1 /* Definitions of x86 tunable features.
2 Copyright (C
) 2013-2015 Free Software Foundation
, Inc.
4 This file is part of GCC.
6 GCC is free software
; you can redistribute it and
/or modify
7 it under the terms of the GNU General Public License as published by
8 the Free Software Foundation
; either version
3, or (at your option
)
11 GCC is distributed in the hope that it will be useful
,
12 but WITHOUT ANY WARRANTY
; without even the implied warranty of
13 MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the
14 GNU General Public License for more details.
16 You should have received a copy of the GNU General Public License and
17 a copy of the GCC Runtime Library Exception along with this program
;
18 see the files COPYING3 and COPYING.RUNTIME respectively. If not
, see
19 <http
://www.gnu.org
/licenses
/>.
*/
21 /* Tuning for a given CPU XXXX consists of
:
22 - adding new CPU into
:
23 - adding PROCESSOR_XXX to
processor_type (in i386.h
)
24 - possibly adding XXX into CPU attribute in i386.md
25 - adding XXX to
processor_alias_table (in i386.c
)
26 - introducing ix86_XXX_cost in i386.c
27 - Stringop generation table can be build based on test_stringop
28 - script (once rest of tuning is complete
)
29 - designing a scheduler model in
31 - Updating ix86_issue_rate and ix86_adjust_cost in i386.md
32 - possibly updating ia32_multipass_dfa_lookahead
, ix86_sched_reorder
33 and ix86_sched_init_global if those tricks are needed.
34 - Tunning the flags bellow. Those are split into sections and each
35 section is very roughly ordered by importance.
*/
37 /*****************************************************************************/
38 /* Scheduling flags.
*/
39 /*****************************************************************************/
41 /* X86_TUNE_SCHEDULE
: Enable scheduling.
*/
42 DEF_TUNE (X86_TUNE_SCHEDULE
, "schedule",
43 m_PENT | m_IAMCU | m_PPRO | m_CORE_ALL | m_BONNELL | m_SILVERMONT
44 | m_INTEL | m_KNL | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC
)
46 /* X86_TUNE_PARTIAL_REG_DEPENDENCY
: Enable more register renaming
47 on modern chips. Preffer stores affecting whole integer register
48 over partial stores. For example preffer MOVZBL or MOVQ to load
8bit
50 DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY
, "partial_reg_dependency",
51 m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_INTEL
52 | m_KNL | m_AMD_MULTIPLE | m_GENERIC
)
54 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY
: This knob promotes all store
55 destinations to be
128bit to allow register renaming on
128bit SSE units
,
56 but usually results in one extra microop on
64bit SSE units.
57 Experimental results shows that disabling this option on P4 brings over
20%
58 SPECfp regression
, while enabling it on K8 brings roughly
2.4% regression
59 that can be partly masked by careful scheduling of moves.
*/
60 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY
, "sse_partial_reg_dependency",
61 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
62 | m_BDVER | m_GENERIC
)
64 /* X86_TUNE_SSE_SPLIT_REGS
: Set for machines where the type and dependencies
65 are resolved on SSE register parts instead of whole registers
, so we may
66 maintain just lower part of scalar values in proper format leaving the
67 upper part undefined.
*/
68 DEF_TUNE (X86_TUNE_SSE_SPLIT_REGS
, "sse_split_regs", m_ATHLON_K8
)
70 /* X86_TUNE_PARTIAL_FLAG_REG_STALL
: this flag disables use of of flags
71 set by instructions affecting just some
flags (in particular shifts
).
72 This is because Core2 resolves dependencies on whole flags register
73 and such sequences introduce false dependency on previous instruction
76 The flags does not affect generation of
INC and
DEC that is controlled
77 by X86_TUNE_USE_INCDEC.
79 This flag may be dropped from generic once core2
-corei5 machines are
81 DEF_TUNE (X86_TUNE_PARTIAL_FLAG_REG_STALL
, "partial_flag_reg_stall",
84 /* X86_TUNE_MOVX
: Enable to zero extend integer registers to avoid
85 partial dependencies.
*/
86 DEF_TUNE (X86_TUNE_MOVX
, "movx",
87 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
88 | m_KNL | m_INTEL | m_GEODE | m_AMD_MULTIPLE | m_GENERIC
)
90 /* X86_TUNE_MEMORY_MISMATCH_STALL
: Avoid partial stores that are followed by
92 DEF_TUNE (X86_TUNE_MEMORY_MISMATCH_STALL
, "memory_mismatch_stall",
93 m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_INTEL
94 | m_KNL | m_AMD_MULTIPLE | m_GENERIC
)
96 /* X86_TUNE_FUSE_CMP_AND_BRANCH_32
: Fuse compare with a subsequent
97 conditional jump instruction for
32 bit TARGET.
98 FIXME
: revisit for generic.
*/
99 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_32
, "fuse_cmp_and_branch_32",
100 m_CORE_ALL | m_BDVER
)
102 /* X86_TUNE_FUSE_CMP_AND_BRANCH_64
: Fuse compare with a subsequent
103 conditional jump instruction for TARGET_64BIT.
104 FIXME
: revisit for generic.
*/
105 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_64
, "fuse_cmp_and_branch_64",
106 m_NEHALEM | m_SANDYBRIDGE | m_HASWELL | m_BDVER
)
108 /* X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS
: Fuse compare with a
109 subsequent conditional jump instruction when the condition jump
110 check sign
flag (SF
) or overflow
flag (OF).
*/
111 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS
, "fuse_cmp_and_branch_soflags",
112 m_NEHALEM | m_SANDYBRIDGE | m_HASWELL | m_BDVER
)
114 /* X86_TUNE_FUSE_ALU_AND_BRANCH
: Fuse alu with a subsequent conditional
115 jump instruction when the alu instruction produces the CCFLAG consumed by
116 the conditional jump instruction.
*/
117 DEF_TUNE (X86_TUNE_FUSE_ALU_AND_BRANCH
, "fuse_alu_and_branch",
118 m_SANDYBRIDGE | m_HASWELL
)
120 /* X86_TUNE_REASSOC_INT_TO_PARALLEL
: Try to produce parallel computations
121 during reassociation of integer computation.
*/
122 DEF_TUNE (X86_TUNE_REASSOC_INT_TO_PARALLEL
, "reassoc_int_to_parallel",
125 /* X86_TUNE_REASSOC_FP_TO_PARALLEL
: Try to produce parallel computations
126 during reassociation of fp computation.
*/
127 DEF_TUNE (X86_TUNE_REASSOC_FP_TO_PARALLEL
, "reassoc_fp_to_parallel",
128 m_BONNELL | m_SILVERMONT | m_HASWELL | m_KNL |m_INTEL | m_BDVER1
129 | m_BDVER2 | m_GENERIC
)
131 /*****************************************************************************/
132 /* Function prologue
, epilogue and function calling sequences.
*/
133 /*****************************************************************************/
135 /* X86_TUNE_ACCUMULATE_OUTGOING_ARGS
: Allocate stack space for outgoing
136 arguments in prologue
/epilogue instead of separately for each call
137 by push
/pop instructions.
138 This increase code size by about
5% in
32bit mode
, less so in
64bit mode
139 because parameters are passed in registers. It is considerable
140 win for targets without stack engine that prevents multple push operations
141 to happen in parallel.
143 FIXME
: the flags is incorrectly enabled for amdfam10
, Bulldozer
,
144 Bobcat and Generic. This is because disabling it causes large
145 regression on mgrid due to IRA limitation leading to unecessary
146 use of the frame pointer in
32bit mode.
*/
147 DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS
, "accumulate_outgoing_args",
148 m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_INTEL
151 /* X86_TUNE_PROLOGUE_USING_MOVE
: Do not use push
/pop in prologues that are
152 considered on critical path.
*/
153 DEF_TUNE (X86_TUNE_PROLOGUE_USING_MOVE
, "prologue_using_move",
154 m_PPRO | m_ATHLON_K8
)
156 /* X86_TUNE_PROLOGUE_USING_MOVE
: Do not use push
/pop in epilogues that are
157 considered on critical path.
*/
158 DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE
, "epilogue_using_move",
159 m_PPRO | m_ATHLON_K8
)
161 /* X86_TUNE_USE_LEAVE
: Use
"leave" instruction in epilogues where it fits.
*/
162 DEF_TUNE (X86_TUNE_USE_LEAVE
, "use_leave",
163 m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC
)
165 /* X86_TUNE_PUSH_MEMORY
: Enable generation of
"push mem" instructions.
166 Some chips
, like
486 and Pentium works faster with separate load
167 and push instructions.
*/
168 DEF_TUNE (X86_TUNE_PUSH_MEMORY
, "push_memory",
169 m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE
172 /* X86_TUNE_SINGLE_PUSH
: Enable if single push insn is preferred
173 over esp subtraction.
*/
174 DEF_TUNE (X86_TUNE_SINGLE_PUSH
, "single_push", m_386 | m_486 | m_PENT
175 | m_IAMCU | m_K6_GEODE
)
177 /* X86_TUNE_DOUBLE_PUSH. Enable if double push insn is preferred
178 over esp subtraction.
*/
179 DEF_TUNE (X86_TUNE_DOUBLE_PUSH
, "double_push", m_PENT | m_IAMCU
182 /* X86_TUNE_SINGLE_POP
: Enable if single pop insn is preferred
183 over esp addition.
*/
184 DEF_TUNE (X86_TUNE_SINGLE_POP
, "single_pop", m_386 | m_486 | m_PENT
187 /* X86_TUNE_DOUBLE_POP
: Enable if double pop insn is preferred
188 over esp addition.
*/
189 DEF_TUNE (X86_TUNE_DOUBLE_POP
, "double_pop", m_PENT | m_IAMCU
)
191 /*****************************************************************************/
192 /* Branch predictor tuning
*/
193 /*****************************************************************************/
195 /* X86_TUNE_PAD_SHORT_FUNCTION
: Make every function to be at least
4
196 instructions long.
*/
197 DEF_TUNE (X86_TUNE_PAD_SHORT_FUNCTION
, "pad_short_function", m_BONNELL
)
199 /* X86_TUNE_PAD_RETURNS
: Place NOP before every RET that is a destination
200 of conditional jump or directly preceded by other jump instruction.
201 This is important for
AND K8
-AMDFAM10 because the branch prediction
202 architecture expect at most one jump per
2 byte window. Failing to
203 pad returns leads to misaligned return stack.
*/
204 DEF_TUNE (X86_TUNE_PAD_RETURNS
, "pad_returns",
205 m_ATHLON_K8 | m_AMDFAM10 | m_GENERIC
)
207 /* X86_TUNE_FOUR_JUMP_LIMIT
: Some CPU cores are not able to predict more
208 than
4 branch instructions in the
16 byte window.
*/
209 DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT
, "four_jump_limit",
210 m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL |m_INTEL |
211 m_ATHLON_K8 | m_AMDFAM10
)
213 /*****************************************************************************/
214 /* Integer instruction selection tuning
*/
215 /*****************************************************************************/
217 /* X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL
: Enable software prefetching
218 at
-O3. For the moment
, the prefetching seems badly tuned for Intel
220 DEF_TUNE (X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL
, "software_prefetching_beneficial",
221 m_K6_GEODE | m_AMD_MULTIPLE
)
223 /* X86_TUNE_LCP_STALL
: Avoid an expensive length
-changing prefix stall
224 on
16-bit immediate moves into memory on Core2 and Corei7.
*/
225 DEF_TUNE (X86_TUNE_LCP_STALL
, "lcp_stall", m_CORE_ALL | m_GENERIC
)
227 /* X86_TUNE_READ_MODIFY
: Enable use of read
-modify instructions such
228 as
"add mem, reg".
*/
229 DEF_TUNE (X86_TUNE_READ_MODIFY
, "read_modify", ~
(m_PENT | m_IAMCU | m_PPRO
))
231 /* X86_TUNE_USE_INCDEC
: Enable use of inc
/dec instructions.
*/
232 DEF_TUNE (X86_TUNE_USE_INCDEC
, "use_incdec",
233 ~
(m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_INTEL
234 | m_KNL | m_GENERIC
))
236 /* X86_TUNE_INTEGER_DFMODE_MOVES
: Enable if integer moves are preferred
238 DEF_TUNE (X86_TUNE_INTEGER_DFMODE_MOVES
, "integer_dfmode_moves",
239 ~
(m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
240 | m_KNL | m_INTEL | m_GEODE | m_AMD_MULTIPLE | m_GENERIC
))
242 /* X86_TUNE_OPT_AGU
: Optimize for Address Generation Unit. This flag
243 will impact LEA instruction selection.
*/
244 DEF_TUNE (X86_TUNE_OPT_AGU
, "opt_agu", m_BONNELL | m_SILVERMONT | m_KNL
247 /* X86_TUNE_AVOID_LEA_FOR_ADDR
: Avoid lea for address computation.
*/
248 DEF_TUNE (X86_TUNE_AVOID_LEA_FOR_ADDR
, "avoid_lea_for_addr",
249 m_BONNELL | m_SILVERMONT | m_KNL
)
251 /* X86_TUNE_SLOW_IMUL_IMM32_MEM
: Imul of
32-bit constant and memory is
252 vector path on AMD machines.
253 FIXME
: Do we need to enable this for core?
*/
254 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM32_MEM
, "slow_imul_imm32_mem",
257 /* X86_TUNE_SLOW_IMUL_IMM8
: Imul of
8-bit constant is vector path on AMD
259 FIXME
: Do we need to enable this for core?
*/
260 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8
, "slow_imul_imm8",
263 /* X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE
: Try to avoid memory operands for
264 a conditional move.
*/
265 DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE
, "avoid_mem_opnd_for_cmove",
266 m_BONNELL | m_SILVERMONT | m_KNL | m_INTEL
)
268 /* X86_TUNE_SINGLE_STRINGOP
: Enable use of single string operations
, such
269 as MOVS and
STOS (without a REP prefix
) to move
/set sequences of bytes.
*/
270 DEF_TUNE (X86_TUNE_SINGLE_STRINGOP
, "single_stringop", m_386 | m_P4_NOCONA
)
272 /* X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES
: Enable generation of
273 compact prologues and epilogues by issuing a misaligned moves. This
274 requires target to handle misaligned moves and partial memory stalls
276 FIXME
: This may actualy be a win on more targets than listed here.
*/
277 DEF_TUNE (X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES
,
278 "misaligned_move_string_pro_epilogues",
279 m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_GENERIC
)
281 /* X86_TUNE_USE_SAHF
: Controls use of SAHF.
*/
282 DEF_TUNE (X86_TUNE_USE_SAHF
, "use_sahf",
283 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
284 | m_KNL | m_INTEL | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_BDVER
285 | m_BTVER | m_GENERIC
)
287 /* X86_TUNE_USE_CLTD
: Controls use of CLTD and CTQO instructions.
*/
288 DEF_TUNE (X86_TUNE_USE_CLTD
, "use_cltd",
289 ~
(m_PENT | m_IAMCU | m_BONNELL | m_SILVERMONT | m_KNL | m_INTEL
292 /* X86_TUNE_USE_BT
: Enable use of
BT (bit test
) instructions.
*/
293 DEF_TUNE (X86_TUNE_USE_BT
, "use_bt",
294 m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_KNL | m_INTEL
295 | m_AMD_MULTIPLE | m_GENERIC
)
297 /*****************************************************************************/
298 /* 387 instruction selection tuning
*/
299 /*****************************************************************************/
301 /* X86_TUNE_USE_HIMODE_FIOP
: Enables use of x87 instructions with
16bit
303 FIXME
: Why this is disabled for modern chips?
*/
304 DEF_TUNE (X86_TUNE_USE_HIMODE_FIOP
, "use_himode_fiop",
305 m_386 | m_486 | m_K6_GEODE
)
307 /* X86_TUNE_USE_SIMODE_FIOP
: Enables use of x87 instructions with
32bit
309 DEF_TUNE (X86_TUNE_USE_SIMODE_FIOP
, "use_simode_fiop",
310 ~
(m_PENT | m_IAMCU | m_PPRO | m_CORE_ALL | m_BONNELL
311 | m_SILVERMONT | m_KNL | m_INTEL | m_AMD_MULTIPLE | m_GENERIC
))
313 /* X86_TUNE_USE_FFREEP
: Use freep instruction instead of fstp.
*/
314 DEF_TUNE (X86_TUNE_USE_FFREEP
, "use_ffreep", m_AMD_MULTIPLE
)
316 /* X86_TUNE_EXT_80387_CONSTANTS
: Use fancy
80387 constants
, such as PI.
*/
317 DEF_TUNE (X86_TUNE_EXT_80387_CONSTANTS
, "ext_80387_constants",
318 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
319 | m_KNL | m_INTEL | m_K6_GEODE | m_ATHLON_K8 | m_GENERIC
)
321 /*****************************************************************************/
322 /* SSE instruction selection tuning
*/
323 /*****************************************************************************/
325 /* X86_TUNE_VECTORIZE_DOUBLE
: Enable double precision vector
327 DEF_TUNE (X86_TUNE_VECTORIZE_DOUBLE
, "vectorize_double", ~m_BONNELL
)
329 /* X86_TUNE_GENERAL_REGS_SSE_SPILL
: Try to spill general regs to SSE
330 regs instead of memory.
*/
331 DEF_TUNE (X86_TUNE_GENERAL_REGS_SSE_SPILL
, "general_regs_sse_spill",
334 /* X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL
: Use movups for misaligned loads instead
335 of a sequence loading registers by parts.
*/
336 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL
, "sse_unaligned_load_optimal",
337 m_NEHALEM | m_SANDYBRIDGE | m_HASWELL | m_AMDFAM10 | m_BDVER
338 | m_BTVER | m_SILVERMONT | m_KNL | m_INTEL | m_GENERIC
)
340 /* X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL
: Use movups for misaligned stores instead
341 of a sequence loading registers by parts.
*/
342 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL
, "sse_unaligned_store_optimal",
343 m_NEHALEM | m_SANDYBRIDGE | m_HASWELL | m_BDVER | m_SILVERMONT
344 | m_KNL | m_INTEL | m_GENERIC
)
346 /* Use packed single precision instructions where posisble. I.e. movups instead
348 DEF_TUNE (X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL
, "sse_packed_single_insn_optimal",
351 /* X86_TUNE_SSE_TYPELESS_STORES
: Always movaps
/movups for
128bit stores.
*/
352 DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES
, "sse_typeless_stores",
353 m_AMD_MULTIPLE | m_CORE_ALL | m_GENERIC
)
355 /* X86_TUNE_SSE_LOAD0_BY_PXOR
: Always use pxor to load0 as opposed to
356 xorps
/xorpd and other variants.
*/
357 DEF_TUNE (X86_TUNE_SSE_LOAD0_BY_PXOR
, "sse_load0_by_pxor",
358 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BDVER | m_BTVER | m_GENERIC
)
360 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC
: Enable moves in from integer
361 to SSE registers. If disabled
, the moves will be done by storing
362 the value to memory and reloading.
*/
363 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_TO_VEC
, "inter_unit_moves_to_vec",
364 ~
(m_AMD_MULTIPLE | m_GENERIC
))
366 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC
: Enable moves in from SSE
367 to integer registers. If disabled
, the moves will be done by storing
368 the value to memory and reloading.
*/
369 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_FROM_VEC
, "inter_unit_moves_from_vec",
372 /* X86_TUNE_INTER_UNIT_CONVERSIONS
: Enable float
<->integer conversions
373 to use both SSE and integer registers at a same time.
374 FIXME
: revisit importance of this for generic.
*/
375 DEF_TUNE (X86_TUNE_INTER_UNIT_CONVERSIONS
, "inter_unit_conversions",
376 ~
(m_AMDFAM10 | m_BDVER
))
378 /* X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS
: Try to split memory operand for
379 fp converts to destination register.
*/
380 DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS
, "split_mem_opnd_for_fp_converts",
381 m_SILVERMONT | m_KNL | m_INTEL
)
383 /* X86_TUNE_USE_VECTOR_FP_CONVERTS
: Prefer vector packed SSE conversion
384 from FP to FP. This form of instructions avoids partial write to the
386 DEF_TUNE (X86_TUNE_USE_VECTOR_FP_CONVERTS
, "use_vector_fp_converts",
389 /* X86_TUNE_USE_VECTOR_CONVERTS
: Prefer vector packed SSE conversion
390 from integer to FP.
*/
391 DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS
, "use_vector_converts", m_AMDFAM10
)
393 /* X86_TUNE_SLOW_SHUFB
: Indicates tunings with slow pshufb instruction.
*/
394 DEF_TUNE (X86_TUNE_SLOW_PSHUFB
, "slow_pshufb",
395 m_BONNELL | m_SILVERMONT | m_KNL | m_INTEL
)
397 /* X86_TUNE_VECTOR_PARALLEL_EXECUTION
: Indicates tunings with ability to
398 execute
2 or more vector instructions in parallel.
*/
399 DEF_TUNE (X86_TUNE_VECTOR_PARALLEL_EXECUTION
, "vec_parallel",
400 m_NEHALEM | m_SANDYBRIDGE | m_HASWELL
)
402 /* X86_TUNE_AVOID_4BYTE_PREFIXES
: Avoid instructions requiring
4+ bytes of prefixes.
*/
403 DEF_TUNE (X86_TUNE_AVOID_4BYTE_PREFIXES
, "avoid_4byte_prefixes",
404 m_SILVERMONT | m_INTEL
)
406 /*****************************************************************************/
407 /* AVX instruction selection
tuning (some of SSE flags affects AVX
, too
) */
408 /*****************************************************************************/
410 /* X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL
: if false
, unaligned loads are
412 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL
, "256_unaligned_load_optimal",
413 ~
(m_NEHALEM | m_SANDYBRIDGE | m_GENERIC
))
415 /* X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL
: if false
, unaligned stores are
417 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL
, "256_unaligned_store_optimal",
418 ~
(m_NEHALEM | m_SANDYBRIDGE | m_BDVER | m_GENERIC
))
420 /* X86_TUNE_AVX128_OPTIMAL
: Enable
128-bit AVX instruction generation for
421 the auto
-vectorizer.
*/
422 DEF_TUNE (X86_TUNE_AVX128_OPTIMAL
, "avx128_optimal", m_BDVER | m_BTVER2
)
424 /*****************************************************************************/
425 /* Historical relics
: tuning flags that helps a specific old CPU designs
*/
426 /*****************************************************************************/
428 /* X86_TUNE_DOUBLE_WITH_ADD
: Use add instead of sal to double value in
429 an integer register.
*/
430 DEF_TUNE (X86_TUNE_DOUBLE_WITH_ADD
, "double_with_add", ~m_386
)
432 /* X86_TUNE_ALWAYS_FANCY_MATH_387
: controls use of fancy
387 operations
,
433 such as fsqrt
, fprem
, fsin
, fcos
, fsincos etc.
434 Should be enabled for all targets that always has coprocesor.
*/
435 DEF_TUNE (X86_TUNE_ALWAYS_FANCY_MATH_387
, "always_fancy_math_387",
438 /* X86_TUNE_UNROLL_STRLEN
: Produce (quite lame
) unrolled sequence for
439 inline strlen. This affects only
-minline
-all
-stringops mode. By
440 default we always dispatch to a library since our internal strlen
442 DEF_TUNE (X86_TUNE_UNROLL_STRLEN
, "unroll_strlen", ~m_386
)
444 /* X86_TUNE_SHIFT1
: Enables use of short encoding of
"sal reg" instead of
445 longer
"sal $1, reg".
*/
446 DEF_TUNE (X86_TUNE_SHIFT1
, "shift1", ~m_486
)
448 /* X86_TUNE_ZERO_EXTEND_WITH_AND
: Use
AND instruction instead
450 DEF_TUNE (X86_TUNE_ZERO_EXTEND_WITH_AND
, "zero_extend_with_and",
451 m_486 | m_PENT | m_IAMCU
)
453 /* X86_TUNE_PROMOTE_HIMODE_IMUL
: Modern CPUs have same latency for HImode
454 and SImode multiply
, but
386 and
486 do HImode multiply faster.
*/
455 DEF_TUNE (X86_TUNE_PROMOTE_HIMODE_IMUL
, "promote_himode_imul",
458 /* X86_TUNE_FAST_PREFIX
: Enable demoting some
32bit or
64bit arithmetic
459 into
16bit
/8bit when resulting sequence is shorter. For example
460 for
"and $-65536, reg" to
16bit store of
0.
*/
461 DEF_TUNE (X86_TUNE_FAST_PREFIX
, "fast_prefix",
462 ~
(m_386 | m_486 | m_PENT | m_IAMCU
))
464 /* X86_TUNE_READ_MODIFY_WRITE
: Enable use of read modify write instructions
465 such as
"add $1, mem".
*/
466 DEF_TUNE (X86_TUNE_READ_MODIFY_WRITE
, "read_modify_write",
469 /* X86_TUNE_MOVE_M1_VIA_OR
: On pentiums
, it is faster to load
-1 via
OR
471 DEF_TUNE (X86_TUNE_MOVE_M1_VIA_OR
, "move_m1_via_or", m_PENT | m_IAMCU
)
473 /* X86_TUNE_NOT_UNPAIRABLE
: NOT is not pairable on Pentium
, while XOR is
,
474 but one byte longer.
*/
475 DEF_TUNE (X86_TUNE_NOT_UNPAIRABLE
, "not_unpairable", m_PENT | m_IAMCU
)
477 /* X86_TUNE_PARTIAL_REG_STALL
: Pentium pro
, unlike later chips
, handled
478 use of partial registers by renaming. This improved performance of
16bit
479 code where upper halves of registers are not used. It also leads to
480 an penalty whenever a
16bit store is followed by
32bit use. This flag
481 disables production of such sequences in common cases.
482 See also X86_TUNE_HIMODE_MATH.
484 In current implementation the partial register stalls are not eliminated
485 very well
- they can be introduced via subregs synthesized by combine
486 and can happen in caller
/callee saving sequences.
*/
487 DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL
, "partial_reg_stall", m_PPRO
)
489 /* X86_TUNE_PROMOTE_QIMODE
: When it is cheap
, turn
8bit arithmetic to
490 corresponding
32bit arithmetic.
*/
491 DEF_TUNE (X86_TUNE_PROMOTE_QIMODE
, "promote_qimode",
494 /* X86_TUNE_PROMOTE_HI_REGS
: Same
, but for
16bit artihmetic. Again we avoid
495 partial register stalls on PentiumPro targets.
*/
496 DEF_TUNE (X86_TUNE_PROMOTE_HI_REGS
, "promote_hi_regs", m_PPRO
)
498 /* X86_TUNE_HIMODE_MATH
: Enable use of
16bit arithmetic.
499 On PPro this flag is meant to avoid partial register stalls.
*/
500 DEF_TUNE (X86_TUNE_HIMODE_MATH
, "himode_math", ~m_PPRO
)
502 /* X86_TUNE_SPLIT_LONG_MOVES
: Avoid instructions moving immediates
503 directly to memory.
*/
504 DEF_TUNE (X86_TUNE_SPLIT_LONG_MOVES
, "split_long_moves", m_PPRO
)
506 /* X86_TUNE_USE_XCHGB
: Use xchgb
%rh
,%rl instead of rolw
/rorw $
8,rx.
*/
507 DEF_TUNE (X86_TUNE_USE_XCHGB
, "use_xchgb", m_PENT4
)
509 /* X86_TUNE_USE_MOV0
: Use
"mov $0, reg" instead of
"xor reg, reg" to clear
511 DEF_TUNE (X86_TUNE_USE_MOV0
, "use_mov0", m_K6
)
513 /* X86_TUNE_NOT_VECTORMODE
: On AMD K6
, NOT is vector decoded with memory
514 operand that cannot be represented using a modRM byte. The XOR
515 replacement is long decoded
, so this split helps here as well.
*/
516 DEF_TUNE (X86_TUNE_NOT_VECTORMODE
, "not_vectormode", m_K6
)
518 /* X86_TUNE_AVOID_VECTOR_DECODE
: Enable splitters that avoid vector decoded
519 forms of instructions on K8 targets.
*/
520 DEF_TUNE (X86_TUNE_AVOID_VECTOR_DECODE
, "avoid_vector_decode",
523 /* X86_TUNE_AVOID_FALSE_DEP_FOR_BMI
: Avoid false dependency
524 for bit
-manipulation instructions.
*/
525 DEF_TUNE (X86_TUNE_AVOID_FALSE_DEP_FOR_BMI
, "avoid_false_dep_for_bmi",
526 m_SANDYBRIDGE | m_HASWELL | m_GENERIC
)
528 /*****************************************************************************/
529 /* This never worked well before.
*/
530 /*****************************************************************************/
532 /* X86_TUNE_BRANCH_PREDICTION_HINTS
: Branch hints were put in P4 based
533 on simulation result. But after P4 was made
, no performance benefit
534 was observed with branch hints. It also increases the code size.
535 As a result
, icc never generates branch hints.
*/
536 DEF_TUNE (X86_TUNE_BRANCH_PREDICTION_HINTS
, "branch_prediction_hints", 0)
538 /* X86_TUNE_QIMODE_MATH
: Enable use of
8bit arithmetic.
*/
539 DEF_TUNE (X86_TUNE_QIMODE_MATH
, "qimode_math", ~
0)
541 /* X86_TUNE_PROMOTE_QI_REGS
: This enables generic code that promotes all
8bit
542 arithmetic to
32bit via PROMOTE_MODE macro. This code generation scheme
543 is usually used for RISC targets.
*/
544 DEF_TUNE (X86_TUNE_PROMOTE_QI_REGS
, "promote_qi_regs", 0)
546 /* X86_TUNE_ADJUST_UNROLL
: This enables adjusting the unroll factor based
547 on hardware capabilities. Bdver3 hardware has a loop buffer which makes
548 unrolling small loop less important. For
, such architectures we adjust
549 the unroll factor so that the unrolled loop fits the loop buffer.
*/
550 DEF_TUNE (X86_TUNE_ADJUST_UNROLL
, "adjust_unroll_factor", m_BDVER3 | m_BDVER4
)