111
|
1 /* Definitions of x86 tunable features.
|
131
|
2 Copyright (C) 2013-2018 Free Software Foundation, Inc.
|
111
|
3
|
|
4 This file is part of GCC.
|
|
5
|
|
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)
|
|
9 any later version.
|
|
10
|
|
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.
|
|
15
|
|
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/>. */
|
|
20
|
|
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
|
|
30 - XXXX.md file
|
|
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. */
|
|
36
|
|
37 /*****************************************************************************/
|
|
38 /* Scheduling flags. */
|
|
39 /*****************************************************************************/
|
|
40
|
|
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
|
131
|
44 | m_INTEL | m_KNL | m_KNM | m_K6_GEODE | m_AMD_MULTIPLE | m_GOLDMONT
|
|
45 | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC)
|
111
|
46
|
|
47 /* X86_TUNE_PARTIAL_REG_DEPENDENCY: Enable more register renaming
|
|
48 on modern chips. Preffer stores affecting whole integer register
|
|
49 over partial stores. For example preffer MOVZBL or MOVQ to load 8bit
|
|
50 value over movb. */
|
|
51 DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY, "partial_reg_dependency",
|
131
|
52 m_P4_NOCONA | m_CORE2 | m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2
|
|
53 | m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_INTEL
|
|
54 | m_KNL | m_KNM | m_AMD_MULTIPLE | m_TREMONT
|
|
55 | m_GENERIC)
|
111
|
56
|
|
57 /* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: This knob promotes all store
|
|
58 destinations to be 128bit to allow register renaming on 128bit SSE units,
|
|
59 but usually results in one extra microop on 64bit SSE units.
|
|
60 Experimental results shows that disabling this option on P4 brings over 20%
|
|
61 SPECfp regression, while enabling it on K8 brings roughly 2.4% regression
|
|
62 that can be partly masked by careful scheduling of moves. */
|
|
63 DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY, "sse_partial_reg_dependency",
|
|
64 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_AMDFAM10
|
|
65 | m_BDVER | m_ZNVER1 | m_GENERIC)
|
|
66
|
|
67 /* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies
|
|
68 are resolved on SSE register parts instead of whole registers, so we may
|
|
69 maintain just lower part of scalar values in proper format leaving the
|
|
70 upper part undefined. */
|
|
71 DEF_TUNE (X86_TUNE_SSE_SPLIT_REGS, "sse_split_regs", m_ATHLON_K8)
|
|
72
|
|
73 /* X86_TUNE_PARTIAL_FLAG_REG_STALL: this flag disables use of of flags
|
|
74 set by instructions affecting just some flags (in particular shifts).
|
|
75 This is because Core2 resolves dependencies on whole flags register
|
|
76 and such sequences introduce false dependency on previous instruction
|
|
77 setting full flags.
|
|
78
|
|
79 The flags does not affect generation of INC and DEC that is controlled
|
131
|
80 by X86_TUNE_USE_INCDEC. */
|
111
|
81
|
|
82 DEF_TUNE (X86_TUNE_PARTIAL_FLAG_REG_STALL, "partial_flag_reg_stall",
|
131
|
83 m_CORE2)
|
111
|
84
|
|
85 /* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid
|
|
86 partial dependencies. */
|
|
87 DEF_TUNE (X86_TUNE_MOVX, "movx",
|
131
|
88 m_PPRO | m_P4_NOCONA | m_CORE2 | m_NEHALEM | m_SANDYBRIDGE
|
|
89 | m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_KNL | m_KNM | m_INTEL
|
|
90 | m_GOLDMONT_PLUS | m_GEODE | m_AMD_MULTIPLE
|
|
91 | m_CORE_AVX2 | m_TREMONT | m_GENERIC)
|
111
|
92
|
|
93 /* X86_TUNE_MEMORY_MISMATCH_STALL: Avoid partial stores that are followed by
|
|
94 full sized loads. */
|
|
95 DEF_TUNE (X86_TUNE_MEMORY_MISMATCH_STALL, "memory_mismatch_stall",
|
|
96 m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_INTEL
|
131
|
97 | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_AMD_MULTIPLE
|
|
98 | m_TREMONT | m_GENERIC)
|
111
|
99
|
|
100 /* X86_TUNE_FUSE_CMP_AND_BRANCH_32: Fuse compare with a subsequent
|
131
|
101 conditional jump instruction for 32 bit TARGET. */
|
111
|
102 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_32, "fuse_cmp_and_branch_32",
|
131
|
103 m_CORE_ALL | m_BDVER | m_ZNVER1 | m_GENERIC)
|
111
|
104
|
|
105 /* X86_TUNE_FUSE_CMP_AND_BRANCH_64: Fuse compare with a subsequent
|
131
|
106 conditional jump instruction for TARGET_64BIT. */
|
111
|
107 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_64, "fuse_cmp_and_branch_64",
|
131
|
108 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_BDVER | m_ZNVER1 | m_GENERIC)
|
111
|
109
|
|
110 /* X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS: Fuse compare with a
|
|
111 subsequent conditional jump instruction when the condition jump
|
|
112 check sign flag (SF) or overflow flag (OF). */
|
|
113 DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH_SOFLAGS, "fuse_cmp_and_branch_soflags",
|
131
|
114 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_BDVER | m_ZNVER1 | m_GENERIC)
|
111
|
115
|
|
116 /* X86_TUNE_FUSE_ALU_AND_BRANCH: Fuse alu with a subsequent conditional
|
|
117 jump instruction when the alu instruction produces the CCFLAG consumed by
|
|
118 the conditional jump instruction. */
|
|
119 DEF_TUNE (X86_TUNE_FUSE_ALU_AND_BRANCH, "fuse_alu_and_branch",
|
131
|
120 m_SANDYBRIDGE | m_CORE_AVX2 | m_GENERIC)
|
111
|
121
|
|
122
|
|
123 /*****************************************************************************/
|
|
124 /* Function prologue, epilogue and function calling sequences. */
|
|
125 /*****************************************************************************/
|
|
126
|
|
127 /* X86_TUNE_ACCUMULATE_OUTGOING_ARGS: Allocate stack space for outgoing
|
|
128 arguments in prologue/epilogue instead of separately for each call
|
|
129 by push/pop instructions.
|
|
130 This increase code size by about 5% in 32bit mode, less so in 64bit mode
|
|
131 because parameters are passed in registers. It is considerable
|
|
132 win for targets without stack engine that prevents multple push operations
|
131
|
133 to happen in parallel. */
|
111
|
134
|
|
135 DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS, "accumulate_outgoing_args",
|
|
136 m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
|
131
|
137 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_ATHLON_K8)
|
111
|
138
|
|
139 /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in prologues that are
|
|
140 considered on critical path. */
|
|
141 DEF_TUNE (X86_TUNE_PROLOGUE_USING_MOVE, "prologue_using_move",
|
|
142 m_PPRO | m_ATHLON_K8)
|
|
143
|
|
144 /* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in epilogues that are
|
|
145 considered on critical path. */
|
|
146 DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE, "epilogue_using_move",
|
|
147 m_PPRO | m_ATHLON_K8)
|
|
148
|
|
149 /* X86_TUNE_USE_LEAVE: Use "leave" instruction in epilogues where it fits. */
|
|
150 DEF_TUNE (X86_TUNE_USE_LEAVE, "use_leave",
|
|
151 m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC)
|
|
152
|
|
153 /* X86_TUNE_PUSH_MEMORY: Enable generation of "push mem" instructions.
|
|
154 Some chips, like 486 and Pentium works faster with separate load
|
|
155 and push instructions. */
|
|
156 DEF_TUNE (X86_TUNE_PUSH_MEMORY, "push_memory",
|
|
157 m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE
|
|
158 | m_GENERIC)
|
|
159
|
|
160 /* X86_TUNE_SINGLE_PUSH: Enable if single push insn is preferred
|
|
161 over esp subtraction. */
|
|
162 DEF_TUNE (X86_TUNE_SINGLE_PUSH, "single_push", m_386 | m_486 | m_PENT
|
|
163 | m_LAKEMONT | m_K6_GEODE)
|
|
164
|
|
165 /* X86_TUNE_DOUBLE_PUSH. Enable if double push insn is preferred
|
|
166 over esp subtraction. */
|
|
167 DEF_TUNE (X86_TUNE_DOUBLE_PUSH, "double_push", m_PENT | m_LAKEMONT
|
|
168 | m_K6_GEODE)
|
|
169
|
|
170 /* X86_TUNE_SINGLE_POP: Enable if single pop insn is preferred
|
|
171 over esp addition. */
|
|
172 DEF_TUNE (X86_TUNE_SINGLE_POP, "single_pop", m_386 | m_486 | m_PENT
|
|
173 | m_LAKEMONT | m_PPRO)
|
|
174
|
|
175 /* X86_TUNE_DOUBLE_POP: Enable if double pop insn is preferred
|
|
176 over esp addition. */
|
|
177 DEF_TUNE (X86_TUNE_DOUBLE_POP, "double_pop", m_PENT | m_LAKEMONT)
|
|
178
|
|
179 /*****************************************************************************/
|
|
180 /* Branch predictor tuning */
|
|
181 /*****************************************************************************/
|
|
182
|
|
183 /* X86_TUNE_PAD_SHORT_FUNCTION: Make every function to be at least 4
|
|
184 instructions long. */
|
|
185 DEF_TUNE (X86_TUNE_PAD_SHORT_FUNCTION, "pad_short_function", m_BONNELL)
|
|
186
|
|
187 /* X86_TUNE_PAD_RETURNS: Place NOP before every RET that is a destination
|
|
188 of conditional jump or directly preceded by other jump instruction.
|
|
189 This is important for AND K8-AMDFAM10 because the branch prediction
|
|
190 architecture expect at most one jump per 2 byte window. Failing to
|
|
191 pad returns leads to misaligned return stack. */
|
|
192 DEF_TUNE (X86_TUNE_PAD_RETURNS, "pad_returns",
|
131
|
193 m_ATHLON_K8 | m_AMDFAM10)
|
111
|
194
|
|
195 /* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more
|
|
196 than 4 branch instructions in the 16 byte window. */
|
|
197 DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT, "four_jump_limit",
|
|
198 m_PPRO | m_P4_NOCONA | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM
|
131
|
199 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL | m_ATHLON_K8
|
|
200 | m_AMDFAM10)
|
111
|
201
|
|
202 /*****************************************************************************/
|
|
203 /* Integer instruction selection tuning */
|
|
204 /*****************************************************************************/
|
|
205
|
|
206 /* X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL: Enable software prefetching
|
|
207 at -O3. For the moment, the prefetching seems badly tuned for Intel
|
|
208 chips. */
|
|
209 DEF_TUNE (X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL, "software_prefetching_beneficial",
|
|
210 m_K6_GEODE | m_ATHLON_K8 | m_AMDFAM10 | m_BDVER | m_BTVER)
|
|
211
|
|
212 /* X86_TUNE_LCP_STALL: Avoid an expensive length-changing prefix stall
|
|
213 on 16-bit immediate moves into memory on Core2 and Corei7. */
|
|
214 DEF_TUNE (X86_TUNE_LCP_STALL, "lcp_stall", m_CORE_ALL | m_GENERIC)
|
|
215
|
|
216 /* X86_TUNE_READ_MODIFY: Enable use of read-modify instructions such
|
|
217 as "add mem, reg". */
|
|
218 DEF_TUNE (X86_TUNE_READ_MODIFY, "read_modify", ~(m_PENT | m_LAKEMONT | m_PPRO))
|
|
219
|
131
|
220 /* X86_TUNE_USE_INCDEC: Enable use of inc/dec instructions.
|
|
221
|
|
222 Core2 and nehalem has stall of 7 cycles for partial flag register stalls.
|
|
223 Sandy bridge and Ivy bridge generate extra uop. On Haswell this extra uop
|
|
224 is output only when the values needs to be really merged, which is not
|
|
225 done by GCC generated code. */
|
111
|
226 DEF_TUNE (X86_TUNE_USE_INCDEC, "use_incdec",
|
131
|
227 ~(m_P4_NOCONA | m_CORE2 | m_NEHALEM | m_SANDYBRIDGE
|
|
228 | m_BONNELL | m_SILVERMONT | m_INTEL | m_KNL | m_KNM | m_GOLDMONT
|
|
229 | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC))
|
111
|
230
|
|
231 /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred
|
|
232 for DFmode copies */
|
|
233 DEF_TUNE (X86_TUNE_INTEGER_DFMODE_MOVES, "integer_dfmode_moves",
|
|
234 ~(m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
|
131
|
235 | m_KNL | m_KNM | m_INTEL | m_GEODE | m_AMD_MULTIPLE | m_GOLDMONT
|
|
236 | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC))
|
111
|
237
|
|
238 /* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag
|
|
239 will impact LEA instruction selection. */
|
|
240 DEF_TUNE (X86_TUNE_OPT_AGU, "opt_agu", m_BONNELL | m_SILVERMONT | m_KNL
|
131
|
241 | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL)
|
111
|
242
|
|
243 /* X86_TUNE_AVOID_LEA_FOR_ADDR: Avoid lea for address computation. */
|
|
244 DEF_TUNE (X86_TUNE_AVOID_LEA_FOR_ADDR, "avoid_lea_for_addr",
|
131
|
245 m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT
|
|
246 | m_KNL | m_KNM)
|
111
|
247
|
|
248 /* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is
|
|
249 vector path on AMD machines.
|
|
250 FIXME: Do we need to enable this for core? */
|
|
251 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM32_MEM, "slow_imul_imm32_mem",
|
|
252 m_K8 | m_AMDFAM10)
|
|
253
|
|
254 /* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD
|
|
255 machines.
|
|
256 FIXME: Do we need to enable this for core? */
|
|
257 DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8, "slow_imul_imm8",
|
|
258 m_K8 | m_AMDFAM10)
|
|
259
|
|
260 /* X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE: Try to avoid memory operands for
|
|
261 a conditional move. */
|
|
262 DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE, "avoid_mem_opnd_for_cmove",
|
131
|
263 m_BONNELL | m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_KNL
|
|
264 | m_KNM | m_TREMONT | m_INTEL)
|
111
|
265
|
|
266 /* X86_TUNE_SINGLE_STRINGOP: Enable use of single string operations, such
|
|
267 as MOVS and STOS (without a REP prefix) to move/set sequences of bytes. */
|
|
268 DEF_TUNE (X86_TUNE_SINGLE_STRINGOP, "single_stringop", m_386 | m_P4_NOCONA)
|
|
269
|
|
270 /* X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES: Enable generation of
|
|
271 compact prologues and epilogues by issuing a misaligned moves. This
|
|
272 requires target to handle misaligned moves and partial memory stalls
|
|
273 reasonably well.
|
|
274 FIXME: This may actualy be a win on more targets than listed here. */
|
|
275 DEF_TUNE (X86_TUNE_MISALIGNED_MOVE_STRING_PRO_EPILOGUES,
|
|
276 "misaligned_move_string_pro_epilogues",
|
|
277 m_386 | m_486 | m_CORE_ALL | m_AMD_MULTIPLE | m_GENERIC)
|
|
278
|
|
279 /* X86_TUNE_USE_SAHF: Controls use of SAHF. */
|
|
280 DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf",
|
|
281 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
|
|
282 | m_KNL | m_KNM | m_INTEL | m_K6_GEODE | m_K8 | m_AMDFAM10 | m_BDVER
|
131
|
283 | m_BTVER | m_ZNVER1 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT
|
|
284 | m_GENERIC)
|
111
|
285
|
|
286 /* X86_TUNE_USE_CLTD: Controls use of CLTD and CTQO instructions. */
|
|
287 DEF_TUNE (X86_TUNE_USE_CLTD, "use_cltd",
|
|
288 ~(m_PENT | m_LAKEMONT | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
|
131
|
289 | m_K6 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT))
|
111
|
290
|
|
291 /* X86_TUNE_USE_BT: Enable use of BT (bit test) instructions. */
|
|
292 DEF_TUNE (X86_TUNE_USE_BT, "use_bt",
|
|
293 m_CORE_ALL | m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_INTEL
|
131
|
294 | m_LAKEMONT | m_AMD_MULTIPLE | m_GOLDMONT | m_GOLDMONT_PLUS
|
|
295 | m_TREMONT | m_GENERIC)
|
111
|
296
|
|
297 /* X86_TUNE_AVOID_FALSE_DEP_FOR_BMI: Avoid false dependency
|
|
298 for bit-manipulation instructions. */
|
|
299 DEF_TUNE (X86_TUNE_AVOID_FALSE_DEP_FOR_BMI, "avoid_false_dep_for_bmi",
|
131
|
300 m_SANDYBRIDGE | m_CORE_AVX2 | m_GENERIC)
|
111
|
301
|
|
302 /* X86_TUNE_ADJUST_UNROLL: This enables adjusting the unroll factor based
|
|
303 on hardware capabilities. Bdver3 hardware has a loop buffer which makes
|
|
304 unrolling small loop less important. For, such architectures we adjust
|
|
305 the unroll factor so that the unrolled loop fits the loop buffer. */
|
|
306 DEF_TUNE (X86_TUNE_ADJUST_UNROLL, "adjust_unroll_factor", m_BDVER3 | m_BDVER4)
|
|
307
|
|
308 /* X86_TUNE_ONE_IF_CONV_INSNS: Restrict a number of cmov insns in
|
|
309 if-converted sequence to one. */
|
|
310 DEF_TUNE (X86_TUNE_ONE_IF_CONV_INSN, "one_if_conv_insn",
|
131
|
311 m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_CORE_ALL | m_GOLDMONT
|
|
312 | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC)
|
111
|
313
|
|
314 /*****************************************************************************/
|
|
315 /* 387 instruction selection tuning */
|
|
316 /*****************************************************************************/
|
|
317
|
|
318 /* X86_TUNE_USE_HIMODE_FIOP: Enables use of x87 instructions with 16bit
|
|
319 integer operand.
|
|
320 FIXME: Why this is disabled for modern chips? */
|
|
321 DEF_TUNE (X86_TUNE_USE_HIMODE_FIOP, "use_himode_fiop",
|
|
322 m_386 | m_486 | m_K6_GEODE)
|
|
323
|
|
324 /* X86_TUNE_USE_SIMODE_FIOP: Enables use of x87 instructions with 32bit
|
|
325 integer operand. */
|
|
326 DEF_TUNE (X86_TUNE_USE_SIMODE_FIOP, "use_simode_fiop",
|
|
327 ~(m_PENT | m_LAKEMONT | m_PPRO | m_CORE_ALL | m_BONNELL
|
131
|
328 | m_SILVERMONT | m_KNL | m_KNM | m_INTEL | m_AMD_MULTIPLE
|
|
329 | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC))
|
111
|
330
|
|
331 /* X86_TUNE_USE_FFREEP: Use freep instruction instead of fstp. */
|
|
332 DEF_TUNE (X86_TUNE_USE_FFREEP, "use_ffreep", m_AMD_MULTIPLE)
|
|
333
|
|
334 /* X86_TUNE_EXT_80387_CONSTANTS: Use fancy 80387 constants, such as PI. */
|
|
335 DEF_TUNE (X86_TUNE_EXT_80387_CONSTANTS, "ext_80387_constants",
|
|
336 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BONNELL | m_SILVERMONT
|
131
|
337 | m_KNL | m_KNM | m_INTEL | m_K6_GEODE | m_ATHLON_K8 | m_GOLDMONT
|
|
338 | m_GOLDMONT_PLUS | m_TREMONT | m_GENERIC)
|
111
|
339
|
|
340 /*****************************************************************************/
|
|
341 /* SSE instruction selection tuning */
|
|
342 /*****************************************************************************/
|
|
343
|
|
344 /* X86_TUNE_GENERAL_REGS_SSE_SPILL: Try to spill general regs to SSE
|
|
345 regs instead of memory. */
|
|
346 DEF_TUNE (X86_TUNE_GENERAL_REGS_SSE_SPILL, "general_regs_sse_spill",
|
|
347 m_CORE_ALL)
|
|
348
|
|
349 /* X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL: Use movups for misaligned loads instead
|
|
350 of a sequence loading registers by parts. */
|
|
351 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL, "sse_unaligned_load_optimal",
|
131
|
352 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_SILVERMONT | m_KNL | m_KNM
|
|
353 | m_INTEL | m_GOLDMONT | m_GOLDMONT_PLUS
|
|
354 | m_TREMONT | m_AMDFAM10 | m_BDVER | m_BTVER | m_ZNVER1 | m_GENERIC)
|
111
|
355
|
|
356 /* X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL: Use movups for misaligned stores instead
|
|
357 of a sequence loading registers by parts. */
|
|
358 DEF_TUNE (X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL, "sse_unaligned_store_optimal",
|
131
|
359 m_NEHALEM | m_SANDYBRIDGE | m_CORE_AVX2 | m_SILVERMONT | m_KNL | m_KNM
|
|
360 | m_INTEL | m_GOLDMONT | m_GOLDMONT_PLUS
|
|
361 | m_TREMONT | m_BDVER | m_ZNVER1 | m_GENERIC)
|
111
|
362
|
|
363 /* Use packed single precision instructions where posisble. I.e. movups instead
|
|
364 of movupd. */
|
|
365 DEF_TUNE (X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL, "sse_packed_single_insn_optimal",
|
|
366 m_BDVER | m_ZNVER1)
|
|
367
|
|
368 /* X86_TUNE_SSE_TYPELESS_STORES: Always movaps/movups for 128bit stores. */
|
|
369 DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES, "sse_typeless_stores",
|
|
370 m_AMD_MULTIPLE | m_CORE_ALL | m_GENERIC)
|
|
371
|
|
372 /* X86_TUNE_SSE_LOAD0_BY_PXOR: Always use pxor to load0 as opposed to
|
|
373 xorps/xorpd and other variants. */
|
|
374 DEF_TUNE (X86_TUNE_SSE_LOAD0_BY_PXOR, "sse_load0_by_pxor",
|
|
375 m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BDVER | m_BTVER | m_ZNVER1
|
|
376 | m_GENERIC)
|
|
377
|
|
378 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from integer
|
|
379 to SSE registers. If disabled, the moves will be done by storing
|
|
380 the value to memory and reloading. */
|
|
381 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_TO_VEC, "inter_unit_moves_to_vec",
|
131
|
382 ~(m_ATHLON_K8 | m_AMDFAM10 | m_BDVER | m_BTVER | m_GENERIC))
|
111
|
383
|
|
384 /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from SSE
|
|
385 to integer registers. If disabled, the moves will be done by storing
|
|
386 the value to memory and reloading. */
|
|
387 DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_FROM_VEC, "inter_unit_moves_from_vec",
|
|
388 ~m_ATHLON_K8)
|
|
389
|
|
390 /* X86_TUNE_INTER_UNIT_CONVERSIONS: Enable float<->integer conversions
|
131
|
391 to use both SSE and integer registers at a same time. */
|
111
|
392 DEF_TUNE (X86_TUNE_INTER_UNIT_CONVERSIONS, "inter_unit_conversions",
|
|
393 ~(m_AMDFAM10 | m_BDVER))
|
|
394
|
|
395 /* X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS: Try to split memory operand for
|
|
396 fp converts to destination register. */
|
|
397 DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS, "split_mem_opnd_for_fp_converts",
|
131
|
398 m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT | m_GOLDMONT_PLUS
|
|
399 | m_TREMONT | m_INTEL)
|
111
|
400
|
|
401 /* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion
|
|
402 from FP to FP. This form of instructions avoids partial write to the
|
|
403 destination. */
|
|
404 DEF_TUNE (X86_TUNE_USE_VECTOR_FP_CONVERTS, "use_vector_fp_converts",
|
|
405 m_AMDFAM10)
|
|
406
|
|
407 /* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion
|
|
408 from integer to FP. */
|
|
409 DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS, "use_vector_converts", m_AMDFAM10)
|
|
410
|
|
411 /* X86_TUNE_SLOW_SHUFB: Indicates tunings with slow pshufb instruction. */
|
|
412 DEF_TUNE (X86_TUNE_SLOW_PSHUFB, "slow_pshufb",
|
131
|
413 m_BONNELL | m_SILVERMONT | m_KNL | m_KNM | m_GOLDMONT
|
|
414 | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL)
|
111
|
415
|
|
416 /* X86_TUNE_AVOID_4BYTE_PREFIXES: Avoid instructions requiring 4+ bytes of prefixes. */
|
|
417 DEF_TUNE (X86_TUNE_AVOID_4BYTE_PREFIXES, "avoid_4byte_prefixes",
|
131
|
418 m_SILVERMONT | m_GOLDMONT | m_GOLDMONT_PLUS | m_TREMONT | m_INTEL)
|
|
419
|
|
420 /* X86_TUNE_USE_GATHER: Use gather instructions. */
|
|
421 DEF_TUNE (X86_TUNE_USE_GATHER, "use_gather",
|
|
422 ~(m_ZNVER1 | m_GENERIC))
|
|
423
|
|
424 /* X86_TUNE_AVOID_128FMA_CHAINS: Avoid creating loops with tight 128bit or
|
|
425 smaller FMA chain. */
|
|
426 DEF_TUNE (X86_TUNE_AVOID_128FMA_CHAINS, "avoid_fma_chains", m_ZNVER1)
|
111
|
427
|
|
428 /*****************************************************************************/
|
|
429 /* AVX instruction selection tuning (some of SSE flags affects AVX, too) */
|
|
430 /*****************************************************************************/
|
|
431
|
|
432 /* X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL: if false, unaligned loads are
|
|
433 split. */
|
|
434 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL, "256_unaligned_load_optimal",
|
|
435 ~(m_NEHALEM | m_SANDYBRIDGE | m_GENERIC))
|
|
436
|
|
437 /* X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL: if false, unaligned stores are
|
|
438 split. */
|
|
439 DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL, "256_unaligned_store_optimal",
|
|
440 ~(m_NEHALEM | m_SANDYBRIDGE | m_BDVER | m_ZNVER1 | m_GENERIC))
|
|
441
|
|
442 /* X86_TUNE_AVX128_OPTIMAL: Enable 128-bit AVX instruction generation for
|
|
443 the auto-vectorizer. */
|
|
444 DEF_TUNE (X86_TUNE_AVX128_OPTIMAL, "avx128_optimal", m_BDVER | m_BTVER2
|
|
445 | m_ZNVER1)
|
|
446
|
131
|
447 /* X86_TUNE_AVX256_OPTIMAL: Use 256-bit AVX instructions instead of 512-bit AVX
|
|
448 instructions in the auto-vectorizer. */
|
|
449 DEF_TUNE (X86_TUNE_AVX256_OPTIMAL, "avx256_optimal", m_CORE_AVX512)
|
|
450
|
111
|
451 /*****************************************************************************/
|
|
452 /* Historical relics: tuning flags that helps a specific old CPU designs */
|
|
453 /*****************************************************************************/
|
|
454
|
|
455 /* X86_TUNE_DOUBLE_WITH_ADD: Use add instead of sal to double value in
|
|
456 an integer register. */
|
|
457 DEF_TUNE (X86_TUNE_DOUBLE_WITH_ADD, "double_with_add", ~m_386)
|
|
458
|
|
459 /* X86_TUNE_ALWAYS_FANCY_MATH_387: controls use of fancy 387 operations,
|
|
460 such as fsqrt, fprem, fsin, fcos, fsincos etc.
|
|
461 Should be enabled for all targets that always has coprocesor. */
|
|
462 DEF_TUNE (X86_TUNE_ALWAYS_FANCY_MATH_387, "always_fancy_math_387",
|
|
463 ~(m_386 | m_486 | m_LAKEMONT))
|
|
464
|
|
465 /* X86_TUNE_UNROLL_STRLEN: Produce (quite lame) unrolled sequence for
|
|
466 inline strlen. This affects only -minline-all-stringops mode. By
|
|
467 default we always dispatch to a library since our internal strlen
|
|
468 is bad. */
|
|
469 DEF_TUNE (X86_TUNE_UNROLL_STRLEN, "unroll_strlen", ~m_386)
|
|
470
|
|
471 /* X86_TUNE_SHIFT1: Enables use of short encoding of "sal reg" instead of
|
|
472 longer "sal $1, reg". */
|
|
473 DEF_TUNE (X86_TUNE_SHIFT1, "shift1", ~m_486)
|
|
474
|
|
475 /* X86_TUNE_ZERO_EXTEND_WITH_AND: Use AND instruction instead
|
|
476 of mozbl/movwl. */
|
|
477 DEF_TUNE (X86_TUNE_ZERO_EXTEND_WITH_AND, "zero_extend_with_and",
|
|
478 m_486 | m_PENT)
|
|
479
|
|
480 /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode
|
|
481 and SImode multiply, but 386 and 486 do HImode multiply faster. */
|
|
482 DEF_TUNE (X86_TUNE_PROMOTE_HIMODE_IMUL, "promote_himode_imul",
|
|
483 ~(m_386 | m_486))
|
|
484
|
|
485 /* X86_TUNE_FAST_PREFIX: Enable demoting some 32bit or 64bit arithmetic
|
|
486 into 16bit/8bit when resulting sequence is shorter. For example
|
|
487 for "and $-65536, reg" to 16bit store of 0. */
|
|
488 DEF_TUNE (X86_TUNE_FAST_PREFIX, "fast_prefix",
|
|
489 ~(m_386 | m_486 | m_PENT | m_LAKEMONT))
|
|
490
|
|
491 /* X86_TUNE_READ_MODIFY_WRITE: Enable use of read modify write instructions
|
|
492 such as "add $1, mem". */
|
|
493 DEF_TUNE (X86_TUNE_READ_MODIFY_WRITE, "read_modify_write",
|
|
494 ~(m_PENT | m_LAKEMONT))
|
|
495
|
|
496 /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR
|
|
497 than a MOV. */
|
|
498 DEF_TUNE (X86_TUNE_MOVE_M1_VIA_OR, "move_m1_via_or", m_PENT | m_LAKEMONT)
|
|
499
|
|
500 /* X86_TUNE_NOT_UNPAIRABLE: NOT is not pairable on Pentium, while XOR is,
|
|
501 but one byte longer. */
|
|
502 DEF_TUNE (X86_TUNE_NOT_UNPAIRABLE, "not_unpairable", m_PENT | m_LAKEMONT)
|
|
503
|
|
504 /* X86_TUNE_PARTIAL_REG_STALL: Pentium pro, unlike later chips, handled
|
|
505 use of partial registers by renaming. This improved performance of 16bit
|
|
506 code where upper halves of registers are not used. It also leads to
|
|
507 an penalty whenever a 16bit store is followed by 32bit use. This flag
|
|
508 disables production of such sequences in common cases.
|
|
509 See also X86_TUNE_HIMODE_MATH.
|
|
510
|
|
511 In current implementation the partial register stalls are not eliminated
|
|
512 very well - they can be introduced via subregs synthesized by combine
|
|
513 and can happen in caller/callee saving sequences. */
|
|
514 DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL, "partial_reg_stall", m_PPRO)
|
|
515
|
|
516 /* X86_TUNE_PROMOTE_QIMODE: When it is cheap, turn 8bit arithmetic to
|
|
517 corresponding 32bit arithmetic. */
|
|
518 DEF_TUNE (X86_TUNE_PROMOTE_QIMODE, "promote_qimode",
|
|
519 ~m_PPRO)
|
|
520
|
|
521 /* X86_TUNE_PROMOTE_HI_REGS: Same, but for 16bit artihmetic. Again we avoid
|
|
522 partial register stalls on PentiumPro targets. */
|
|
523 DEF_TUNE (X86_TUNE_PROMOTE_HI_REGS, "promote_hi_regs", m_PPRO)
|
|
524
|
|
525 /* X86_TUNE_HIMODE_MATH: Enable use of 16bit arithmetic.
|
|
526 On PPro this flag is meant to avoid partial register stalls. */
|
|
527 DEF_TUNE (X86_TUNE_HIMODE_MATH, "himode_math", ~m_PPRO)
|
|
528
|
|
529 /* X86_TUNE_SPLIT_LONG_MOVES: Avoid instructions moving immediates
|
|
530 directly to memory. */
|
|
531 DEF_TUNE (X86_TUNE_SPLIT_LONG_MOVES, "split_long_moves", m_PPRO)
|
|
532
|
|
533 /* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */
|
|
534 DEF_TUNE (X86_TUNE_USE_XCHGB, "use_xchgb", m_PENT4)
|
|
535
|
|
536 /* X86_TUNE_USE_MOV0: Use "mov $0, reg" instead of "xor reg, reg" to clear
|
|
537 integer register. */
|
|
538 DEF_TUNE (X86_TUNE_USE_MOV0, "use_mov0", m_K6)
|
|
539
|
|
540 /* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory
|
|
541 operand that cannot be represented using a modRM byte. The XOR
|
|
542 replacement is long decoded, so this split helps here as well. */
|
|
543 DEF_TUNE (X86_TUNE_NOT_VECTORMODE, "not_vectormode", m_K6)
|
|
544
|
|
545 /* X86_TUNE_AVOID_VECTOR_DECODE: Enable splitters that avoid vector decoded
|
|
546 forms of instructions on K8 targets. */
|
|
547 DEF_TUNE (X86_TUNE_AVOID_VECTOR_DECODE, "avoid_vector_decode",
|
|
548 m_K8)
|
|
549
|
|
550 /*****************************************************************************/
|
|
551 /* This never worked well before. */
|
|
552 /*****************************************************************************/
|
|
553
|
|
554 /* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based
|
|
555 on simulation result. But after P4 was made, no performance benefit
|
|
556 was observed with branch hints. It also increases the code size.
|
|
557 As a result, icc never generates branch hints. */
|
|
558 DEF_TUNE (X86_TUNE_BRANCH_PREDICTION_HINTS, "branch_prediction_hints", 0U)
|
|
559
|
|
560 /* X86_TUNE_QIMODE_MATH: Enable use of 8bit arithmetic. */
|
|
561 DEF_TUNE (X86_TUNE_QIMODE_MATH, "qimode_math", ~0U)
|
|
562
|
|
563 /* X86_TUNE_PROMOTE_QI_REGS: This enables generic code that promotes all 8bit
|
|
564 arithmetic to 32bit via PROMOTE_MODE macro. This code generation scheme
|
|
565 is usually used for RISC targets. */
|
|
566 DEF_TUNE (X86_TUNE_PROMOTE_QI_REGS, "promote_qi_regs", 0U)
|
131
|
567
|
|
568 /* X86_TUNE_EMIT_VZEROUPPER: This enables vzeroupper instruction insertion
|
|
569 before a transfer of control flow out of the function. */
|
|
570 DEF_TUNE (X86_TUNE_EMIT_VZEROUPPER, "emit_vzeroupper", ~m_KNL)
|