From 616cbae76230eae695946595f68a8eb6b64f82ff Mon Sep 17 00:00:00 2001 From: hubicka Date: Mon, 21 Oct 2013 11:02:08 +0000 Subject: [PATCH] * config/i386/i386-tune.def: Add comment; organize into categories git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@203888 138bc75d-0d04-0410-961f-82ee72b054a4 --- gcc/ChangeLog | 4 + gcc/config/i386/x86-tune.def | 609 +++++++++++++++++++++++-------------------- 2 files changed, 333 insertions(+), 280 deletions(-) diff --git a/gcc/ChangeLog b/gcc/ChangeLog index a8cbb5a..bf54da9 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,3 +1,7 @@ +2013-10-20 Jan Hubicka + + * config/i386/i386-tune.def: Add comment; organize into categories + 2013-10-21 Michael Zolotukhin * config/i386/i386.c (expand_set_or_movmem_via_loop): Add issetmem diff --git a/gcc/config/i386/x86-tune.def b/gcc/config/i386/x86-tune.def index 4d238af..a6e70f5 100644 --- a/gcc/config/i386/x86-tune.def +++ b/gcc/config/i386/x86-tune.def @@ -18,61 +18,54 @@ a copy of the GCC Runtime Library Exception along with this program; see the files COPYING3 and COPYING.RUNTIME respectively. If not, see . */ -/* X86_TUNE_USE_LEAVE: Use "leave" instruction in epilogues where it fits. */ -DEF_TUNE (X86_TUNE_USE_LEAVE, "use_leave", - m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC) - -/* X86_TUNE_PUSH_MEMORY: Enable generation of "push mem" instructions. - Some chips, like 486 and Pentium works faster with separate load - and push instructions. */ -DEF_TUNE (X86_TUNE_PUSH_MEMORY, "push_memory", - m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE - | m_GENERIC) - -/* X86_TUNE_ZERO_EXTEND_WITH_AND: Use AND instruction instead - of mozbl/movwl. */ -DEF_TUNE (X86_TUNE_ZERO_EXTEND_WITH_AND, "zero_extend_with_and", m_486 | m_PENT) - -/* X86_TUNE_UNROLL_STRLEN: Produce (quite lame) unrolled sequence for - inline strlen. This affects only -minline-all-stringops mode. By - default we always dispatch to a library since our internal strlen - is bad. */ -DEF_TUNE (X86_TUNE_UNROLL_STRLEN, "unroll_strlen", - m_486 | m_PENT | m_PPRO | m_ATOM | m_SLM | m_CORE_ALL | m_K6 - | m_AMD_MULTIPLE | m_GENERIC) +/* Tuning for a given CPU XXXX consists of: + - adding new CPU into: + - adding PROCESSOR_XXX to processor_type (in i386.h) + - possibly adding XXX into CPU attribute in i386.md + - adding XXX to processor_alias_table (in i386.c) + - introducing ix86_XXX_cost in i386.c + - Stringop generation table can be build based on test_stringop + - script (once rest of tuning is complete) + - designing a scheduler model in + - XXXX.md file + - Updating ix86_issue_rate and ix86_adjust_cost in i386.md + - possibly updating ia32_multipass_dfa_lookahead, ix86_sched_reorder + and ix86_sched_init_global if those tricks are needed. + - Tunning the flags bellow. Those are split into sections and each + section is very roughly ordered by importance. */ + +/*****************************************************************************/ +/* Scheduling flags. */ +/*****************************************************************************/ -/* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based - on simulation result. But after P4 was made, no performance benefit - was observed with branch hints. It also increases the code size. - As a result, icc never generates branch hints. */ -DEF_TUNE (X86_TUNE_BRANCH_PREDICTION_HINTS, "branch_prediction_hints", 0) - -/* X86_TUNE_DOUBLE_WITH_ADD: Use add instead of sal to double value in - an integer register. */ -DEF_TUNE (X86_TUNE_DOUBLE_WITH_ADD, "double_with_add", ~m_386) - -/* X86_TUNE_USE_SAHF: Controls use of SAHF. */ -DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf", - m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_K6_GEODE - | m_K8 | m_AMDFAM10 | m_BDVER | m_BTVER | m_GENERIC) +/* X86_TUNE_SCHEDULE: Enable scheduling. */ +DEF_TUNE (X86_TUNE_SCHEDULE, "schedule", + m_PENT | m_PPRO | m_CORE_ALL | m_ATOM | m_SLM | m_K6_GEODE + | m_AMD_MULTIPLE | m_GENERIC) -/* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid - partial dependencies. */ -DEF_TUNE (X86_TUNE_MOVX, "movx", - m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_GEODE - | m_AMD_MULTIPLE | m_GENERIC) +/* X86_TUNE_PARTIAL_REG_DEPENDENCY: Enable more register renaming + on modern chips. Preffer stores affecting whole integer register + over partial stores. For example preffer MOVZBL or MOVQ to load 8bit + value over movb. */ +DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY, "partial_reg_dependency", + m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_AMD_MULTIPLE + | m_GENERIC) -/* X86_TUNE_PARTIAL_REG_STALL: Pentium pro, unlike later chips, handled - use of partial registers by renaming. This improved performance of 16bit - code where upper halves of registers are not used. It also leads to - an penalty whenever a 16bit store is followed by 32bit use. This flag - disables production of such sequences in common cases. - See also X86_TUNE_HIMODE_MATH. +/* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: This knob promotes all store + destinations to be 128bit to allow register renaming on 128bit SSE units, + but usually results in one extra microop on 64bit SSE units. + Experimental results shows that disabling this option on P4 brings over 20% + SPECfp regression, while enabling it on K8 brings roughly 2.4% regression + that can be partly masked by careful scheduling of moves. */ +DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY, "sse_partial_reg_dependency", + m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_AMDFAM10 + | m_BDVER | m_GENERIC) - In current implementation the partial register stalls are not eliminated - very well - they can be introduced via subregs synthesized by combine - and can happen in caller/callee saving sequences. */ -DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL, "partial_reg_stall", m_PPRO) +/* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies + are resolved on SSE register parts instead of whole registers, so we may + maintain just lower part of scalar values in proper format leaving the + upper part undefined. */ +DEF_TUNE (X86_TUNE_SSE_SPLIT_REGS, "sse_split_regs", m_ATHLON_K8) /* X86_TUNE_PARTIAL_FLAG_REG_STALL: this flag disables use of of flags set by instructions affecting just some flags (in particular shifts). @@ -88,74 +81,81 @@ DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL, "partial_reg_stall", m_PPRO) DEF_TUNE (X86_TUNE_PARTIAL_FLAG_REG_STALL, "partial_flag_reg_stall", m_CORE2 | m_GENERIC) -/* X86_TUNE_LCP_STALL: Avoid an expensive length-changing prefix stall - on 16-bit immediate moves into memory on Core2 and Corei7. */ -DEF_TUNE (X86_TUNE_LCP_STALL, "lcp_stall", m_CORE_ALL | m_GENERIC) - -/* X86_TUNE_USE_HIMODE_FIOP: Enables use of x87 instructions with 16bit - integer operand. - FIXME: Why this is disabled for modern chips? */ -DEF_TUNE (X86_TUNE_USE_HIMODE_FIOP, "use_himode_fiop", - m_386 | m_486 | m_K6_GEODE) - -/* X86_TUNE_USE_SIMODE_FIOP: Enables use of x87 instructions with 32bit - integer operand. */ -DEF_TUNE (X86_TUNE_USE_SIMODE_FIOP, "use_simode_fiop", - ~(m_PENT | m_PPRO | m_CORE_ALL | m_ATOM - | m_SLM | m_AMD_MULTIPLE | m_GENERIC)) +/* X86_TUNE_MOVX: Enable to zero extend integer registers to avoid + partial dependencies. */ +DEF_TUNE (X86_TUNE_MOVX, "movx", + m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_GEODE + | m_AMD_MULTIPLE | m_GENERIC) -/* X86_TUNE_USE_MOV0: Use "mov $0, reg" instead of "xor reg, reg" to clear - integer register. */ -DEF_TUNE (X86_TUNE_USE_MOV0, "use_mov0", m_K6) +/* X86_TUNE_MEMORY_MISMATCH_STALL: Avoid partial stores that are followed by + full sized loads. */ +DEF_TUNE (X86_TUNE_MEMORY_MISMATCH_STALL, "memory_mismatch_stall", + m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_AMD_MULTIPLE | m_GENERIC) -/* X86_TUNE_USE_CLTD: Controls use of CLTD and CTQO instructions. */ -DEF_TUNE (X86_TUNE_USE_CLTD, "use_cltd", ~(m_PENT | m_ATOM | m_SLM | m_K6)) +/* X86_TUNE_FUSE_CMP_AND_BRANCH: Fuse a compare or test instruction + with a subsequent conditional jump instruction into a single + compare-and-branch uop. + FIXME: revisit for generic. */ +DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH, "fuse_cmp_and_branch", m_BDVER | m_CORE_ALL) -/* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */ -DEF_TUNE (X86_TUNE_USE_XCHGB, "use_xchgb", m_PENT4) +/* X86_TUNE_REASSOC_INT_TO_PARALLEL: Try to produce parallel computations + during reassociation of integer computation. */ +DEF_TUNE (X86_TUNE_REASSOC_INT_TO_PARALLEL, "reassoc_int_to_parallel", + m_ATOM) -/* X86_TUNE_SPLIT_LONG_MOVES: Avoid instructions moving immediates - directly to memory. */ -DEF_TUNE (X86_TUNE_SPLIT_LONG_MOVES, "split_long_moves", m_PPRO) +/* X86_TUNE_REASSOC_FP_TO_PARALLEL: Try to produce parallel computations + during reassociation of fp computation. */ +DEF_TUNE (X86_TUNE_REASSOC_FP_TO_PARALLEL, "reassoc_fp_to_parallel", + m_ATOM | m_SLM | m_HASWELL | m_BDVER1 | m_BDVER2 | m_GENERIC) -/* X86_TUNE_READ_MODIFY_WRITE: Enable use of read modify write instructions - such as "add $1, mem". */ -DEF_TUNE (X86_TUNE_READ_MODIFY_WRITE, "read_modify_write", ~m_PENT) +/*****************************************************************************/ +/* Function prologue, epilogue and function calling sequences. */ +/*****************************************************************************/ -/* X86_TUNE_READ_MODIFY: Enable use of read-modify instructions such - as "add mem, reg". */ -DEF_TUNE (X86_TUNE_READ_MODIFY, "read_modify", ~(m_PENT | m_PPRO)) +/* X86_TUNE_ACCUMULATE_OUTGOING_ARGS: Allocate stack space for outgoing + arguments in prologue/epilogue instead of separately for each call + by push/pop instructions. + This increase code size by about 5% in 32bit mode, less so in 64bit mode + because parameters are passed in registers. It is considerable + win for targets without stack engine that prevents multple push operations + to happen in parallel. -/* X86_TUNE_PROMOTE_QIMODE: When it is cheap, turn 8bit arithmetic to - corresponding 32bit arithmetic. */ -DEF_TUNE (X86_TUNE_PROMOTE_QIMODE, "promote_qimode", - m_386 | m_486 | m_PENT | m_CORE_ALL | m_ATOM | m_SLM - | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC) + FIXME: the flags is incorrectly enabled for amdfam10, Bulldozer, + Bobcat and Generic. This is because disabling it causes large + regression on mgrid due to IRA limitation leading to unecessary + use of the frame pointer in 32bit mode. */ +DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS, "accumulate_outgoing_args", + m_PPRO | m_P4_NOCONA | m_ATOM | m_SLM | m_AMD_MULTIPLE | m_GENERIC) -/* X86_TUNE_FAST_PREFIX: Enable demoting some 32bit or 64bit arithmetic - into 16bit/8bit when resulting sequence is shorter. For example - for "and $-65536, reg" to 16bit store of 0. */ -DEF_TUNE (X86_TUNE_FAST_PREFIX, "fast_prefix", ~(m_386 | m_486 | m_PENT)) +/* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in prologues that are + considered on critical path. */ +DEF_TUNE (X86_TUNE_PROLOGUE_USING_MOVE, "prologue_using_move", + m_PPRO | m_ATHLON_K8) -/* X86_TUNE_SINGLE_STRINGOP: Enable use of single string operations, such - as MOVS and STOS (without a REP prefix) to move/set sequences of bytes. */ -DEF_TUNE (X86_TUNE_SINGLE_STRINGOP, "single_stringop", m_386 | m_P4_NOCONA) +/* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in epilogues that are + considered on critical path. */ +DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE, "epilogue_using_move", + m_PPRO | m_ATHLON_K8) -/* X86_TUNE_QIMODE_MATH: Enable use of 8bit arithmetic. */ -DEF_TUNE (X86_TUNE_QIMODE_MATH, "qimode_math", ~0) +/* X86_TUNE_USE_LEAVE: Use "leave" instruction in epilogues where it fits. */ +DEF_TUNE (X86_TUNE_USE_LEAVE, "use_leave", + m_386 | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE | m_GENERIC) -/* X86_TUNE_HIMODE_MATH: Enable use of 16bit arithmetic. - On PPro this flag is meant to avoid partial register stalls. */ -DEF_TUNE (X86_TUNE_HIMODE_MATH, "himode_math", ~m_PPRO) +/* X86_TUNE_PUSH_MEMORY: Enable generation of "push mem" instructions. + Some chips, like 486 and Pentium works faster with separate load + and push instructions. */ +DEF_TUNE (X86_TUNE_PUSH_MEMORY, "push_memory", + m_386 | m_P4_NOCONA | m_CORE_ALL | m_K6_GEODE | m_AMD_MULTIPLE + | m_GENERIC) -/* X86_TUNE_PROMOTE_QI_REGS: This enables generic code that promotes all 8bit - arithmetic to 32bit via PROMOTE_MODE macro. This code generation scheme - is usually used for RISC targets. */ -DEF_TUNE (X86_TUNE_PROMOTE_QI_REGS, "promote_qi_regs", 0) +/* X86_TUNE_SINGLE_PUSH: Enable if single push insn is preferred + over esp subtraction. */ +DEF_TUNE (X86_TUNE_SINGLE_PUSH, "single_push", m_386 | m_486 | m_PENT + | m_K6_GEODE) -/* X86_TUNE_PROMOTE_HI_REGS: Same, but for 16bit artihmetic. Again we avoid - partial register stalls on PentiumPro targets. */ -DEF_TUNE (X86_TUNE_PROMOTE_HI_REGS, "promote_hi_regs", m_PPRO) +/* X86_TUNE_DOUBLE_PUSH. Enable if double push insn is preferred + over esp subtraction. */ +DEF_TUNE (X86_TUNE_DOUBLE_PUSH, "double_push", m_PENT | m_K6_GEODE) /* X86_TUNE_SINGLE_POP: Enable if single pop insn is preferred over esp addition. */ @@ -165,14 +165,48 @@ DEF_TUNE (X86_TUNE_SINGLE_POP, "single_pop", m_386 | m_486 | m_PENT | m_PPRO) over esp addition. */ DEF_TUNE (X86_TUNE_DOUBLE_POP, "double_pop", m_PENT) -/* X86_TUNE_SINGLE_PUSH: Enable if single push insn is preferred - over esp subtraction. */ -DEF_TUNE (X86_TUNE_SINGLE_PUSH, "single_push", m_386 | m_486 | m_PENT - | m_K6_GEODE) +/*****************************************************************************/ +/* Branch predictor tuning */ +/*****************************************************************************/ -/* X86_TUNE_DOUBLE_PUSH. Enable if double push insn is preferred - over esp subtraction. */ -DEF_TUNE (X86_TUNE_DOUBLE_PUSH, "double_push", m_PENT | m_K6_GEODE) +/* X86_TUNE_PAD_SHORT_FUNCTION: Make every function to be at least 4 + instructions long. */ +DEF_TUNE (X86_TUNE_PAD_SHORT_FUNCTION, "pad_short_function", m_ATOM) + +/* X86_TUNE_PAD_RETURNS: Place NOP before every RET that is a destination + of conditional jump or directly preceded by other jump instruction. + This is important for AND K8-AMDFAM10 because the branch prediction + architecture expect at most one jump per 2 byte window. Failing to + pad returns leads to misaligned return stack. */ +DEF_TUNE (X86_TUNE_PAD_RETURNS, "pad_returns", + m_ATHLON_K8 | m_AMDFAM10 | m_GENERIC) + +/* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more + than 4 branch instructions in the 16 byte window. */ +DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT, "four_jump_limit", + m_PPRO | m_P4_NOCONA | m_ATOM | m_SLM | m_ATHLON_K8 | m_AMDFAM10) + +/*****************************************************************************/ +/* Integer instruction selection tuning */ +/*****************************************************************************/ + +/* X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL: Enable software prefetching + at -O3. For the moment, the prefetching seems badly tuned for Intel + chips. */ +DEF_TUNE (X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL, "software_prefetching_beneficial", + m_K6_GEODE | m_AMD_MULTIPLE) + +/* X86_TUNE_LCP_STALL: Avoid an expensive length-changing prefix stall + on 16-bit immediate moves into memory on Core2 and Corei7. */ +DEF_TUNE (X86_TUNE_LCP_STALL, "lcp_stall", m_CORE_ALL | m_GENERIC) + +/* X86_TUNE_READ_MODIFY: Enable use of read-modify instructions such + as "add mem, reg". */ +DEF_TUNE (X86_TUNE_READ_MODIFY, "read_modify", ~(m_PENT | m_PPRO)) + +/* X86_TUNE_USE_INCDEC: Enable use of inc/dec instructions. */ +DEF_TUNE (X86_TUNE_USE_INCDEC, "use_incdec", + ~(m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_GENERIC)) /* X86_TUNE_INTEGER_DFMODE_MOVES: Enable if integer moves are preferred for DFmode copies */ @@ -180,23 +214,79 @@ DEF_TUNE (X86_TUNE_INTEGER_DFMODE_MOVES, "integer_dfmode_moves", ~(m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_GEODE | m_AMD_MULTIPLE | m_GENERIC)) -/* X86_TUNE_PARTIAL_REG_DEPENDENCY: Enable more register renaming - on modern chips. Preffer stores affecting whole integer register - over partial stores. For example preffer MOVZBL or MOVQ to load 8bit - value over movb. */ -DEF_TUNE (X86_TUNE_PARTIAL_REG_DEPENDENCY, "partial_reg_dependency", - m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_AMD_MULTIPLE - | m_GENERIC) +/* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag + will impact LEA instruction selection. */ +DEF_TUNE (X86_TUNE_OPT_AGU, "opt_agu", m_ATOM | m_SLM) -/* X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY: This knob promotes all store - destinations to be 128bit to allow register renaming on 128bit SSE units, - but usually results in one extra microop on 64bit SSE units. - Experimental results shows that disabling this option on P4 brings over 20% - SPECfp regression, while enabling it on K8 brings roughly 2.4% regression - that can be partly masked by careful scheduling of moves. */ -DEF_TUNE (X86_TUNE_SSE_PARTIAL_REG_DEPENDENCY, "sse_partial_reg_dependency", - m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_AMDFAM10 - | m_BDVER | m_GENERIC) +/* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is + vector path on AMD machines. + FIXME: Do we need to enable this for core? */ +DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM32_MEM, "slow_imul_imm32_mem", + m_K8 | m_AMDFAM10) + +/* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD + machines. + FIXME: Do we need to enable this for core? */ +DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8, "slow_imul_imm8", + m_K8 | m_AMDFAM10) + +/* X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE: Try to avoid memory operands for + a conditional move. */ +DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE, "avoid_mem_opnd_for_cmove", + m_ATOM | m_SLM) + +/* X86_TUNE_SINGLE_STRINGOP: Enable use of single string operations, such + as MOVS and STOS (without a REP prefix) to move/set sequences of bytes. */ +DEF_TUNE (X86_TUNE_SINGLE_STRINGOP, "single_stringop", m_386 | m_P4_NOCONA) + +/* X86_TUNE_USE_SAHF: Controls use of SAHF. */ +DEF_TUNE (X86_TUNE_USE_SAHF, "use_sahf", + m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_K6_GEODE + | m_K8 | m_AMDFAM10 | m_BDVER | m_BTVER | m_GENERIC) + +/* X86_TUNE_USE_CLTD: Controls use of CLTD and CTQO instructions. */ +DEF_TUNE (X86_TUNE_USE_CLTD, "use_cltd", ~(m_PENT | m_ATOM | m_SLM | m_K6)) + +/* X86_TUNE_USE_BT: Enable use of BT (bit test) instructions. */ +DEF_TUNE (X86_TUNE_USE_BT, "use_bt", + m_CORE_ALL | m_ATOM | m_SLM | m_AMD_MULTIPLE | m_GENERIC) + +/*****************************************************************************/ +/* 387 instruction selection tuning */ +/*****************************************************************************/ + +/* X86_TUNE_USE_HIMODE_FIOP: Enables use of x87 instructions with 16bit + integer operand. + FIXME: Why this is disabled for modern chips? */ +DEF_TUNE (X86_TUNE_USE_HIMODE_FIOP, "use_himode_fiop", + m_386 | m_486 | m_K6_GEODE) + +/* X86_TUNE_USE_SIMODE_FIOP: Enables use of x87 instructions with 32bit + integer operand. */ +DEF_TUNE (X86_TUNE_USE_SIMODE_FIOP, "use_simode_fiop", + ~(m_PENT | m_PPRO | m_CORE_ALL | m_ATOM + | m_SLM | m_AMD_MULTIPLE | m_GENERIC)) + +/* X86_TUNE_USE_FFREEP: Use freep instruction instead of fstp. */ +DEF_TUNE (X86_TUNE_USE_FFREEP, "use_ffreep", m_AMD_MULTIPLE) + +/* X86_TUNE_EXT_80387_CONSTANTS: Use fancy 80387 constants, such as PI. */ +DEF_TUNE (X86_TUNE_EXT_80387_CONSTANTS, "ext_80387_constants", + m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_K6_GEODE + | m_ATHLON_K8 | m_GENERIC) + +/*****************************************************************************/ +/* SSE instruction selection tuning */ +/*****************************************************************************/ + +/* X86_TUNE_VECTORIZE_DOUBLE: Enable double precision vector + instructions. */ +DEF_TUNE (X86_TUNE_VECTORIZE_DOUBLE, "vectorize_double", ~m_ATOM) + +/* X86_TUNE_GENERAL_REGS_SSE_SPILL: Try to spill general regs to SSE + regs instead of memory. */ +DEF_TUNE (X86_TUNE_GENERAL_REGS_SSE_SPILL, "general_regs_sse_spill", + m_CORE_ALL) /* X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL: Use movups for misaligned loads instead of a sequence loading registers by parts. */ @@ -208,27 +298,11 @@ DEF_TUNE (X86_TUNE_SSE_UNALIGNED_LOAD_OPTIMAL, "sse_unaligned_load_optimal", DEF_TUNE (X86_TUNE_SSE_UNALIGNED_STORE_OPTIMAL, "sse_unaligned_store_optimal", m_COREI7 | m_BDVER | m_SLM | m_GENERIC) -/* X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL: if true, unaligned loads are - split. */ -DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL, "256_unaligned_load_optimal", - ~(m_COREI7 | m_GENERIC)) - -/* X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL: if true, unaligned loads are - split. */ -DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL, "256_unaligned_load_optimal", - ~(m_COREI7 | m_BDVER | m_GENERIC)) - /* Use packed single precision instructions where posisble. I.e. movups instead of movupd. */ DEF_TUNE (X86_TUNE_SSE_PACKED_SINGLE_INSN_OPTIMAL, "sse_packed_single_insn_optimal", m_BDVER) -/* X86_TUNE_SSE_SPLIT_REGS: Set for machines where the type and dependencies - are resolved on SSE register parts instead of whole registers, so we may - maintain just lower part of scalar values in proper format leaving the - upper part undefined. */ -DEF_TUNE (X86_TUNE_SSE_SPLIT_REGS, "sse_split_regs", m_ATHLON_K8) - /* X86_TUNE_SSE_TYPELESS_STORES: Always movaps/movups for 128bit stores. */ DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES, "sse_typeless_stores", m_AMD_MULTIPLE | m_CORE_ALL | m_GENERIC) @@ -238,28 +312,6 @@ DEF_TUNE (X86_TUNE_SSE_TYPELESS_STORES, "sse_typeless_stores", DEF_TUNE (X86_TUNE_SSE_LOAD0_BY_PXOR, "sse_load0_by_pxor", m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_BDVER | m_BTVER | m_GENERIC) -/* X86_TUNE_MEMORY_MISMATCH_STALL: Avoid partial stores that are followed by - full sized loads. */ -DEF_TUNE (X86_TUNE_MEMORY_MISMATCH_STALL, "memory_mismatch_stall", - m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_AMD_MULTIPLE | m_GENERIC) - -/* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in prologues that are - considered on critical path. */ -DEF_TUNE (X86_TUNE_PROLOGUE_USING_MOVE, "prologue_using_move", - m_PPRO | m_ATHLON_K8) - -/* X86_TUNE_PROLOGUE_USING_MOVE: Do not use push/pop in epilogues that are - considered on critical path. */ -DEF_TUNE (X86_TUNE_EPILOGUE_USING_MOVE, "epilogue_using_move", - m_PPRO | m_ATHLON_K8) - -/* X86_TUNE_SHIFT1: Enables use of short encoding of "sal reg" instead of - longer "sal $1, reg". */ -DEF_TUNE (X86_TUNE_SHIFT1, "shift1", ~m_486) - -/* X86_TUNE_USE_FFREEP: Use freep instruction instead of fstp. */ -DEF_TUNE (X86_TUNE_USE_FFREEP, "use_ffreep", m_AMD_MULTIPLE) - /* X86_TUNE_INTER_UNIT_MOVES_TO_VEC: Enable moves in from integer to SSE registers. If disabled, the moves will be done by storing the value to memory and reloading. */ @@ -278,62 +330,80 @@ DEF_TUNE (X86_TUNE_INTER_UNIT_MOVES_FROM_VEC, "inter_unit_moves_from_vec", DEF_TUNE (X86_TUNE_INTER_UNIT_CONVERSIONS, "inter_unit_conversions", ~(m_AMDFAM10 | m_BDVER)) -/* X86_TUNE_FOUR_JUMP_LIMIT: Some CPU cores are not able to predict more - than 4 branch instructions in the 16 byte window. */ -DEF_TUNE (X86_TUNE_FOUR_JUMP_LIMIT, "four_jump_limit", - m_PPRO | m_P4_NOCONA | m_ATOM | m_SLM | m_ATHLON_K8 | m_AMDFAM10) +/* X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS: Try to split memory operand for + fp converts to destination register. */ +DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS, "split_mem_opnd_for_fp_converts", + m_SLM) -/* X86_TUNE_SCHEDULE: Enable scheduling. */ -DEF_TUNE (X86_TUNE_SCHEDULE, "schedule", - m_PENT | m_PPRO | m_CORE_ALL | m_ATOM | m_SLM | m_K6_GEODE - | m_AMD_MULTIPLE | m_GENERIC) +/* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion + from FP to FP. This form of instructions avoids partial write to the + destination. */ +DEF_TUNE (X86_TUNE_USE_VECTOR_FP_CONVERTS, "use_vector_fp_converts", + m_AMDFAM10) -/* X86_TUNE_USE_BT: Enable use of BT (bit test) instructions. */ -DEF_TUNE (X86_TUNE_USE_BT, "use_bt", - m_CORE_ALL | m_ATOM | m_SLM | m_AMD_MULTIPLE | m_GENERIC) +/* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion + from integer to FP. */ +DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS, "use_vector_converts", m_AMDFAM10) -/* X86_TUNE_USE_INCDEC: Enable use of inc/dec instructions. */ -DEF_TUNE (X86_TUNE_USE_INCDEC, "use_incdec", - ~(m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_GENERIC)) +/*****************************************************************************/ +/* AVX instruction selection tuning (some of SSE flags affects AVX, too) */ +/*****************************************************************************/ -/* X86_TUNE_PAD_RETURNS: Place NOP before every RET that is a destination - of conditional jump or directly preceded by other jump instruction. - This is important for AND K8-AMDFAM10 because the branch prediction - architecture expect at most one jump per 2 byte window. Failing to - pad returns leads to misaligned return stack. */ -DEF_TUNE (X86_TUNE_PAD_RETURNS, "pad_returns", - m_ATHLON_K8 | m_AMDFAM10 | m_GENERIC) +/* X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL: if true, unaligned loads are + split. */ +DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_LOAD_OPTIMAL, "256_unaligned_load_optimal", + ~(m_COREI7 | m_GENERIC)) -/* X86_TUNE_PAD_SHORT_FUNCTION: Make every function to be at least 4 - instructions long. */ -DEF_TUNE (X86_TUNE_PAD_SHORT_FUNCTION, "pad_short_function", m_ATOM) +/* X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL: if true, unaligned loads are + split. */ +DEF_TUNE (X86_TUNE_AVX256_UNALIGNED_STORE_OPTIMAL, "256_unaligned_load_optimal", + ~(m_COREI7 | m_BDVER | m_GENERIC)) -/* X86_TUNE_EXT_80387_CONSTANTS: Use fancy 80387 constants, such as PI. */ -DEF_TUNE (X86_TUNE_EXT_80387_CONSTANTS, "ext_80387_constants", - m_PPRO | m_P4_NOCONA | m_CORE_ALL | m_ATOM | m_SLM | m_K6_GEODE - | m_ATHLON_K8 | m_GENERIC) +/* X86_TUNE_AVX128_OPTIMAL: Enable 128-bit AVX instruction generation for + the auto-vectorizer. */ +DEF_TUNE (X86_TUNE_AVX128_OPTIMAL, "avx128_optimal", m_BDVER | m_BTVER2) -/* X86_TUNE_AVOID_VECTOR_DECODE: Enable splitters that avoid vector decoded - forms of instructions on K8 targets. */ -DEF_TUNE (X86_TUNE_AVOID_VECTOR_DECODE, "avoid_vector_decode", - m_K8) +/*****************************************************************************/ +/* Historical relics: tuning flags that helps a specific old CPU designs */ +/*****************************************************************************/ + +/* X86_TUNE_DOUBLE_WITH_ADD: Use add instead of sal to double value in + an integer register. */ +DEF_TUNE (X86_TUNE_DOUBLE_WITH_ADD, "double_with_add", ~m_386) + +/* X86_TUNE_ALWAYS_FANCY_MATH_387: controls use of fancy 387 operations, + such as fsqrt, fprem, fsin, fcos, fsincos etc. + Should be enabled for all targets that always has coprocesor. */ +DEF_TUNE (X86_TUNE_ALWAYS_FANCY_MATH_387, "always_fancy_math_387", + ~(m_386 | m_486)) + +/* X86_TUNE_UNROLL_STRLEN: Produce (quite lame) unrolled sequence for + inline strlen. This affects only -minline-all-stringops mode. By + default we always dispatch to a library since our internal strlen + is bad. */ +DEF_TUNE (X86_TUNE_UNROLL_STRLEN, "unroll_strlen", ~m_386) + +/* X86_TUNE_SHIFT1: Enables use of short encoding of "sal reg" instead of + longer "sal $1, reg". */ +DEF_TUNE (X86_TUNE_SHIFT1, "shift1", ~m_486) + +/* X86_TUNE_ZERO_EXTEND_WITH_AND: Use AND instruction instead + of mozbl/movwl. */ +DEF_TUNE (X86_TUNE_ZERO_EXTEND_WITH_AND, "zero_extend_with_and", m_486 | m_PENT) /* X86_TUNE_PROMOTE_HIMODE_IMUL: Modern CPUs have same latency for HImode and SImode multiply, but 386 and 486 do HImode multiply faster. */ DEF_TUNE (X86_TUNE_PROMOTE_HIMODE_IMUL, "promote_himode_imul", ~(m_386 | m_486)) -/* X86_TUNE_SLOW_IMUL_IMM32_MEM: Imul of 32-bit constant and memory is - vector path on AMD machines. - FIXME: Do we need to enable this for core? */ -DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM32_MEM, "slow_imul_imm32_mem", - m_K8 | m_AMDFAM10) +/* X86_TUNE_FAST_PREFIX: Enable demoting some 32bit or 64bit arithmetic + into 16bit/8bit when resulting sequence is shorter. For example + for "and $-65536, reg" to 16bit store of 0. */ +DEF_TUNE (X86_TUNE_FAST_PREFIX, "fast_prefix", ~(m_386 | m_486 | m_PENT)) -/* X86_TUNE_SLOW_IMUL_IMM8: Imul of 8-bit constant is vector path on AMD - machines. - FIXME: Do we need to enable this for core? */ -DEF_TUNE (X86_TUNE_SLOW_IMUL_IMM8, "slow_imul_imm8", - m_K8 | m_AMDFAM10) +/* X86_TUNE_READ_MODIFY_WRITE: Enable use of read modify write instructions + such as "add $1, mem". */ +DEF_TUNE (X86_TUNE_READ_MODIFY_WRITE, "read_modify_write", ~m_PENT) /* X86_TUNE_MOVE_M1_VIA_OR: On pentiums, it is faster to load -1 via OR than a MOV. */ @@ -343,87 +413,66 @@ DEF_TUNE (X86_TUNE_MOVE_M1_VIA_OR, "move_m1_via_or", m_PENT) but one byte longer. */ DEF_TUNE (X86_TUNE_NOT_UNPAIRABLE, "not_unpairable", m_PENT) -/* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory - operand that cannot be represented using a modRM byte. The XOR - replacement is long decoded, so this split helps here as well. */ -DEF_TUNE (X86_TUNE_NOT_VECTORMODE, "not_vectormode", m_K6) - -/* X86_TUNE_USE_VECTOR_FP_CONVERTS: Prefer vector packed SSE conversion - from FP to FP. This form of instructions avoids partial write to the - destination. */ -DEF_TUNE (X86_TUNE_USE_VECTOR_FP_CONVERTS, "use_vector_fp_converts", - m_AMDFAM10) - -/* X86_TUNE_USE_VECTOR_CONVERTS: Prefer vector packed SSE conversion - from integer to FP. */ -DEF_TUNE (X86_TUNE_USE_VECTOR_CONVERTS, "use_vector_converts", m_AMDFAM10) +/* X86_TUNE_PARTIAL_REG_STALL: Pentium pro, unlike later chips, handled + use of partial registers by renaming. This improved performance of 16bit + code where upper halves of registers are not used. It also leads to + an penalty whenever a 16bit store is followed by 32bit use. This flag + disables production of such sequences in common cases. + See also X86_TUNE_HIMODE_MATH. -/* X86_TUNE_FUSE_CMP_AND_BRANCH: Fuse a compare or test instruction - with a subsequent conditional jump instruction into a single - compare-and-branch uop. - FIXME: revisit for generic. */ -DEF_TUNE (X86_TUNE_FUSE_CMP_AND_BRANCH, "fuse_cmp_and_branch", m_BDVER | m_CORE_ALL) + In current implementation the partial register stalls are not eliminated + very well - they can be introduced via subregs synthesized by combine + and can happen in caller/callee saving sequences. */ +DEF_TUNE (X86_TUNE_PARTIAL_REG_STALL, "partial_reg_stall", m_PPRO) -/* X86_TUNE_OPT_AGU: Optimize for Address Generation Unit. This flag - will impact LEA instruction selection. */ -DEF_TUNE (X86_TUNE_OPT_AGU, "opt_agu", m_ATOM | m_SLM) +/* X86_TUNE_PROMOTE_QIMODE: When it is cheap, turn 8bit arithmetic to + corresponding 32bit arithmetic. */ +DEF_TUNE (X86_TUNE_PROMOTE_QIMODE, "promote_qimode", + ~m_PPRO) -/* X86_TUNE_VECTORIZE_DOUBLE: Enable double precision vector - instructions. */ -DEF_TUNE (X86_TUNE_VECTORIZE_DOUBLE, "vectorize_double", ~m_ATOM) +/* X86_TUNE_PROMOTE_HI_REGS: Same, but for 16bit artihmetic. Again we avoid + partial register stalls on PentiumPro targets. */ +DEF_TUNE (X86_TUNE_PROMOTE_HI_REGS, "promote_hi_regs", m_PPRO) -/* X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL: Enable software prefetching - at -O3. For the moment, the prefetching seems badly tuned for Intel - chips. */ -DEF_TUNE (X86_TUNE_SOFTWARE_PREFETCHING_BENEFICIAL, "software_prefetching_beneficial", - m_K6_GEODE | m_AMD_MULTIPLE) +/* X86_TUNE_HIMODE_MATH: Enable use of 16bit arithmetic. + On PPro this flag is meant to avoid partial register stalls. */ +DEF_TUNE (X86_TUNE_HIMODE_MATH, "himode_math", ~m_PPRO) -/* X86_TUNE_AVX128_OPTIMAL: Enable 128-bit AVX instruction generation for - the auto-vectorizer. */ -DEF_TUNE (X86_TUNE_AVX128_OPTIMAL, "avx128_optimal", m_BDVER | m_BTVER2) +/* X86_TUNE_SPLIT_LONG_MOVES: Avoid instructions moving immediates + directly to memory. */ +DEF_TUNE (X86_TUNE_SPLIT_LONG_MOVES, "split_long_moves", m_PPRO) -/* X86_TUNE_REASSOC_INT_TO_PARALLEL: Try to produce parallel computations - during reassociation of integer computation. */ -DEF_TUNE (X86_TUNE_REASSOC_INT_TO_PARALLEL, "reassoc_int_to_parallel", - m_ATOM) +/* X86_TUNE_USE_XCHGB: Use xchgb %rh,%rl instead of rolw/rorw $8,rx. */ +DEF_TUNE (X86_TUNE_USE_XCHGB, "use_xchgb", m_PENT4) -/* X86_TUNE_REASSOC_FP_TO_PARALLEL: Try to produce parallel computations - during reassociation of fp computation. */ -DEF_TUNE (X86_TUNE_REASSOC_FP_TO_PARALLEL, "reassoc_fp_to_parallel", - m_ATOM | m_SLM | m_HASWELL | m_BDVER1 | m_BDVER2 | m_GENERIC) +/* X86_TUNE_USE_MOV0: Use "mov $0, reg" instead of "xor reg, reg" to clear + integer register. */ +DEF_TUNE (X86_TUNE_USE_MOV0, "use_mov0", m_K6) -/* X86_TUNE_GENERAL_REGS_SSE_SPILL: Try to spill general regs to SSE - regs instead of memory. */ -DEF_TUNE (X86_TUNE_GENERAL_REGS_SSE_SPILL, "general_regs_sse_spill", - m_CORE_ALL) +/* X86_TUNE_NOT_VECTORMODE: On AMD K6, NOT is vector decoded with memory + operand that cannot be represented using a modRM byte. The XOR + replacement is long decoded, so this split helps here as well. */ +DEF_TUNE (X86_TUNE_NOT_VECTORMODE, "not_vectormode", m_K6) -/* X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE: Try to avoid memory operands for - a conditional move. */ -DEF_TUNE (X86_TUNE_AVOID_MEM_OPND_FOR_CMOVE, "avoid_mem_opnd_for_cmove", - m_ATOM | m_SLM) +/* X86_TUNE_AVOID_VECTOR_DECODE: Enable splitters that avoid vector decoded + forms of instructions on K8 targets. */ +DEF_TUNE (X86_TUNE_AVOID_VECTOR_DECODE, "avoid_vector_decode", + m_K8) -/* X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS: Try to split memory operand for - fp converts to destination register. */ -DEF_TUNE (X86_TUNE_SPLIT_MEM_OPND_FOR_FP_CONVERTS, "split_mem_opnd_for_fp_converts", - m_SLM) +/*****************************************************************************/ +/* This never worked well before. */ +/*****************************************************************************/ -/* X86_TUNE_ACCUMULATE_OUTGOING_ARGS: Allocate stack space for outgoing - arguments in prologue/epilogue instead of separately for each call - by push/pop instructions. - This increase code size by about 5% in 32bit mode, less so in 64bit mode - because parameters are passed in registers. It is considerable - win for targets without stack engine that prevents multple push operations - to happen in parallel. +/* X86_TUNE_BRANCH_PREDICTION_HINTS: Branch hints were put in P4 based + on simulation result. But after P4 was made, no performance benefit + was observed with branch hints. It also increases the code size. + As a result, icc never generates branch hints. */ +DEF_TUNE (X86_TUNE_BRANCH_PREDICTION_HINTS, "branch_prediction_hints", 0) - FIXME: the flags is incorrectly enabled for amdfam10, Bulldozer, - Bobcat and Generic. This is because disabling it causes large - regression on mgrid due to IRA limitation leading to unecessary - use of the frame pointer in 32bit mode. */ -DEF_TUNE (X86_TUNE_ACCUMULATE_OUTGOING_ARGS, "accumulate_outgoing_args", - m_PPRO | m_P4_NOCONA | m_ATOM | m_SLM | m_AMD_MULTIPLE | m_GENERIC) +/* X86_TUNE_QIMODE_MATH: Enable use of 8bit arithmetic. */ +DEF_TUNE (X86_TUNE_QIMODE_MATH, "qimode_math", ~0) -/* X86_TUNE_ALWAYS_FANCY_MATH_387: controls use of fancy 387 operations, - such as fsqrt, fprem, fsin, fcos, fsincos etc. - Should be enabled for all targets that always has coprocesor. */ -DEF_TUNE (X86_TUNE_ALWAYS_FANCY_MATH_387, "always_fancy_math_387", - ~(m_386 | m_486)) +/* X86_TUNE_PROMOTE_QI_REGS: This enables generic code that promotes all 8bit + arithmetic to 32bit via PROMOTE_MODE macro. This code generation scheme + is usually used for RISC targets. */ +DEF_TUNE (X86_TUNE_PROMOTE_QI_REGS, "promote_qi_regs", 0) -- 2.7.4