_t : +let mayLoad = 1, UseNamedOperandTable = 1 in +defm LOAD_V128 : + SIMD_I<(outs V128:$dst), (ins P2Align:$p2align, offset32_op:$off, I32:$addr), + (outs), (ins P2Align:$p2align, offset32_op:$off), [], + "v128.load\t$dst, ${off}(${addr})$p2align", + "v128.load\t$off$p2align", 0>; + +// Def load and store patterns from WebAssemblyInstrMemory.td for vector types +foreach vec_t = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in { +def : LoadPatNoOffset<vec_t, load, LOAD_V128>; +def : LoadPatImmOff<vec_t, load, regPlusImm, LOAD_V128>; +def : LoadPatImmOff<vec_t, load, or_is_add, LOAD_V128>; +def : LoadPatOffsetOnly<vec_t, load, LOAD_V128>; +def : LoadPatGlobalAddrOffOnly<vec_t, load, LOAD_V128>; +} + +// vNxM.load_splat +multiclass SIMDLoadSplat<string vec, bits<32> simdop> { + let mayLoad = 1, UseNamedOperandTable = 1, + Predicates = [HasUnimplementedSIMD128] in + defm LOAD_SPLAT_#vec : SIMD_I<(outs V128:$dst), (ins P2Align:$p2align, offset32_op:$off, I32:$addr), (outs), (ins P2Align:$p2align, offset32_op:$off), [], - "v128.load\t$dst, ${off}(${addr})$p2align", - "v128.load\t$off$p2align", 0>; + vec#".load_splat\t$dst, ${off}(${addr})$p2align", + vec#".load_splat\t$off$p2align", simdop>; } -foreach vec_t = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in { -defm "" : SIMDLoad<vec_t>; - -// Def load and store patterns from WebAssemblyInstrMemory.td for vector types -def : LoadPatNoOffset<vec_t, load, !cast<NI>("LOAD_"#vec_t)>; -def : LoadPatImmOff<vec_t, load, regPlusImm, !cast<NI>("LOAD_"#vec_t)>; -def : LoadPatImmOff<vec_t, load, or_is_add, !cast<NI>("LOAD_"#vec_t)>; -def : LoadPatGlobalAddr<vec_t, load, !cast<NI>("LOAD_"#vec_t)>; -def : LoadPatOffsetOnly<vec_t, load, !cast<NI>("LOAD_"#vec_t)>; -def : LoadPatGlobalAddrOffOnly<vec_t, load, !cast<NI>("LOAD_"#vec_t)>; +defm "" : SIMDLoadSplat<"v8x16", 194>; +defm "" : SIMDLoadSplat<"v16x8", 195>; +defm "" : SIMDLoadSplat<"v32x4", 196>; +defm "" : SIMDLoadSplat<"v64x2", 197>; + +def wasm_load_splat_t : SDTypeProfile<1, 1, []>; +def wasm_load_splat : SDNode<"WebAssemblyISD::LOAD_SPLAT", wasm_load_splat_t>; + +foreach args = [["v16i8", "i32", "extloadi8"], ["v8i16", "i32", "extloadi16"], + ["v4i32", "i32", "load"], ["v2i64", "i64", "load"], + ["v4f32", "f32", "load"], ["v2f64", "f64", "load"]] in +def load_splat_#args[0] : + PatFrag<(ops node:$addr), (wasm_load_splat + (!cast<ValueType>(args[1]) (!cast<PatFrag>(args[2]) node:$addr)))>; + +let Predicates = [HasUnimplementedSIMD128] in +foreach args = [["v16i8", "v8x16"], ["v8i16", "v16x8"], ["v4i32", "v32x4"], + ["v2i64", "v64x2"], ["v4f32", "v32x4"], ["v2f64", "v64x2"]] in { +def : LoadPatNoOffset<!cast<ValueType>(args[0]), + !cast<PatFrag>("load_splat_"#args[0]), + !cast<NI>("LOAD_SPLAT_"#args[1])>; +def : LoadPatImmOff<!cast<ValueType>(args[0]), + !cast<PatFrag>("load_splat_"#args[0]), + regPlusImm, + !cast<NI>("LOAD_SPLAT_"#args[1])>; +def : LoadPatImmOff<!cast<ValueType>(args[0]), + !cast<PatFrag>("load_splat_"#args[0]), + or_is_add, + !cast<NI>("LOAD_SPLAT_"#args[1])>; +def : LoadPatOffsetOnly<!cast<ValueType>(args[0]), + !cast<PatFrag>("load_splat_"#args[0]), + !cast<NI>("LOAD_SPLAT_"#args[1])>; +def : LoadPatGlobalAddrOffOnly<!cast<ValueType>(args[0]), + !cast<PatFrag>("load_splat_"#args[0]), + !cast<NI>("LOAD_SPLAT_"#args[1])>; } -// Store: v128.store -multiclass SIMDStore<ValueType vec_t> { - let mayStore = 1, UseNamedOperandTable = 1 in - defm STORE_#vec_t : - SIMD_I<(outs), (ins P2Align:$p2align, offset32_op:$off, I32:$addr, V128:$vec), +// Load and extend +multiclass SIMDLoadExtend<ValueType vec_t, string name, bits<32> simdop> { + let mayLoad = 1, UseNamedOperandTable = 1, + Predicates = [HasUnimplementedSIMD128] in { + defm LOAD_EXTEND_S_#vec_t : + SIMD_I<(outs V128:$dst), (ins P2Align:$p2align, offset32_op:$off, I32:$addr), + (outs), (ins P2Align:$p2align, offset32_op:$off), [], + name#"_s\t$dst, ${off}(${addr})$p2align", + name#"_s\t$off$p2align", simdop>; + defm LOAD_EXTEND_U_#vec_t : + SIMD_I<(outs V128:$dst), (ins P2Align:$p2align, offset32_op:$off, I32:$addr), (outs), (ins P2Align:$p2align, offset32_op:$off), [], - "v128.store\t${off}(${addr})$p2align, $vec", - "v128.store\t$off$p2align", 1>; + name#"_u\t$dst, ${off}(${addr})$p2align", + name#"_u\t$off$p2align", !add(simdop, 1)>; + } } -foreach vec_t = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in { -defm "" : SIMDStore<vec_t>; +defm "" : SIMDLoadExtend<v8i16, "i16x8.load8x8", 210>; +defm "" : SIMDLoadExtend<v4i32, "i32x4.load16x4", 212>; +defm "" : SIMDLoadExtend<v2i64, "i64x2.load32x2", 214>; + +let Predicates = [HasUnimplementedSIMD128] in +foreach types = [[v8i16, i8], [v4i32, i16], [v2i64, i32]] in +foreach exts = [["sextloadv", "_S"], + ["zextloadv", "_U"], + ["extloadv", "_U"]] in { +def : LoadPatNoOffset<types[0], !cast<PatFrag>(exts[0]#types[1]), + !cast<NI>("LOAD_EXTEND"#exts[1]#"_"#types[0])>; +def : LoadPatImmOff<types[0], !cast<PatFrag>(exts[0]#types[1]), regPlusImm, + !cast<NI>("LOAD_EXTEND"#exts[1]#"_"#types[0])>; +def : LoadPatImmOff<types[0], !cast<PatFrag>(exts[0]#types[1]), or_is_add, + !cast<NI>("LOAD_EXTEND"#exts[1]#"_"#types[0])>; +def : LoadPatOffsetOnly<types[0], !cast<PatFrag>(exts[0]#types[1]), + !cast<NI>("LOAD_EXTEND"#exts[1]#"_"#types[0])>; +def : LoadPatGlobalAddrOffOnly<types[0], !cast<PatFrag>(exts[0]#types[1]), + !cast<NI>("LOAD_EXTEND"#exts[1]#"_"#types[0])>; +} + + +// Store: v128.store +let mayStore = 1, UseNamedOperandTable = 1 in +defm STORE_V128 : + SIMD_I<(outs), (ins P2Align:$p2align, offset32_op:$off, I32:$addr, V128:$vec), + (outs), (ins P2Align:$p2align, offset32_op:$off), [], + "v128.store\t${off}(${addr})$p2align, $vec", + "v128.store\t$off$p2align", 1>; +foreach vec_t = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in { // Def load and store patterns from WebAssemblyInstrMemory.td for vector types -def : StorePatNoOffset<vec_t, store, !cast<NI>("STORE_"#vec_t)>; -def : StorePatImmOff<vec_t, store, regPlusImm, !cast<NI>("STORE_"#vec_t)>; -def : StorePatImmOff<vec_t, store, or_is_add, !cast<NI>("STORE_"#vec_t)>; -def : StorePatGlobalAddr<vec_t, store, !cast<NI>("STORE_"#vec_t)>; -def : StorePatOffsetOnly<vec_t, store, !cast<NI>("STORE_"#vec_t)>; -def : StorePatGlobalAddrOffOnly<vec_t, store, !cast<NI>("STORE_"#vec_t)>; +def : StorePatNoOffset<vec_t, store, STORE_V128>; +def : StorePatImmOff<vec_t, store, regPlusImm, STORE_V128>; +def : StorePatImmOff<vec_t, store, or_is_add, STORE_V128>; +def : StorePatOffsetOnly<vec_t, store, STORE_V128>; +def : StorePatGlobalAddrOffOnly<vec_t, store, STORE_V128>; } //===----------------------------------------------------------------------===// @@ -90,7 +167,7 @@ def : StorePatGlobalAddrOffOnly<vec_t, store, !cast<NI>("STORE_"#vec_t)>; // Constant: v128.const multiclass ConstVec<ValueType vec_t, dag ops, dag pat, string args> { let isMoveImm = 1, isReMaterializable = 1, - Predicates = [HasSIMD128, HasUnimplementedSIMD128] in + Predicates = [HasUnimplementedSIMD128] in defm CONST_V128_#vec_t : SIMD_I<(outs V128:$dst), ops, (outs), ops, [(set V128:$dst, (vec_t pat))], "v128.const\t$dst, "#args, @@ -198,6 +275,19 @@ def : Pat<(vec_t (wasm_shuffle (vec_t V128:$x), (vec_t V128:$y), (i32 LaneIdx32:$mE), (i32 LaneIdx32:$mF)))>; } +// Swizzle lanes: v8x16.swizzle +def wasm_swizzle_t : SDTypeProfile<1, 2, []>; +def wasm_swizzle : SDNode<"WebAssemblyISD::SWIZZLE", wasm_swizzle_t>; +let Predicates = [HasUnimplementedSIMD128] in +defm SWIZZLE : + SIMD_I<(outs V128:$dst), (ins V128:$src, V128:$mask), (outs), (ins), + [(set (v16i8 V128:$dst), + (wasm_swizzle (v16i8 V128:$src), (v16i8 V128:$mask)))], + "v8x16.swizzle\t$dst, $src, $mask", "v8x16.swizzle", 192>; + +def : Pat<(int_wasm_swizzle (v16i8 V128:$src), (v16i8 V128:$mask)), + (SWIZZLE V128:$src, V128:$mask)>; + // Create vector with identical lanes: splat def splat2 : PatFrag<(ops node:$x), (build_vector node:$x, node:$x)>; def splat4 : PatFrag<(ops node:$x), (build_vector @@ -286,7 +376,7 @@ multiclass ExtractLaneExtended<string sign, bits<32> baseInst> { } defm "" : ExtractLaneExtended<"_s", 5>; -let Predicates = [HasSIMD128, HasUnimplementedSIMD128] in +let Predicates = [HasUnimplementedSIMD128] in defm "" : ExtractLaneExtended<"_u", 6>; defm "" : ExtractLane<v4i32, "i32x4", LaneIdx4, I32, 13>; defm "" : ExtractLane<v2i64, "i64x2", LaneIdx2, I64, 16>; @@ -472,6 +562,11 @@ defm OR : SIMDBitwise<or, "or", 78>; defm XOR : SIMDBitwise<xor, "xor", 79>; } // isCommutable = 1 +// Bitwise logic: v128.andnot +def andnot : PatFrag<(ops node:$left, node:$right), (and $left, (vnot $right))>; +let Predicates = [HasUnimplementedSIMD128] in +defm ANDNOT : SIMDBitwise<andnot, "andnot", 216>; + // Bitwise select: v128.bitselect foreach vec_t = [v16i8, v8i16, v4i32, v2i64, v4f32, v2f64] in defm BITSELECT_#vec_t : @@ -655,7 +750,7 @@ defm ABS : SIMDUnaryFP<fabs, "abs", 149>; defm NEG : SIMDUnaryFP<fneg, "neg", 150>; // Square root: sqrt -let Predicates = [HasSIMD128, HasUnimplementedSIMD128] in +let Predicates = [HasUnimplementedSIMD128] in defm SQRT : SIMDUnaryFP<fsqrt, "sqrt", 151>; //===----------------------------------------------------------------------===// @@ -679,7 +774,7 @@ let isCommutable = 1 in defm MUL : SIMDBinaryFP<fmul, "mul", 156>; // Division: div -let Predicates = [HasSIMD128, HasUnimplementedSIMD128] in +let Predicates = [HasUnimplementedSIMD128] in defm DIV : SIMDBinaryFP<fdiv, "div", 157>; // NaN-propagating minimum: min @@ -712,6 +807,42 @@ defm "" : SIMDConvert<v4i32, v4f32, fp_to_uint, "i32x4.trunc_sat_f32x4_u", 172>; defm "" : SIMDConvert<v2i64, v2f64, fp_to_sint, "i64x2.trunc_sat_f64x2_s", 173>; defm "" : SIMDConvert<v2i64, v2f64, fp_to_uint, "i64x2.trunc_sat_f64x2_u", 174>; +// Widening operations +multiclass SIMDWiden<ValueType vec_t, string vec, ValueType arg_t, string arg, + bits<32> baseInst> { + defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_low_signed, + vec#".widen_low_"#arg#"_s", baseInst>; + defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_high_signed, + vec#".widen_high_"#arg#"_s", !add(baseInst, 1)>; + defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_low_unsigned, + vec#".widen_low_"#arg#"_u", !add(baseInst, 2)>; + defm "" : SIMDConvert<vec_t, arg_t, int_wasm_widen_high_unsigned, + vec#".widen_high_"#arg#"_u", !add(baseInst, 3)>; +} + +defm "" : SIMDWiden<v8i16, "i16x8", v16i8, "i8x16", 202>; +defm "" : SIMDWiden<v4i32, "i32x4", v8i16, "i16x8", 206>; + +// Narrowing operations +multiclass SIMDNarrow<ValueType vec_t, string vec, ValueType arg_t, string arg, + bits<32> baseInst> { + defm NARROW_S_#vec_t : + SIMD_I<(outs V128:$dst), (ins V128:$low, V128:$high), (outs), (ins), + [(set (vec_t V128:$dst), (vec_t (int_wasm_narrow_signed + (arg_t V128:$low), (arg_t V128:$high))))], + vec#".narrow_"#arg#"_s\t$dst, $low, $high", vec#".narrow_"#arg#"_s", + baseInst>; + defm NARROW_U_#vec_t : + SIMD_I<(outs V128:$dst), (ins V128:$low, V128:$high), (outs), (ins), + [(set (vec_t V128:$dst), (vec_t (int_wasm_narrow_unsigned + (arg_t V128:$low), (arg_t V128:$high))))], + vec#".narrow_"#arg#"_u\t$dst, $low, $high", vec#".narrow_"#arg#"_u", + !add(baseInst, 1)>; +} + +defm "" : SIMDNarrow<v16i8, "i8x16", v8i16, "i16x8", 198>; +defm "" : SIMDNarrow<v8i16, "i16x8", v4i32, "i32x4", 200>; + // Lower llvm.wasm.trunc.saturate.* to saturating instructions def : Pat<(v4i32 (int_wasm_trunc_saturate_signed (v4f32 V128:$src))), (fp_to_sint_v4i32_v4f32 (v4f32 V128:$src))>; @@ -732,3 +863,25 @@ foreach t2 = !foldl( ) ) in def : Pat<(t1 (bitconvert (t2 V128:$v))), (t1 V128:$v)>; + +//===----------------------------------------------------------------------===// +// Quasi-Fused Multiply- Add and Subtract (QFMA/QFMS) +//===----------------------------------------------------------------------===// + +multiclass SIMDQFM<ValueType vec_t, string vec, bits<32> baseInst> { + defm QFMA_#vec_t : + SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c), + (outs), (ins), + [(set (vec_t V128:$dst), + (int_wasm_qfma (vec_t V128:$a), (vec_t V128:$b), (vec_t V128:$c)))], + vec#".qfma\t$dst, $a, $b, $c", vec#".qfma", baseInst>; + defm QFMS_#vec_t : + SIMD_I<(outs V128:$dst), (ins V128:$a, V128:$b, V128:$c), + (outs), (ins), + [(set (vec_t V128:$dst), + (int_wasm_qfms (vec_t V128:$a), (vec_t V128:$b), (vec_t V128:$c)))], + vec#".qfms\t$dst, $a, $b, $c", vec#".qfms", !add(baseInst, 1)>; +} + +defm "" : SIMDQFM<v4f32, "f32x4", 0x98>; +defm "" : SIMDQFM<v2f64, "f64x2", 0xa3>; |