1 /* Definitions of x86 tunable features.
2 Copyright (C) 2013-2023 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.cc)
26 - introducing ix86_XXX_cost in i386.cc
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_LAKEMONT | m_PPRO | m_CORE_ALL | m_BONNELL | m_SILVERMONT
44 | m_INTEL | m_KNL | m_KNM | m_K6_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
45 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
46 | m_CORE_ATOM | m_GENERIC)
48 /* X86_TUNE_PARTIAL_REG_DEPENDENCY: Enable more register renaming
49 on modern chips. Prefer stores affecting whole integer register
50 over partial stores. For example prefer MOVZBL or MOVQ to load 8bit
52 DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY, "partial_reg_dependency",
53 m_P4_NOCONA | m_CORE2 | m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2
54 | m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL
55 | m_KNL | m_KNM | m_AMD_MULTIPLE | m_LUJIAZUI | m_TREMONT
56 | m_ALDERLAKE | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC)
58 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: This knob promotes all store
59 destinations to be 128bit to allow register renaming on 128bit SSE units,
60 but usually results in one extra microop on 64bit SSE units.
61 Experimental results shows that disabling this option on P4 brings over 20%
62 SPECfp regression, while enabling it on K8 brings roughly 2.4% regression
63 that can be partly masked by careful scheduling of moves. */
64 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY, "sse_partial_reg_dependency",
65 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
66 | m_BDVER | m_ZNVER | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE
67 | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC)
69 /* X86_TUNE_SSE_PARTIAL_REG_FP_CONVERTS_DEPENDENCY: This knob avoids
70 partial write to the destination in scalar SSE conversion from FP
72 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_FP_CONVERTS_DEPENDENCY,
73 "sse_partial_reg_fp_converts_dependency",
74 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
75 | m_BDVER | m_ZNVER | m_LUJIAZUI | m_ALDERLAKE | m_ARROWLAKE
76 | m_CORE_ATOM | m_GENERIC)
78 /* X86_TUNE_SSE_PARTIAL_REG_CONVERTS_DEPENDENCY: This knob avoids partial
79 write to the destination in scalar SSE conversion from integer to FP. */
80 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_CONVERTS_DEPENDENCY,
81 "sse_partial_reg_converts_dependency",
82 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
83 | m_BDVER | m_ZNVER | m_LUJIAZUI | m_ALDERLAKE | m_ARROWLAKE
84 | m_CORE_ATOM | m_GENERIC)
86 /* X86_TUNE_DEST_FALSE_DEP_FOR_GLC: This knob inserts zero-idiom before
87 several insns to break false dependency on the dest register for GLC
88 micro-architecture. */
89 DEF_TUNE (X86_TUNE_DEST_FALSE_DEP_FOR_GLC,
90 "dest_false_dep_for_glc", m_SAPPHIRERAPIDS | m_ALDERLAKE | m_ARROWLAKE
93 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
94 are resolved on SSE register parts instead of whole registers, so we may
95 maintain just lower part of scalar values in proper format leaving the
96 upper part undefined. */
97 DEF_TUNE (X86_TUNE_SSE_SPLIT_REGS, "sse_split_regs", m_ATHLON_K8)
99 /* X86_TUNE_PARTIAL_FLAG_REG_STALL: this flag disables use of flags
100 set by instructions affecting just some flags (in particular shifts).
101 This is because Core2 resolves dependencies on whole flags register
102 and such sequences introduce false dependency on previous instruction
105 The flags does not affect generation of INC and DEC that is controlled
106 by X86_TUNE_USE_INCDEC. */
108 DEF_TUNE (X86_TUNE_PARTIAL_FLAG_REG_STALL, "partial_flag_reg_stall",
111 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
112 partial dependencies. */
113 DEF_TUNE (X86_TUNE_MOVX, "movx",
114 m_PPRO | m_P4_NOCONA | m_CORE2 | m_NEHALEM | m_SANDYBRIDGE
115 | m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_KNL | m_KNM | m_INTEL
116 | m_GOLDMONT_PLUS | m_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
117 | m_CORE_AVX2 | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
118 | m_CORE_ATOM | m_GENERIC)
120 /* X86_TUNE_MEMORY_MISMATCH_STALL: Avoid partial stores that are followed by
122 DEF_TUNE (X86_TUNE_MEMORY_MISMATCH_STALL, "memory_mismatch_stall",
123 m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_INTEL
124 | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_AMD_MULTIPLE
125 | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
126 | m_CORE_ATOM | m_GENERIC)
128 /* X86_TUNE_FUSE_CMP_AND_BRANCH_32: Fuse compare with a subsequent
129 conditional jump instruction for 32 bit TARGET. */
130 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_32, "fuse_cmp_and_branch_32",
131 m_CORE_ALL | m_BDVER | m_ZNVER | m_LUJIAZUI | m_GENERIC)
133 /* X86_TUNE_FUSE_CMP_AND_BRANCH_64: Fuse compare with a subsequent
134 conditional jump instruction for TARGET_64BIT. */
135 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_64, "fuse_cmp_and_branch_64",
136 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_BDVER
137 | m_ZNVER | m_LUJIAZUI | m_GENERIC)
139 /* X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS: Fuse compare with a
140 subsequent conditional jump instruction when the condition jump
141 check sign flag (SF) or overflow flag (OF). */
142 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS, "fuse_cmp_and_branch_soflags",
143 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_BDVER
144 | m_ZNVER | m_LUJIAZUI | m_GENERIC)
146 /* X86_TUNE_FUSE_ALU_AND_BRANCH: Fuse alu with a subsequent conditional
147 jump instruction when the alu instruction produces the CCFLAG consumed by
148 the conditional jump instruction. */
149 DEF_TUNE (X86_TUNE_FUSE_ALU_AND_BRANCH, "fuse_alu_and_branch",
150 m_SANDYBRIDGE | m_CORE_AVX2 | m_LUJIAZUI | m_GENERIC)
153 /*****************************************************************************/
154 /* Function prologue, epilogue and function calling sequences. */
155 /*****************************************************************************/
157 /* X86_TUNE_ACCUMULATE_OUTGOING_ARGS: Allocate stack space for outgoing
158 arguments in prologue/epilogue instead of separately for each call
159 by push/pop instructions.
160 This increase code size by about 5% in 32bit mode, less so in 64bit mode
161 because parameters are passed in registers. It is considerable
162 win for targets without stack engine that prevents multple push operations
163 to happen in parallel. */
165 DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS, "accumulate_outgoing_args",
166 m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
167 | m_GOLDMONT | m_GOLDMONT_PLUS | m_ATHLON_K8 | m_LUJIAZUI)
169 /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in prologues that are
170 considered on critical path. */
171 DEF_TUNE (X86_TUNE_PROLOGUE_USING_MOVE, "prologue_using_move",
172 m_PPRO | m_ATHLON_K8)
174 /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in epilogues that are
175 considered on critical path. */
176 DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE, "epilogue_using_move",
177 m_PPRO | m_ATHLON_K8)
179 /* X86_TUNE_USE_LEAVE: Use "leave" instruction in epilogues where it fits. */
180 DEF_TUNE (X86_TUNE_USE_LEAVE, "use_leave",
181 m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
182 | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC)
184 /* X86_TUNE_PUSH_MEMORY: Enable generation of "push mem" instructions.
185 Some chips, like 486 and Pentium works faster with separate load
186 and push instructions. */
187 DEF_TUNE (X86_TUNE_PUSH_MEMORY, "push_memory",
188 m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE
189 | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
190 | m_CORE_ATOM | m_GENERIC)
192 /* X86_TUNE_SINGLE_PUSH: Enable if single push insn is preferred
193 over esp subtraction. */
194 DEF_TUNE (X86_TUNE_SINGLE_PUSH, "single_push", m_386 | m_486 | m_PENT
195 | m_LAKEMONT | m_K6_GEODE)
197 /* X86_TUNE_DOUBLE_PUSH. Enable if double push insn is preferred
198 over esp subtraction. */
199 DEF_TUNE (X86_TUNE_DOUBLE_PUSH, "double_push", m_PENT | m_LAKEMONT
202 /* X86_TUNE_SINGLE_POP: Enable if single pop insn is preferred
203 over esp addition. */
204 DEF_TUNE (X86_TUNE_SINGLE_POP, "single_pop", m_386 | m_486 | m_PENT
205 | m_LAKEMONT | m_PPRO)
207 /* X86_TUNE_DOUBLE_POP: Enable if double pop insn is preferred
208 over esp addition. */
209 DEF_TUNE (X86_TUNE_DOUBLE_POP, "double_pop", m_PENT | m_LAKEMONT)
211 /*****************************************************************************/
212 /* Branch predictor tuning */
213 /*****************************************************************************/
215 /* X86_TUNE_PAD_SHORT_FUNCTION: Make every function to be at least 4
216 instructions long. */
217 DEF_TUNE (X86_TUNE_PAD_SHORT_FUNCTION, "pad_short_function", m_BONNELL)
219 /* X86_TUNE_PAD_RETURNS: Place NOP before every RET that is a destination
220 of conditional jump or directly preceded by other jump instruction.
221 This is important for AND K8-AMDFAM10 because the branch prediction
222 architecture expect at most one jump per 2 byte window. Failing to
223 pad returns leads to misaligned return stack. */
224 DEF_TUNE (X86_TUNE_PAD_RETURNS, "pad_returns",
225 m_ATHLON_K8 | m_AMDFAM10)
227 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
228 than 4 branch instructions in the 16 byte window. */
229 DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT, "four_jump_limit",
230 m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM
231 | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL | m_ATHLON_K8 | m_AMDFAM10)
233 /*****************************************************************************/
234 /* Integer instruction selection tuning */
235 /*****************************************************************************/
237 /* X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL: Enable software prefetching
238 at -O3. For the moment, the prefetching seems badly tuned for Intel
240 DEF_TUNE (X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL, "software_prefetching_beneficial",
241 m_K6_GEODE | m_ATHLON_K8 | m_AMDFAM10 | m_BDVER | m_BTVER)
243 /* X86_TUNE_LCP_STALL: Avoid an expensive length-changing prefix stall
244 on 16-bit immediate moves into memory on Core2 and Corei7. */
245 DEF_TUNE (X86_TUNE_LCP_STALL, "lcp_stall", m_CORE_ALL | m_LUJIAZUI | m_GENERIC)
247 /* X86_TUNE_READ_MODIFY: Enable use of read-modify instructions such
248 as "add mem, reg". */
249 DEF_TUNE (X86_TUNE_READ_MODIFY, "read_modify", ~(m_PENT | m_LAKEMONT | m_PPRO))
251 /* X86_TUNE_USE_INCDEC: Enable use of inc/dec instructions.
253 Core2 and nehalem has stall of 7 cycles for partial flag register stalls.
254 Sandy bridge and Ivy bridge generate extra uop. On Haswell this extra uop
255 is output only when the values needs to be really merged, which is not
256 done by GCC generated code. */
257 DEF_TUNE (X86_TUNE_USE_INCDEC, "use_incdec",
258 ~(m_P4_NOCONA | m_CORE2 | m_NEHALEM | m_SANDYBRIDGE
259 | m_BONNELL | m_SILVERMONT | m_INTEL | m_KNL | m_KNM | m_GOLDMONT
260 | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
261 | m_CORE_ATOM | m_LUJIAZUI | m_GENERIC))
263 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
265 DEF_TUNE (X86_TUNE_INTEGER_DFMODE_MOVES, "integer_dfmode_moves",
266 ~(m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
267 | m_KNL | m_KNM | m_INTEL | m_GEODE | m_AMD_MULTIPLE | m_LUJIAZUI
268 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
269 | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC))
271 /* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag
272 will impact LEA instruction selection. */
273 DEF_TUNE (X86_TUNE_OPT_AGU, "opt_agu", m_BONNELL | m_SILVERMONT | m_KNL
274 | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL | m_LUJIAZUI)
276 /* X86_TUNE_AVOID_LEA_FOR_ADDR: Avoid lea for address computation. */
277 DEF_TUNE (X86_TUNE_AVOID_LEA_FOR_ADDR, "avoid_lea_for_addr",
278 m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS
281 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
282 vector path on AMD machines.
283 FIXME: Do we need to enable this for core? */
284 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM32_MEM, "slow_imul_imm32_mem",
287 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
289 FIXME: Do we need to enable this for core? */
290 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8, "slow_imul_imm8",
293 /* X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE: Try to avoid memory operands for
294 a conditional move. */
295 DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE, "avoid_mem_opnd_for_cmove",
296 m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_KNL
299 /* X86_TUNE_SINGLE_STRINGOP: Enable use of single string operations, such
300 as MOVS and STOS (without a REP prefix) to move/set sequences of bytes. */
301 DEF_TUNE (X86_TUNE_SINGLE_STRINGOP, "single_stringop", m_386 | m_P4_NOCONA)
303 /* X86_TUNE_PREFER_KNOWN_REP_MOVSB_STOSB: Enable use of REP MOVSB/STOSB to
304 move/set sequences of bytes with known size. */
305 DEF_TUNE (X86_TUNE_PREFER_KNOWN_REP_MOVSB_STOSB,
306 "prefer_known_rep_movsb_stosb",
307 m_SKYLAKE | m_ALDERLAKE | m_ARROWLAKE | m_CORE_ATOM
308 | m_TREMONT | m_CORE_AVX512 | m_LUJIAZUI)
310 /* X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES: Enable generation of
311 compact prologues and epilogues by issuing a misaligned moves. This
312 requires target to handle misaligned moves and partial memory stalls
314 FIXME: This may actualy be a win on more targets than listed here. */
315 DEF_TUNE (X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES,
316 "misaligned_move_string_pro_epilogues",
317 m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_LUJIAZUI | m_TREMONT
318 | m_ALDERLAKE | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC)
320 /* X86_TUNE_USE_SAHF: Controls use of SAHF. */
321 DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf",
322 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
323 | m_KNL | m_KNM | m_INTEL | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_BDVER
324 | m_BTVER | m_ZNVER | m_LUJIAZUI | m_GOLDMONT | m_GOLDMONT_PLUS
325 | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE | m_CORE_ATOM
328 /* X86_TUNE_USE_CLTD: Controls use of CLTD and CTQO instructions. */
329 DEF_TUNE (X86_TUNE_USE_CLTD, "use_cltd",
330 ~(m_PENT | m_LAKEMONT | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
331 | m_K6 | m_GOLDMONT | m_GOLDMONT_PLUS))
333 /* X86_TUNE_USE_BT: Enable use of BT (bit test) instructions. */
334 DEF_TUNE (X86_TUNE_USE_BT, "use_bt",
335 m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
336 | m_LAKEMONT | m_AMD_MULTIPLE | m_LUJIAZUI | m_GOLDMONT
337 | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
338 | m_CORE_ATOM | m_GENERIC)
340 /* X86_TUNE_AVOID_FALSE_DEP_FOR_BMI: Avoid false dependency
341 for bit-manipulation instructions. */
342 DEF_TUNE (X86_TUNE_AVOID_FALSE_DEP_FOR_BMI, "avoid_false_dep_for_bmi",
343 m_SANDYBRIDGE | m_HASWELL | m_SKYLAKE | m_SKYLAKE_AVX512
344 | m_CANNONLAKE | m_CASCADELAKE | m_COOPERLAKE
345 | m_LUJIAZUI | m_GENERIC)
347 /* X86_TUNE_ADJUST_UNROLL: This enables adjusting the unroll factor based
348 on hardware capabilities. Bdver3 hardware has a loop buffer which makes
349 unrolling small loop less important. For, such architectures we adjust
350 the unroll factor so that the unrolled loop fits the loop buffer. */
351 DEF_TUNE (X86_TUNE_ADJUST_UNROLL, "adjust_unroll_factor", m_BDVER3 | m_BDVER4)
353 /* X86_TUNE_ONE_IF_CONV_INSNS: Restrict a number of cmov insns in
354 if-converted sequence to one. */
355 DEF_TUNE (X86_TUNE_ONE_IF_CONV_INSN, "one_if_conv_insn",
356 m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_CORE_ALL | m_GOLDMONT
357 | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
358 | m_CORE_ATOM | m_LUJIAZUI | m_GENERIC)
360 /* X86_TUNE_AVOID_MFENCE: Use lock prefixed instructions instead of mfence. */
361 DEF_TUNE (X86_TUNE_AVOID_MFENCE, "avoid_mfence",
362 m_CORE_ALL | m_BDVER | m_ZNVER | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE
363 | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC)
365 /* X86_TUNE_EXPAND_ABS: This enables a new abs pattern by
366 generating instructions for abs (x) = (((signed) x >> (W-1) ^ x) -
367 (signed) x >> (W-1)) instead of cmove or SSE max/abs instructions. */
368 DEF_TUNE (X86_TUNE_EXPAND_ABS, "expand_abs",
369 m_CORE_ALL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
370 | m_GOLDMONT_PLUS | m_LUJIAZUI)
372 /*****************************************************************************/
373 /* 387 instruction selection tuning */
374 /*****************************************************************************/
376 /* X86_TUNE_USE_HIMODE_FIOP: Enables use of x87 instructions with 16bit
378 FIXME: Why this is disabled for modern chips? */
379 DEF_TUNE (X86_TUNE_USE_HIMODE_FIOP, "use_himode_fiop",
380 m_386 | m_486 | m_K6_GEODE)
382 /* X86_TUNE_USE_SIMODE_FIOP: Enables use of x87 instructions with 32bit
384 DEF_TUNE (X86_TUNE_USE_SIMODE_FIOP, "use_simode_fiop",
385 ~(m_PENT | m_LAKEMONT | m_PPRO | m_CORE_ALL | m_BONNELL
386 | m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_AMD_MULTIPLE
387 | m_LUJIAZUI | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT
388 | m_ALDERLAKE | m_ARROWLAKE | m_CORE_ATOM
391 /* X86_TUNE_USE_FFREEP: Use freep instruction instead of fstp. */
392 DEF_TUNE (X86_TUNE_USE_FFREEP, "use_ffreep", m_AMD_MULTIPLE | m_LUJIAZUI)
394 /* X86_TUNE_EXT_80387_CONSTANTS: Use fancy 80387 constants, such as PI. */
395 DEF_TUNE (X86_TUNE_EXT_80387_CONSTANTS, "ext_80387_constants",
396 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
397 | m_KNL | m_KNM | m_INTEL | m_K6_GEODE | m_ATHLON_K8 | m_LUJIAZUI
398 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
399 | m_CORE_ATOM | m_GENERIC)
401 /*****************************************************************************/
402 /* SSE instruction selection tuning */
403 /*****************************************************************************/
405 /* X86_TUNE_GENERAL_REGS_SSE_SPILL: Try to spill general regs to SSE
406 regs instead of memory. */
407 DEF_TUNE (X86_TUNE_GENERAL_REGS_SSE_SPILL, "general_regs_sse_spill",
410 /* X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL: Use movups for misaligned loads instead
411 of a sequence loading registers by parts. */
412 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL, "sse_unaligned_load_optimal",
413 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_SILVERMONT | m_KNL | m_KNM
414 | m_INTEL | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
415 | m_ARROWLAKE | m_CORE_ATOM | m_AMDFAM10 | m_BDVER
416 | m_BTVER | m_ZNVER | m_LUJIAZUI | m_GENERIC)
418 /* X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL: Use movups for misaligned stores
419 instead of a sequence loading registers by parts. */
420 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL, "sse_unaligned_store_optimal",
421 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_SILVERMONT | m_KNL | m_KNM
422 | m_INTEL | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
423 | m_ARROWLAKE | m_CORE_ATOM | m_BDVER | m_ZNVER
424 | m_LUJIAZUI | m_GENERIC)
426 /* X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL: Use packed single
427 precision 128bit instructions instead of double where possible. */
428 DEF_TUNE (X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL, "sse_packed_single_insn_optimal",
431 /* X86_TUNE_SSE_TYPELESS_STORES: Always movaps/movups for 128bit stores. */
432 DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES, "sse_typeless_stores",
433 m_AMD_MULTIPLE | m_LUJIAZUI | m_CORE_ALL | m_TREMONT | m_ALDERLAKE
434 | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC)
436 /* X86_TUNE_SSE_LOAD0_BY_PXOR: Always use pxor to load0 as opposed to
437 xorps/xorpd and other variants. */
438 DEF_TUNE (X86_TUNE_SSE_LOAD0_BY_PXOR, "sse_load0_by_pxor",
439 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BDVER | m_BTVER | m_ZNVER
440 | m_LUJIAZUI | m_TREMONT | m_ALDERLAKE | m_ARROWLAKE
441 | m_CORE_ATOM | m_GENERIC)
443 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from integer
444 to SSE registers. If disabled, the moves will be done by storing
445 the value to memory and reloading.
446 Enable this flag for generic - the only relevant architecture preferring
447 no inter-unit moves is Buldozer. While this makes small regression on SPECfp
448 scores (sub 0.3%), disabling inter-unit moves penalizes noticeably hand
449 written vectorized code which use i.e. _mm_set_epi16. */
450 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_TO_VEC, "inter_unit_moves_to_vec",
451 ~(m_ATHLON_K8 | m_AMDFAM10 | m_BDVER | m_BTVER))
453 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from SSE
454 to integer registers. If disabled, the moves will be done by storing
455 the value to memory and reloading. */
456 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_FROM_VEC, "inter_unit_moves_from_vec",
459 /* X86_TUNE_INTER_UNIT_CONVERSIONS: Enable float<->integer conversions
460 to use both SSE and integer registers at a same time. */
461 DEF_TUNE (X86_TUNE_INTER_UNIT_CONVERSIONS, "inter_unit_conversions",
462 ~(m_AMDFAM10 | m_BDVER))
464 /* X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS: Try to split memory operand for
465 fp converts to destination register. */
466 DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS, "split_mem_opnd_for_fp_converts",
467 m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS
470 /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
471 from FP to FP. This form of instructions avoids partial write to the
473 DEF_TUNE (X86_TUNE_USE_VECTOR_FP_CONVERTS, "use_vector_fp_converts",
476 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
477 from integer to FP. */
478 DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS, "use_vector_converts", m_AMDFAM10)
480 /* X86_TUNE_SLOW_SHUFB: Indicates tunings with slow pshufb instruction. */
481 DEF_TUNE (X86_TUNE_SLOW_PSHUFB, "slow_pshufb",
482 m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
483 | m_GOLDMONT_PLUS | m_INTEL)
485 /* X86_TUNE_AVOID_4BYTE_PREFIXES: Avoid instructions requiring 4+ bytes of prefixes. */
486 DEF_TUNE (X86_TUNE_AVOID_4BYTE_PREFIXES, "avoid_4byte_prefixes",
487 m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ALDERLAKE
488 | m_ARROWLAKE | m_CORE_ATOM | m_INTEL)
490 /* X86_TUNE_USE_GATHER_2PARTS: Use gather instructions for vectors with 2
492 DEF_TUNE (X86_TUNE_USE_GATHER_2PARTS, "use_gather_2parts",
493 ~(m_ZNVER1 | m_ZNVER2 | m_ZNVER3 | m_ZNVER4 | m_ALDERLAKE
494 | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC | m_GDS))
496 /* X86_TUNE_USE_SCATTER_2PARTS: Use scater instructions for vectors with 2
498 DEF_TUNE (X86_TUNE_USE_SCATTER_2PARTS, "use_scatter_2parts",
501 /* X86_TUNE_USE_GATHER_4PARTS: Use gather instructions for vectors with 4
503 DEF_TUNE (X86_TUNE_USE_GATHER_4PARTS, "use_gather_4parts",
504 ~(m_ZNVER1 | m_ZNVER2 | m_ZNVER3 | m_ZNVER4 | m_ALDERLAKE
505 | m_ARROWLAKE | m_CORE_ATOM | m_GENERIC | m_GDS))
507 /* X86_TUNE_USE_SCATTER_4PARTS: Use scater instructions for vectors with 4
509 DEF_TUNE (X86_TUNE_USE_SCATTER_4PARTS, "use_scatter_4parts",
512 /* X86_TUNE_USE_GATHER: Use gather instructions for vectors with 8 or more
514 DEF_TUNE (X86_TUNE_USE_GATHER_8PARTS, "use_gather_8parts",
515 ~(m_ZNVER1 | m_ZNVER2 | m_ZNVER4 | m_ALDERLAKE | m_ARROWLAKE
516 | m_CORE_ATOM | m_GENERIC | m_GDS))
518 /* X86_TUNE_USE_SCATTER: Use scater instructions for vectors with 8 or more
520 DEF_TUNE (X86_TUNE_USE_SCATTER_8PARTS, "use_scatter_8parts",
523 /* X86_TUNE_AVOID_128FMA_CHAINS: Avoid creating loops with tight 128bit or
524 smaller FMA chain. */
525 DEF_TUNE (X86_TUNE_AVOID_128FMA_CHAINS, "avoid_fma_chains", m_ZNVER1 | m_ZNVER2 | m_ZNVER3)
527 /* X86_TUNE_AVOID_256FMA_CHAINS: Avoid creating loops with tight 256bit or
528 smaller FMA chain. */
529 DEF_TUNE (X86_TUNE_AVOID_256FMA_CHAINS, "avoid_fma256_chains", m_ZNVER2 | m_ZNVER3
530 | m_ALDERLAKE | m_ARROWLAKE | m_SAPPHIRERAPIDS
533 /* X86_TUNE_AVOID_512FMA_CHAINS: Avoid creating loops with tight 512bit or
534 smaller FMA chain. */
535 DEF_TUNE (X86_TUNE_AVOID_512FMA_CHAINS, "avoid_fma512_chains", m_NONE)
537 /* X86_TUNE_V2DF_REDUCTION_PREFER_PHADDPD: Prefer haddpd
538 for v2df vector reduction. */
539 DEF_TUNE (X86_TUNE_V2DF_REDUCTION_PREFER_HADDPD,
540 "v2df_reduction_prefer_haddpd", m_NONE)
542 /*****************************************************************************/
543 /* AVX instruction selection tuning (some of SSE flags affects AVX, too) */
544 /*****************************************************************************/
546 /* X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL: if false, unaligned loads are
548 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL, "256_unaligned_load_optimal",
549 ~(m_NEHALEM | m_SANDYBRIDGE))
551 /* X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL: if false, unaligned stores are
553 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL, "256_unaligned_store_optimal",
554 ~(m_NEHALEM | m_SANDYBRIDGE | m_BDVER | m_ZNVER1))
556 /* X86_TUNE_AVX256_SPLIT_REGS: if true, AVX256 ops are split into two AVX128 ops. */
557 DEF_TUNE (X86_TUNE_AVX256_SPLIT_REGS, "avx256_split_regs",m_BDVER | m_BTVER2
560 /* X86_TUNE_AVX128_OPTIMAL: Enable 128-bit AVX instruction generation for
561 the auto-vectorizer. */
562 DEF_TUNE (X86_TUNE_AVX128_OPTIMAL, "avx128_optimal", m_BDVER | m_BTVER2
565 /* X86_TUNE_AVX256_OPTIMAL: Use 256-bit AVX instructions instead of 512-bit AVX
566 instructions in the auto-vectorizer. */
567 DEF_TUNE (X86_TUNE_AVX256_OPTIMAL, "avx256_optimal", m_CORE_AVX512)
569 /* X86_TUNE_AVX256_SPLIT_REGS: if true, AVX512 ops are split into two AVX256 ops. */
570 DEF_TUNE (X86_TUNE_AVX512_SPLIT_REGS, "avx512_split_regs", m_ZNVER4)
572 /* X86_TUNE_AVX256_MOVE_BY_PIECES: Optimize move_by_pieces with 256-bit
574 DEF_TUNE (X86_TUNE_AVX256_MOVE_BY_PIECES, "avx256_move_by_pieces",
575 m_ALDERLAKE | m_ARROWLAKE | m_CORE_AVX2 | m_ZNVER1
576 | m_ZNVER2 | m_ZNVER3)
578 /* X86_TUNE_AVX256_STORE_BY_PIECES: Optimize store_by_pieces with 256-bit
580 DEF_TUNE (X86_TUNE_AVX256_STORE_BY_PIECES, "avx256_store_by_pieces",
581 m_ALDERLAKE | m_ARROWLAKE | m_CORE_AVX2 | m_ZNVER1
582 | m_ZNVER2 | m_ZNVER3)
584 /* X86_TUNE_AVX512_MOVE_BY_PIECES: Optimize move_by_pieces with 512-bit
586 DEF_TUNE (X86_TUNE_AVX512_MOVE_BY_PIECES, "avx512_move_by_pieces",
587 m_SAPPHIRERAPIDS | m_ZNVER4)
589 /* X86_TUNE_AVX512_STORE_BY_PIECES: Optimize store_by_pieces with 512-bit
591 DEF_TUNE (X86_TUNE_AVX512_STORE_BY_PIECES, "avx512_store_by_pieces",
592 m_SAPPHIRERAPIDS | m_ZNVER4)
594 /*****************************************************************************/
595 /*****************************************************************************/
596 /* Historical relics: tuning flags that helps a specific old CPU designs */
597 /*****************************************************************************/
599 /* X86_TUNE_DOUBLE_WITH_ADD: Use add instead of sal to double value in
600 an integer register. */
601 DEF_TUNE (X86_TUNE_DOUBLE_WITH_ADD, "double_with_add", ~m_386)
603 /* X86_TUNE_ALWAYS_FANCY_MATH_387: controls use of fancy 387 operations,
604 such as fsqrt, fprem, fsin, fcos, fsincos etc.
605 Should be enabled for all targets that always has coprocesor. */
606 DEF_TUNE (X86_TUNE_ALWAYS_FANCY_MATH_387, "always_fancy_math_387",
607 ~(m_386 | m_486 | m_LAKEMONT))
609 /* X86_TUNE_UNROLL_STRLEN: Produce (quite lame) unrolled sequence for
610 inline strlen. This affects only -minline-all-stringops mode. By
611 default we always dispatch to a library since our internal strlen
613 DEF_TUNE (X86_TUNE_UNROLL_STRLEN, "unroll_strlen", ~m_386)
615 /* X86_TUNE_SHIFT1: Enables use of short encoding of "sal reg" instead of
616 longer "sal $1, reg". */
617 DEF_TUNE (X86_TUNE_SHIFT1, "shift1", ~m_486)
619 /* X86_TUNE_ZERO_EXTEND_WITH_AND: Use AND instruction instead
621 DEF_TUNE (X86_TUNE_ZERO_EXTEND_WITH_AND, "zero_extend_with_and",
624 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
625 and SImode multiply, but 386 and 486 do HImode multiply faster. */
626 DEF_TUNE (X86_TUNE_PROMOTE_HIMODE_IMUL, "promote_himode_imul",
629 /* X86_TUNE_FAST_PREFIX: Enable demoting some 32bit or 64bit arithmetic
630 into 16bit/8bit when resulting sequence is shorter. For example
631 for "and $-65536, reg" to 16bit store of 0. */
632 DEF_TUNE (X86_TUNE_FAST_PREFIX, "fast_prefix",
633 ~(m_386 | m_486 | m_PENT | m_LAKEMONT))
635 /* X86_TUNE_READ_MODIFY_WRITE: Enable use of read modify write instructions
636 such as "add $1, mem". */
637 DEF_TUNE (X86_TUNE_READ_MODIFY_WRITE, "read_modify_write",
638 ~(m_PENT | m_LAKEMONT))
640 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
642 DEF_TUNE (X86_TUNE_MOVE_M1_VIA_OR, "move_m1_via_or", m_PENT | m_LAKEMONT)
644 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
645 but one byte longer. */
646 DEF_TUNE (X86_TUNE_NOT_UNPAIRABLE, "not_unpairable", m_PENT | m_LAKEMONT)
648 /* X86_TUNE_PARTIAL_REG_STALL: Pentium pro, unlike later chips, handled
649 use of partial registers by renaming. This improved performance of 16bit
650 code where upper halves of registers are not used. It also leads to
651 an penalty whenever a 16bit store is followed by 32bit use. This flag
652 disables production of such sequences in common cases.
653 See also X86_TUNE_HIMODE_MATH.
655 In current implementation the partial register stalls are not eliminated
656 very well - they can be introduced via subregs synthesized by combine
657 and can happen in caller/callee saving sequences. */
658 DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL, "partial_reg_stall", m_PPRO)
660 /* X86_TUNE_PROMOTE_QIMODE: When it is cheap, turn 8bit arithmetic to
661 corresponding 32bit arithmetic. */
662 DEF_TUNE (X86_TUNE_PROMOTE_QIMODE, "promote_qimode",
665 /* X86_TUNE_PROMOTE_HI_REGS: Same, but for 16bit artihmetic. Again we avoid
666 partial register stalls on PentiumPro targets. */
667 DEF_TUNE (X86_TUNE_PROMOTE_HI_REGS, "promote_hi_regs", m_PPRO)
669 /* X86_TUNE_HIMODE_MATH: Enable use of 16bit arithmetic.
670 On PPro this flag is meant to avoid partial register stalls. */
671 DEF_TUNE (X86_TUNE_HIMODE_MATH, "himode_math", ~m_PPRO)
673 /* X86_TUNE_SPLIT_LONG_MOVES: Avoid instructions moving immediates
674 directly to memory. */
675 DEF_TUNE (X86_TUNE_SPLIT_LONG_MOVES, "split_long_moves", m_PPRO)
677 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
678 DEF_TUNE (X86_TUNE_USE_XCHGB, "use_xchgb", m_PENT4)
680 /* X86_TUNE_USE_MOV0: Use "mov $0, reg" instead of "xor reg, reg" to clear
682 DEF_TUNE (X86_TUNE_USE_MOV0, "use_mov0", m_K6)
684 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
685 operand that cannot be represented using a modRM byte. The XOR
686 replacement is long decoded, so this split helps here as well. */
687 DEF_TUNE (X86_TUNE_NOT_VECTORMODE, "not_vectormode", m_K6)
689 /* X86_TUNE_AVOID_VECTOR_DECODE: Enable splitters that avoid vector decoded
690 forms of instructions on K8 targets. */
691 DEF_TUNE (X86_TUNE_AVOID_VECTOR_DECODE, "avoid_vector_decode",
694 /*****************************************************************************/
695 /* This never worked well before. */
696 /*****************************************************************************/
698 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
699 on simulation result. But after P4 was made, no performance benefit
700 was observed with branch hints. It also increases the code size.
701 As a result, icc never generates branch hints. */
702 DEF_TUNE (X86_TUNE_BRANCH_PREDICTION_HINTS, "branch_prediction_hints", m_NONE)
704 /* X86_TUNE_QIMODE_MATH: Enable use of 8bit arithmetic. */
705 DEF_TUNE (X86_TUNE_QIMODE_MATH, "qimode_math", m_ALL)
707 /* X86_TUNE_PROMOTE_QI_REGS: This enables generic code that promotes all 8bit
708 arithmetic to 32bit via PROMOTE_MODE macro. This code generation scheme
709 is usually used for RISC targets. */
710 DEF_TUNE (X86_TUNE_PROMOTE_QI_REGS, "promote_qi_regs", m_NONE)
712 /* X86_TUNE_EMIT_VZEROUPPER: This enables vzeroupper instruction insertion
713 before a transfer of control flow out of the function. */
714 DEF_TUNE (X86_TUNE_EMIT_VZEROUPPER, "emit_vzeroupper", ~m_KNL)
716 /* X86_TUNE_SLOW_STC: This disables use of stc, clc and cmc carry flag
717 modifications on architectures where theses operations are slow. */
718 DEF_TUNE (X86_TUNE_SLOW_STC, "slow_stc", m_PENT4)