This is an archive of the discontinued LLVM Phabricator instance.

Differential D37455

[X86] Introduce a new td file to hold patterns some of the non instruction patterns from SSE and AVX512
ClosedPublic

Authored by craig.topper on Sep 4 2017, 6:06 PM.

Details

Reviewers
zvi
spatel
RKSimon
guyblank
aymanmus
igorb
Commits
rGcf1d8a55f264: [X86] Introduce a new td file to hold patterns some of the non instruction…
rL312649: [X86] Introduce a new td file to hold patterns some of the non instruction…
Summary

This patch moves some of similar non-instruction patterns from X86InstrSSE.td and X86InstrAVX512.td to a common file.

This is intended as a starting point. There are many other optimization patterns that exist in both files that we could move here.

Diff Detail

Event Timeline

craig.topper created this revision.Sep 4 2017, 6:06 PM
craig.topper added reviewers: spatel, RKSimon, guyblank, aymanmus, igorb.
aymanmus edited edge metadata.Sep 5 2017, 1:31 AM

Why introduce new file instead of copying to X86InstrCompiler.td?

lib/Target/X86/X86InstrSSE.td
323

Remove comment please

craig.topper added a comment.Sep 5 2017, 8:58 AM

I created a new file because X86InstrCompiler.td is already 2000 lines long. There are still more things I want to move to the new file. So I was just trying to keep things manageable.

craig.topper updated this revision to Diff 113875.Sep 5 2017, 9:44 AM

Remove leftover comment

craig.topper mentioned this in D37503: [AVX512] Move more code to the proposed X86InstrVecCompiler.td.Sep 5 2017, 10:16 PM
aymanmus accepted this revision.Sep 6 2017, 12:50 AM

LGTM

This revision is now accepted and ready to land.Sep 6 2017, 12:50 AM
Closed by commit rL312649: [X86] Introduce a new td file to hold patterns some of the non instruction… (authored by ctopper). · Explain WhySep 6 2017, 9:58 AM
This revision was automatically updated to reflect the committed changes.

Revision Contents

PathSize
lib/
Target/
X86/
93 lines
1 line
115 lines
179 lines

Diff 113799

lib/Target/X86/X86InstrAVX512.td

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 422 Lines • ▼ Show 20 Lines AVX512_maskable_custom<O, F, Outs, Ins,
[(set _.RC:$dst, RHS)], [(set _.RC:$dst, RHS)],
[(set _.RC:$dst, [(set _.RC:$dst,
(Select _.KRCWM:$mask, MaskedRHS, _.RC:$src0))], (Select _.KRCWM:$mask, MaskedRHS, _.RC:$src0))],
[(set _.RC:$dst, [(set _.RC:$dst,
(Select _.KRCWM:$mask, MaskedRHS, (Select _.KRCWM:$mask, MaskedRHS,
_.ImmAllZerosV))], _.ImmAllZerosV))],
"$src0 = $dst", itin, IsCommutable>; "$src0 = $dst", itin, IsCommutable>;
// Bitcasts between 512-bit vector types. Return the original type since
// no instruction is needed for the conversion.
def : Pat<(v8f64 (bitconvert (v8i64 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v16i32 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v32i16 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v64i8 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v16f32 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v8i64 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v16i32 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v32i16 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v64i8 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v8f64 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v16i32 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v32i16 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v64i8 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v8f64 VR512:$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:$src)>;
def : Pat<(v32i16 (bitconvert (v8i64 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v16i32 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v64i8 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v8f64 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v16f32 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v16f32 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v8i64 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v16i32 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v32i16 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v8f64 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v16f32 VR512:$src))), (v64i8 VR512:$src)>;
// Alias instruction that maps zero vector to pxor / xorp* for AVX-512. // Alias instruction that maps zero vector to pxor / xorp* for AVX-512.
// This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then // This is expanded by ExpandPostRAPseudos to an xorps / vxorps, and then
// swizzled by ExecutionDepsFix to pxor. // swizzled by ExecutionDepsFix to pxor.
// We set canFoldAsLoad because this can be converted to a constant-pool // We set canFoldAsLoad because this can be converted to a constant-pool
// load of an all-zeros value if folding it would be beneficial. // load of an all-zeros value if folding it would be beneficial.
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
isPseudo = 1, Predicates = [HasAVX512], SchedRW = [WriteZero] in { isPseudo = 1, Predicates = [HasAVX512], SchedRW = [WriteZero] in {
▲ Show 20 Lines • Show All 326 Lines • ▼ Show 20 Lines
defm : vextract_for_size_lowering<"VEXTRACTI32x4Z", v64i8_info, v16i8x_info, defm : vextract_for_size_lowering<"VEXTRACTI32x4Z", v64i8_info, v16i8x_info,
vextract128_extract, EXTRACT_get_vextract128_imm, [HasAVX512]>; vextract128_extract, EXTRACT_get_vextract128_imm, [HasAVX512]>;
// Codegen pattern with the alternative types extract VEC256 from VEC512 // Codegen pattern with the alternative types extract VEC256 from VEC512
defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v32i16_info, v16i16x_info, defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v32i16_info, v16i16x_info,
vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>; vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>;
defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v64i8_info, v32i8x_info, defm : vextract_for_size_lowering<"VEXTRACTI64x4Z", v64i8_info, v32i8x_info,
vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>; vextract256_extract, EXTRACT_get_vextract256_imm, [HasAVX512]>;
// A 128-bit subvector extract from the first 256-bit vector position
// is a subregister copy that needs no instruction.
def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))),
(v2i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_xmm))>;
def : Pat<(v2f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))),
(v2f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_xmm))>;
def : Pat<(v4i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))),
(v4i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_xmm))>;
def : Pat<(v4f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))),
(v4f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_xmm))>;
def : Pat<(v8i16 (extract_subvector (v32i16 VR512:$src), (iPTR 0))),
(v8i16 (EXTRACT_SUBREG (v32i16 VR512:$src), sub_xmm))>;
def : Pat<(v16i8 (extract_subvector (v64i8 VR512:$src), (iPTR 0))),
(v16i8 (EXTRACT_SUBREG (v64i8 VR512:$src), sub_xmm))>;
// A 128-bit extract from bits [255:128] of a 512-bit vector should use a // A 128-bit extract from bits [255:128] of a 512-bit vector should use a
// smaller extract to enable EVEX->VEX. // smaller extract to enable EVEX->VEX.
let Predicates = [NoVLX] in { let Predicates = [NoVLX] in {
def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 2))), def : Pat<(v2i64 (extract_subvector (v8i64 VR512:$src), (iPTR 2))),
(v2i64 (VEXTRACTI128rr (v2i64 (VEXTRACTI128rr
(v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm)), (v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm)),
(iPTR 1)))>; (iPTR 1)))>;
▲ Show 20 Lines • Show All 43 Lines • ▼ Show 20 Linesdef : Pat<(v8i16 (extract_subvector (v32i16 VR512:$src), (iPTR 8))),
(v16i16 (EXTRACT_SUBREG (v32i16 VR512:$src), sub_ymm)), (v16i16 (EXTRACT_SUBREG (v32i16 VR512:$src), sub_ymm)),
(iPTR 1)))>; (iPTR 1)))>;
def : Pat<(v16i8 (extract_subvector (v64i8 VR512:$src), (iPTR 16))), def : Pat<(v16i8 (extract_subvector (v64i8 VR512:$src), (iPTR 16))),
(v16i8 (VEXTRACTI32x4Z256rr (v16i8 (VEXTRACTI32x4Z256rr
(v32i8 (EXTRACT_SUBREG (v64i8 VR512:$src), sub_ymm)), (v32i8 (EXTRACT_SUBREG (v64i8 VR512:$src), sub_ymm)),
(iPTR 1)))>; (iPTR 1)))>;
} }
// A 256-bit subvector extract from the first 256-bit vector position
// is a subregister copy that needs no instruction.
def : Pat<(v4i64 (extract_subvector (v8i64 VR512:$src), (iPTR 0))),
(v4i64 (EXTRACT_SUBREG (v8i64 VR512:$src), sub_ymm))>;
def : Pat<(v4f64 (extract_subvector (v8f64 VR512:$src), (iPTR 0))),
(v4f64 (EXTRACT_SUBREG (v8f64 VR512:$src), sub_ymm))>;
def : Pat<(v8i32 (extract_subvector (v16i32 VR512:$src), (iPTR 0))),
(v8i32 (EXTRACT_SUBREG (v16i32 VR512:$src), sub_ymm))>;
def : Pat<(v8f32 (extract_subvector (v16f32 VR512:$src), (iPTR 0))),
(v8f32 (EXTRACT_SUBREG (v16f32 VR512:$src), sub_ymm))>;
def : Pat<(v16i16 (extract_subvector (v32i16 VR512:$src), (iPTR 0))),
(v16i16 (EXTRACT_SUBREG (v32i16 VR512:$src), sub_ymm))>;
def : Pat<(v32i8 (extract_subvector (v64i8 VR512:$src), (iPTR 0))),
(v32i8 (EXTRACT_SUBREG (v64i8 VR512:$src), sub_ymm))>;
let AddedComplexity = 25 in { // to give priority over vinsertf128rm
// A 128-bit subvector insert to the first 512-bit vector position
// is a subregister copy that needs no instruction.
def : Pat<(v8i64 (insert_subvector undef, (v2i64 VR128X:$src), (iPTR 0))),
(INSERT_SUBREG (v8i64 (IMPLICIT_DEF)), VR128X:$src, sub_xmm)>;
def : Pat<(v8f64 (insert_subvector undef, (v2f64 VR128X:$src), (iPTR 0))),
(INSERT_SUBREG (v8f64 (IMPLICIT_DEF)), VR128X:$src, sub_xmm)>;
def : Pat<(v16i32 (insert_subvector undef, (v4i32 VR128X:$src), (iPTR 0))),
(INSERT_SUBREG (v16i32 (IMPLICIT_DEF)), VR128X:$src, sub_xmm)>;
def : Pat<(v16f32 (insert_subvector undef, (v4f32 VR128X:$src), (iPTR 0))),
(INSERT_SUBREG (v16f32 (IMPLICIT_DEF)), VR128X:$src, sub_xmm)>;
def : Pat<(v32i16 (insert_subvector undef, (v8i16 VR128X:$src), (iPTR 0))),
(INSERT_SUBREG (v32i16 (IMPLICIT_DEF)), VR128X:$src, sub_xmm)>;
def : Pat<(v64i8 (insert_subvector undef, (v16i8 VR128X:$src), (iPTR 0))),
(INSERT_SUBREG (v64i8 (IMPLICIT_DEF)), VR128X:$src, sub_xmm)>;
// A 256-bit subvector insert to the first 512-bit vector position
// is a subregister copy that needs no instruction.
def : Pat<(v8i64 (insert_subvector undef, (v4i64 VR256X:$src), (iPTR 0))),
(INSERT_SUBREG (v8i64 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
def : Pat<(v8f64 (insert_subvector undef, (v4f64 VR256X:$src), (iPTR 0))),
(INSERT_SUBREG (v8f64 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
def : Pat<(v16i32 (insert_subvector undef, (v8i32 VR256X:$src), (iPTR 0))),
(INSERT_SUBREG (v16i32 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
def : Pat<(v16f32 (insert_subvector undef, (v8f32 VR256X:$src), (iPTR 0))),
(INSERT_SUBREG (v16f32 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
def : Pat<(v32i16 (insert_subvector undef, (v16i16 VR256X:$src), (iPTR 0))),
(INSERT_SUBREG (v32i16 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
def : Pat<(v64i8 (insert_subvector undef, (v32i8 VR256X:$src), (iPTR 0))),
(INSERT_SUBREG (v64i8 (IMPLICIT_DEF)), VR256X:$src, sub_ymm)>;
}
// Additional patterns for handling a bitcast between the vselect and the // Additional patterns for handling a bitcast between the vselect and the
// extract_subvector. // extract_subvector.
multiclass vextract_for_mask_cast<string InstrStr, X86VectorVTInfo From, multiclass vextract_for_mask_cast<string InstrStr, X86VectorVTInfo From,
X86VectorVTInfo To, X86VectorVTInfo Cast, X86VectorVTInfo To, X86VectorVTInfo Cast,
PatFrag vextract_extract, PatFrag vextract_extract,
SDNodeXForm EXTRACT_get_vextract_imm, SDNodeXForm EXTRACT_get_vextract_imm,
list<Predicate> p> { list<Predicate> p> {
▲ Show 20 Lines • Show All 9,834 Lines • Show Last 20 Lines

lib/Target/X86/X86InstrInfo.td

Show First 20 Lines • Show All 2,761 Lines • ▼ Show 20 Lines
include "X86InstrTSX.td" include "X86InstrTSX.td"
include "X86InstrSGX.td" include "X86InstrSGX.td"
// System instructions. // System instructions.
include "X86InstrSystem.td" include "X86InstrSystem.td"
// Compiler Pseudo Instructions and Pat Patterns // Compiler Pseudo Instructions and Pat Patterns
include "X86InstrCompiler.td" include "X86InstrCompiler.td"
include "X86InstrVecCompiler.td"
//===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===//
// Assembler Mnemonic Aliases // Assembler Mnemonic Aliases
//===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===//
def : MnemonicAlias<"call", "callw", "att">, Requires<[In16BitMode]>; def : MnemonicAlias<"call", "callw", "att">, Requires<[In16BitMode]>;
def : MnemonicAlias<"call", "calll", "att">, Requires<[In32BitMode]>; def : MnemonicAlias<"call", "calll", "att">, Requires<[In32BitMode]>;
def : MnemonicAlias<"call", "callq", "att">, Requires<[In64BitMode]>; def : MnemonicAlias<"call", "callq", "att">, Requires<[In64BitMode]>;
▲ Show 20 Lines • Show All 543 Lines • Show Last 20 Lines

lib/Target/X86/X86InstrSSE.td

  • This file is larger than 256 KB, so syntax highlighting is disabled by default.
Show First 20 Lines • Show All 314 Lines • ▼ Show 20 Lines def rm : PI<opc, MRMSrcMem, (outs RC:$dst), (ins RC:$src1, x86memop:$src2),
!if(Is2Addr, !if(Is2Addr,
!strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"), !strconcat(OpcodeStr, "\t{$src2, $dst|$dst, $src2}"),
!strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")), !strconcat(OpcodeStr, "\t{$src2, $src1, $dst|$dst, $src1, $src2}")),
pat_rm, IIC_SSE_BIT_P_RM, d>, pat_rm, IIC_SSE_BIT_P_RM, d>,
Sched<[WriteVecLogicLd, ReadAfterLd]>; Sched<[WriteVecLogicLd, ReadAfterLd]>;
} }
//===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===//
// Non-instruction patterns // Non-instruction patterns
aymanmusUnsubmitted
Not Done
ReplyInline Actions

Remove comment please

aymanmus: Remove comment please
//===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===//
// A vector extract of the first f32/f64 position is a subregister copy
def : Pat<(f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
(COPY_TO_REGCLASS (v4f32 VR128:$src), FR32)>;
def : Pat<(f64 (extractelt (v2f64 VR128:$src), (iPTR 0))),
(COPY_TO_REGCLASS (v2f64 VR128:$src), FR64)>;
// A 128-bit subvector extract from the first 256-bit vector position
// is a subregister copy that needs no instruction.
def : Pat<(v4i32 (extract_subvector (v8i32 VR256:$src), (iPTR 0))),
(v4i32 (EXTRACT_SUBREG (v8i32 VR256:$src), sub_xmm))>;
def : Pat<(v4f32 (extract_subvector (v8f32 VR256:$src), (iPTR 0))),
(v4f32 (EXTRACT_SUBREG (v8f32 VR256:$src), sub_xmm))>;
def : Pat<(v2i64 (extract_subvector (v4i64 VR256:$src), (iPTR 0))),
(v2i64 (EXTRACT_SUBREG (v4i64 VR256:$src), sub_xmm))>;
def : Pat<(v2f64 (extract_subvector (v4f64 VR256:$src), (iPTR 0))),
(v2f64 (EXTRACT_SUBREG (v4f64 VR256:$src), sub_xmm))>;
def : Pat<(v8i16 (extract_subvector (v16i16 VR256:$src), (iPTR 0))),
(v8i16 (EXTRACT_SUBREG (v16i16 VR256:$src), sub_xmm))>;
def : Pat<(v16i8 (extract_subvector (v32i8 VR256:$src), (iPTR 0))),
(v16i8 (EXTRACT_SUBREG (v32i8 VR256:$src), sub_xmm))>;
// A 128-bit subvector insert to the first 256-bit vector position
// is a subregister copy that needs no instruction.
let AddedComplexity = 25 in { // to give priority over vinsertf128rm
def : Pat<(insert_subvector undef, (v2i64 VR128:$src), (iPTR 0)),
(INSERT_SUBREG (v4i64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
def : Pat<(insert_subvector undef, (v2f64 VR128:$src), (iPTR 0)),
(INSERT_SUBREG (v4f64 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
def : Pat<(insert_subvector undef, (v4i32 VR128:$src), (iPTR 0)),
(INSERT_SUBREG (v8i32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
def : Pat<(insert_subvector undef, (v4f32 VR128:$src), (iPTR 0)),
(INSERT_SUBREG (v8f32 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
def : Pat<(insert_subvector undef, (v8i16 VR128:$src), (iPTR 0)),
(INSERT_SUBREG (v16i16 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
def : Pat<(insert_subvector undef, (v16i8 VR128:$src), (iPTR 0)),
(INSERT_SUBREG (v32i8 (IMPLICIT_DEF)), VR128:$src, sub_xmm)>;
}
// Implicitly promote a 32-bit scalar to a vector.
def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
(COPY_TO_REGCLASS FR32:$src, VR128)>;
// Implicitly promote a 64-bit scalar to a vector.
def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
(COPY_TO_REGCLASS FR64:$src, VR128)>;
// Bitcasts between 128-bit vector types. Return the original type since
// no instruction is needed for the conversion
def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(f128 (bitconvert (i128 FR128:$src))), (f128 FR128:$src)>;
def : Pat<(i128 (bitconvert (f128 FR128:$src))), (i128 FR128:$src)>;
// Bitcasts between 256-bit vector types. Return the original type since
// no instruction is needed for the conversion
def : Pat<(v4i64 (bitconvert (v8i32 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v16i16 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v32i8 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v8f32 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v4f64 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v4i64 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v16i16 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v32i8 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v4f64 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v8f32 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v4i64 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v8i32 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v32i8 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v4f64 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v8f32 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v4i64 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v8i32 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v16i16 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v4f64 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v8f32 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v4i64 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v8i32 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v16i16 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v32i8 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v4f64 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v4i64 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v8i32 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v16i16 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v32i8 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v8f32 VR256:$src))), (v4f64 VR256:$src)>;
// Alias instructions that map fld0 to xorps for sse or vxorps for avx. // Alias instructions that map fld0 to xorps for sse or vxorps for avx.
// This is expanded by ExpandPostRAPseudos. // This is expanded by ExpandPostRAPseudos.
let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1, let isReMaterializable = 1, isAsCheapAsAMove = 1, canFoldAsLoad = 1,
isPseudo = 1, SchedRW = [WriteZero] in { isPseudo = 1, SchedRW = [WriteZero] in {
def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "", def FsFLD0SS : I<0, Pseudo, (outs FR32:$dst), (ins), "",
[(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1, NoAVX512]>; [(set FR32:$dst, fp32imm0)]>, Requires<[HasSSE1, NoAVX512]>;
def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "", def FsFLD0SD : I<0, Pseudo, (outs FR64:$dst), (ins), "",
[(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2, NoAVX512]>; [(set FR64:$dst, fpimm0)]>, Requires<[HasSSE2, NoAVX512]>;
▲ Show 20 Lines • Show All 8,221 Lines • Show Last 20 Lines

lib/Target/X86/X86InstrVecCompiler.td

  • This file was added.
//===- X86InstrVecCompiler.td - Vector Compiler Patterns ---*- tablegen -*-===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file describes the various vector pseudo instructions used by the
// compiler, as well as Pat patterns used during instruction selection.
//
//===----------------------------------------------------------------------===//
//===----------------------------------------------------------------------===//
// Non-instruction patterns
//===----------------------------------------------------------------------===//
// A vector extract of the first f32/f64 position is a subregister copy
def : Pat<(f32 (extractelt (v4f32 VR128:$src), (iPTR 0))),
(COPY_TO_REGCLASS (v4f32 VR128:$src), FR32)>;
def : Pat<(f64 (extractelt (v2f64 VR128:$src), (iPTR 0))),
(COPY_TO_REGCLASS (v2f64 VR128:$src), FR64)>;
// Implicitly promote a 32-bit scalar to a vector.
def : Pat<(v4f32 (scalar_to_vector FR32:$src)),
(COPY_TO_REGCLASS FR32:$src, VR128)>;
// Implicitly promote a 64-bit scalar to a vector.
def : Pat<(v2f64 (scalar_to_vector FR64:$src)),
(COPY_TO_REGCLASS FR64:$src, VR128)>;
// Patterns for insert_subvector/extract_subvector to/from index=0
multiclass subvector_subreg_lowering<RegisterClass subRC, ValueType subVT,
RegisterClass RC, ValueType VT,
SubRegIndex subIdx> {
def : Pat<(subVT (extract_subvector (VT RC:$src), (iPTR 0))),
(subVT (EXTRACT_SUBREG RC:$src, subIdx))>;
let AddedComplexity = 25 in // to give priority over vinsertf128rm
def : Pat<(VT (insert_subvector undef, subRC:$src, (iPTR 0))),
(VT (INSERT_SUBREG (IMPLICIT_DEF), subRC:$src, subIdx))>;
}
// A 128-bit subvector extract from the first 256-bit vector position is a
// subregister copy that needs no instruction. Likewise, a 128-bit subvector
// insert to the first 256-bit vector position is a subregister copy that needs
// no instruction.
defm : subvector_subreg_lowering<VR128, v4i32, VR256, v8i32, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v4f32, VR256, v8f32, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v2i64, VR256, v4i64, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v2f64, VR256, v4f64, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v8i16, VR256, v16i16, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v16i8, VR256, v32i8, sub_xmm>;
// A 128-bit subvector extract from the first 512-bit vector position is a
// subregister copy that needs no instruction. Likewise, a 128-bit subvector
// insert to the first 512-bit vector position is a subregister copy that needs
// no instruction.
defm : subvector_subreg_lowering<VR128, v4i32, VR512, v16i32, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v4f32, VR512, v16f32, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v2i64, VR512, v8i64, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v2f64, VR512, v8f64, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v8i16, VR512, v32i16, sub_xmm>;
defm : subvector_subreg_lowering<VR128, v16i8, VR512, v64i8, sub_xmm>;
// A 128-bit subvector extract from the first 512-bit vector position is a
// subregister copy that needs no instruction. Likewise, a 128-bit subvector
// insert to the first 512-bit vector position is a subregister copy that needs
// no instruction.
defm : subvector_subreg_lowering<VR256, v8i32, VR512, v16i32, sub_ymm>;
defm : subvector_subreg_lowering<VR256, v8f32, VR512, v16f32, sub_ymm>;
defm : subvector_subreg_lowering<VR256, v4i64, VR512, v8i64, sub_ymm>;
defm : subvector_subreg_lowering<VR256, v4f64, VR512, v8f64, sub_ymm>;
defm : subvector_subreg_lowering<VR256, v16i16, VR512, v32i16, sub_ymm>;
defm : subvector_subreg_lowering<VR256, v32i8, VR512, v64i8, sub_ymm>;
// Bitcasts between 128-bit vector types. Return the original type since
// no instruction is needed for the conversion
def : Pat<(v2i64 (bitconvert (v4i32 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v8i16 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v16i8 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v2f64 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v2i64 (bitconvert (v4f32 VR128:$src))), (v2i64 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v2i64 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v8i16 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v16i8 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v2f64 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v4i32 (bitconvert (v4f32 VR128:$src))), (v4i32 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v2i64 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v4i32 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v16i8 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v2f64 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v8i16 (bitconvert (v4f32 VR128:$src))), (v8i16 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v2i64 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v4i32 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v8i16 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v2f64 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v16i8 (bitconvert (v4f32 VR128:$src))), (v16i8 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v2i64 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v4i32 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v8i16 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v16i8 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v4f32 (bitconvert (v2f64 VR128:$src))), (v4f32 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v2i64 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v4i32 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v8i16 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v16i8 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(v2f64 (bitconvert (v4f32 VR128:$src))), (v2f64 VR128:$src)>;
def : Pat<(f128 (bitconvert (i128 FR128:$src))), (f128 FR128:$src)>;
def : Pat<(i128 (bitconvert (f128 FR128:$src))), (i128 FR128:$src)>;
// Bitcasts between 256-bit vector types. Return the original type since
// no instruction is needed for the conversion
def : Pat<(v4i64 (bitconvert (v8i32 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v16i16 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v32i8 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v8f32 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v4i64 (bitconvert (v4f64 VR256:$src))), (v4i64 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v4i64 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v16i16 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v32i8 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v4f64 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v8i32 (bitconvert (v8f32 VR256:$src))), (v8i32 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v4i64 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v8i32 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v32i8 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v4f64 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v16i16 (bitconvert (v8f32 VR256:$src))), (v16i16 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v4i64 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v8i32 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v16i16 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v4f64 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v32i8 (bitconvert (v8f32 VR256:$src))), (v32i8 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v4i64 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v8i32 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v16i16 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v32i8 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v8f32 (bitconvert (v4f64 VR256:$src))), (v8f32 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v4i64 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v8i32 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v16i16 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v32i8 VR256:$src))), (v4f64 VR256:$src)>;
def : Pat<(v4f64 (bitconvert (v8f32 VR256:$src))), (v4f64 VR256:$src)>;
// Bitcasts between 512-bit vector types. Return the original type since
// no instruction is needed for the conversion.
def : Pat<(v8f64 (bitconvert (v8i64 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v16i32 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v32i16 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v64i8 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v8f64 (bitconvert (v16f32 VR512:$src))), (v8f64 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v8i64 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v16i32 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v32i16 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v64i8 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v16f32 (bitconvert (v8f64 VR512:$src))), (v16f32 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v16i32 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v32i16 VR512:$src))), (v8i64 VR512:$src)>;
def : Pat<(v8i64 (bitconvert (v64i8 VR512:$src))), (v8i64 VR512:$src)>;
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)>;
def : Pat<(v32i16 (bitconvert (v8i64 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v16i32 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v64i8 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v8f64 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v16f32 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v32i16 (bitconvert (v16f32 VR512:$src))), (v32i16 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v8i64 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v16i32 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v32i16 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v8f64 VR512:$src))), (v64i8 VR512:$src)>;
def : Pat<(v64i8 (bitconvert (v16f32 VR512:$src))), (v64i8 VR512:$src)>;