无论是投影还是别的计算,表达式执行的入口和计算逻辑都是统一的,这里已投影为分析表达式执行的流程。
1 投影函数
用例
create table t1(i int primary key, j int, k int);
insert into t1 select i, i % 10, i % 100 from generate_series(1,10000000) t(i);
explain analyze select abs(k),abs(k),abs(k),abs(k),abs(k),exp(k),exp(k),exp(k),exp(k),exp(k) from t1;
对于这样这一条查询来说,每扫描一行,都会调用投影函数ExecProject,完成最终结果的构造。
投影函数:
static inline TupleTableSlot *
ExecProject(ProjectionInfo *projInfo)
{ExprContext *econtext = projInfo->pi_exprContext;ExprState *state = &projInfo->pi_state;TupleTableSlot *slot = state->resultslot;bool isnull;/** Clear any former contents of the result slot. This makes it safe for* us to use the slot's Datum/isnull arrays as workspace.*/ExecClearTuple(slot);/* Run the expression, discarding scalar result from the last column. */(void) ExecEvalExprSwitchContext(state, econtext, &isnull);/** Successfully formed a result row. Mark the result slot as containing a* valid virtual tuple (inlined version of ExecStoreVirtualTuple()).*/slot->tts_flags &= ~TTS_FLAG_EMPTY;slot->tts_nvalid = slot->tts_tupleDescriptor->natts;return slot;
}
总结:
- projInfo->pi_exprContext记录了需要执行表达式的上下文信息
- 具体存放:t1表扫出来的一行
- 具体存放:表达式执行的内存上下文
- projInfo->pi_state记录了表达式执行状态
- 具体存放:表达式执行的每一个step,每一个step放到ExecInterpExpr中按顺序执行得到最终结果tuple。
2.1 表达式执行上下文结构ExprContext
其中ecxt_scantuple用来保存待处理输入tuple,通常指向一个 tuple table slot。
ExprContext *econtext = {type = T_ExprContext,ecxt_scantuple = 0x1520918, <<<<<<<< 待处理tupleecxt_innertuple = 0x0,ecxt_outertuple = 0x0,ecxt_per_query_memory = 0x15203e0,ecxt_per_tuple_memory = 0x151e3d0,ecxt_param_exec_vals = 0x0,ecxt_param_list_info = 0x0,ecxt_aggvalues = 0x0,ecxt_aggnulls = 0x0,caseValue_datum = 0,caseValue_isNull = true,domainValue_datum = 0,domainValue_isNull = true,ecxt_estate = 0x15204e0,ecxt_callbacks = 0x0
}
2.2 表达式执行状态ExprState
ExprState *state = {type = T_ExprState,flags = 6 '\006',resnull = false,resvalue = 0,resultslot = 0x151f278,steps = 0x1525c38,evalfunc = 0x74d749 <ExecInterpExprStillValid>,expr = 0x24c9618,evalfunc_private = 0x74af83 <ExecInterpExpr>,steps_len = 37,steps_alloc = 64,parent = 0x151e718,ext_params = 0x0,innermost_caseval = 0x0,innermost_casenull = 0x0,innermost_domainval = 0x0,innermost_domainnull = 0x0,escontext = 0x0
}
注意实际执行使用ExecInterpExpr函数完成表达式计算:
2.3 执行表达式
执行时,通过state->evalfunc函数完成具体的表达式计算动作。具体怎么计算流程放在steps中:
(gdb) p/x state->steps[0]->opcode
$15 = 0x74b0ad
(gdb) p/x state->steps[1]->opcode
$16 = 0x74b1ec
(gdb) p/x state->steps[2]->opcode
$17 = 0x74b784
(gdb) p/x state->steps[3]->opcode
$18 = 0x74b591
(gdb) p/x state->steps[4]->opcode
$19 = 0x74b1ec
(gdb) p/x state->steps[5]->opcode
$20 = 0x74b784
(gdb) p/x state->steps[6]->opcode
$21 = 0x74b591
(gdb) p/x state->steps[7]->opcode
$22 = 0x74b1ec
...
...
opcode需要再reverse_dispatch_table中确定当前计算走哪个ExecInterpExpr中的分支:
(这里不太方便调试因为这里把一个超大的switch case改成了goto,为了性能!)
p reverse_dispatch_table
$13 = {{opcode = 0x74b01a <ExecInterpExpr+151>,op = EEOP_DONE}, {opcode = 0x74b041 <ExecInterpExpr+190>,op = EEOP_INNER_FETCHSOME}, {opcode = 0x74b077 <ExecInterpExpr+244>,op = EEOP_OUTER_FETCHSOME}, {opcode = 0x74b0ad <ExecInterpExpr+298>,op = EEOP_SCAN_FETCHSOME}, {opcode = 0x74b0e6 <ExecInterpExpr+355>,op = EEOP_INNER_VAR}, {opcode = 0x74b169 <ExecInterpExpr+486>,op = EEOP_OUTER_VAR}, {opcode = 0x74b1ec <ExecInterpExpr+617>,op = EEOP_SCAN_VAR}, {opcode = 0x74b26f <ExecInterpExpr+748>,op = EEOP_INNER_SYSVAR}, {opcode = 0x74b29e <ExecInterpExpr+795>,op = EEOP_OUTER_SYSVAR}, {opcode = 0x74b2cd <ExecInterpExpr+842>,op = EEOP_SCAN_SYSVAR}, {opcode = 0x74b2fc <ExecInterpExpr+889>,op = EEOP_WHOLEROW}, {opcode = 0x74b32a <ExecInterpExpr+935>,op = EEOP_ASSIGN_INNER_VAR}, {opcode = 0x74b3f7 <ExecInterpExpr+1140>,op = EEOP_ASSIGN_OUTER_VAR}, {opcode = 0x74b4c4 <ExecInterpExpr+1345>,op = EEOP_ASSIGN_SCAN_VAR}, {opcode = 0x74b591 <ExecInterpExpr+1550>,op = EEOP_ASSIGN_TMP}, {opcode = 0x74b615 <ExecInterpExpr+1682>,op = EEOP_ASSIGN_TMP_MAKE_RO}, {opcode = 0x74b6e1 <ExecInterpExpr+1886>,op = EEOP_CONST}, {opcode = 0x74b717 <ExecInterpExpr+1940>,op = EEOP_FUNCEXPR}, {opcode = 0x74b784 <ExecInterpExpr+2049>,op = EEOP_FUNCEXPR_STRICT}, {opcode = 0x74b853 <ExecInterpExpr+2256>,op = EEOP_FUNCEXPR_FUSAGE}, {opcode = 0x74b881 <ExecInterpExpr+2302>,op = EEOP_FUNCEXPR_STRICT_FUSAGE}, {opcode = 0x74b8af <ExecInterpExpr+2348>,op = EEOP_BOOL_AND_STEP_FIRST}, {opcode = 0x74b8ba <ExecInterpExpr+2359>,op = EEOP_BOOL_AND_STEP}, {opcode = 0x74b92c <ExecInterpExpr+2473>,op = EEOP_BOOL_AND_STEP_LAST}, {opcode = 0x74b98f <ExecInterpExpr+2572>,op = EEOP_BOOL_OR_STEP_FIRST}, {opcode = 0x74b99a <ExecInterpExpr+2583>,op = EEOP_BOOL_OR_STEP}, {opcode = 0x74ba09 <ExecInterpExpr+2694>,op = EEOP_BOOL_OR_STEP_LAST}, {opcode = 0x74ba69 <ExecInterpExpr+2790>,op = EEOP_BOOL_NOT_STEP}, {opcode = 0x74bab3 <ExecInterpExpr+2864>,op = EEOP_QUAL}, {opcode = 0x74bb38 <ExecInterpExpr+2997>,op = EEOP_JUMP}, {opcode = 0x74bb63 <ExecInterpExpr+3040>,op = EEOP_JUMP_IF_NULL}, {opcode = 0x74bbae <ExecInterpExpr+3115>,op = EEOP_JUMP_IF_NOT_NULL}, {opcode = 0x74bbfc <ExecInterpExpr+3193>,op = EEOP_JUMP_IF_NOT_TRUE}, {opcode = 0x74bc61 <ExecInterpExpr+3294>,op = EEOP_NULLTEST_ISNULL}, {opcode = 0x74bc9d <ExecInterpExpr+3354>,op = EEOP_NULLTEST_ISNOTNULL}, {opcode = 0x74bcea <ExecInterpExpr+3431>,op = EEOP_NULLTEST_ROWISNULL}, {opcode = 0x74bd18 <ExecInterpExpr+3477>,op = EEOP_NULLTEST_ROWISNOTNULL}, {opcode = 0x74bd46 <ExecInterpExpr+3523>,op = EEOP_BOOLTEST_IS_TRUE}, {opcode = 0x74bd86 <ExecInterpExpr+3587>,op = EEOP_BOOLTEST_IS_NOT_TRUE}, {opcode = 0x74be01 <ExecInterpExpr+3710>,op = EEOP_BOOLTEST_IS_FALSE}, {opcode = 0x74be7c <ExecInterpExpr+3833>,op = EEOP_BOOLTEST_IS_NOT_FALSE}, {opcode = 0x74bebc <ExecInterpExpr+3897>,op = EEOP_PARAM_EXEC}, {opcode = 0x74beea <ExecInterpExpr+3943>,op = EEOP_PARAM_EXTERN}, {opcode = 0x74bf18 <ExecInterpExpr+3989>,op = EEOP_PARAM_CALLBACK}, {opcode = 0x74bf48 <ExecInterpExpr+4037>,op = EEOP_CASE_TESTVAL}, {opcode = 0x74bfbe <ExecInterpExpr+4155>,op = EEOP_DOMAIN_TESTVAL}, {opcode = 0x74c034 <ExecInterpExpr+4273>,op = EEOP_MAKE_READONLY}, {opcode = 0x74c08a <ExecInterpExpr+4359>,op = EEOP_IOCOERCE}, {opcode = 0x74c26c <ExecInterpExpr+4841>,op = EEOP_IOCOERCE_SAFE}, {opcode = 0x74c293 <ExecInterpExpr+4880>,op = EEOP_DISTINCT}, {opcode = 0x74c3a6 <ExecInterpExpr+5155>,op = EEOP_NOT_DISTINCT}, {opcode = 0x74c496 <ExecInterpExpr+5395>,op = EEOP_NULLIF}, {opcode = 0x74c57f <ExecInterpExpr+5628>,op = EEOP_SQLVALUEFUNCTION}, {opcode = 0x74c5a6 <ExecInterpExpr+5667>,op = EEOP_CURRENTOFEXPR}, {opcode = 0x74c5cd <ExecInterpExpr+5706>,op = EEOP_NEXTVALUEEXPR}, {opcode = 0x74c5f4 <ExecInterpExpr+5745>,op = EEOP_ARRAYEXPR}, {opcode = 0x74c61b <ExecInterpExpr+5784>,op = EEOP_ARRAYCOERCE}, {opcode = 0x74c649 <ExecInterpExpr+5830>,op = EEOP_ROW}, {opcode = 0x74c670 <ExecInterpExpr+5869>,op = EEOP_ROWCOMPARE_STEP}, {opcode = 0x74c7be <ExecInterpExpr+6203>,op = EEOP_ROWCOMPARE_FINAL}, {opcode = 0x74c8cf <ExecInterpExpr+6476>,op = EEOP_MINMAX}, {opcode = 0x74c8f6 <ExecInterpExpr+6515>,op = EEOP_FIELDSELECT}, {opcode = 0x74c924 <ExecInterpExpr+6561>,op = EEOP_FIELDSTORE_DEFORM}, {opcode = 0x74c952 <ExecInterpExpr+6607>,op = EEOP_FIELDSTORE_FORM}, {opcode = 0x74c980 <ExecInterpExpr+6653>,op = EEOP_SBSREF_SUBSCRIPTS}, {opcode = 0x74c9df <ExecInterpExpr+6748>,op = EEOP_SBSREF_OLD}, {opcode = 0x74c9e1 <ExecInterpExpr+6750>,op = EEOP_SBSREF_ASSIGN}, {opcode = 0x74c9e1 <ExecInterpExpr+6750>,op = EEOP_SBSREF_FETCH}, {opcode = 0x74ca11 <ExecInterpExpr+6798>,op = EEOP_CONVERT_ROWTYPE}, {opcode = 0x74ca3f <ExecInterpExpr+6844>,op = EEOP_SCALARARRAYOP}, {opcode = 0x74ca66 <ExecInterpExpr+6883>,op = EEOP_HASHED_SCALARARRAYOP}, {opcode = 0x74ca94 <ExecInterpExpr+6929>,op = EEOP_DOMAIN_NOTNULL}, {opcode = 0x74cabb <ExecInterpExpr+6968>,op = EEOP_DOMAIN_CHECK}, {opcode = 0x74cae2 <ExecInterpExpr+7007>,op = EEOP_XMLEXPR}, {opcode = 0x74cb09 <ExecInterpExpr+7046>,op = EEOP_JSON_CONSTRUCTOR}, {opcode = 0x74cb37 <ExecInterpExpr+7092>,op = EEOP_IS_JSON}, {opcode = 0x74cb5e <ExecInterpExpr+7131>,op = EEOP_JSONEXPR_PATH}, {opcode = 0x74cb9f <ExecInterpExpr+7196>,op = EEOP_JSONEXPR_COERCION}, {opcode = 0x74cbcd <ExecInterpExpr+7242>,op = EEOP_JSONEXPR_COERCION_FINISH}, {opcode = 0x74cbf4 <ExecInterpExpr+7281>,op = EEOP_AGGREF}, {opcode = 0x74cc82 <ExecInterpExpr+7423>,op = EEOP_GROUPING_FUNC}, {opcode = 0x74cca9 <ExecInterpExpr+7462>,op = EEOP_WINDOW_FUNC}, {opcode = 0x74cd40 <ExecInterpExpr+7613>,op = EEOP_MERGE_SUPPORT_FUNC}, {opcode = 0x74cd6e <ExecInterpExpr+7659>,op = EEOP_SUBPLAN}, {opcode = 0x74cd9c <ExecInterpExpr+7705>,op = EEOP_AGG_STRICT_DESERIALIZE}, {opcode = 0x74cdd7 <ExecInterpExpr+7764>,op = EEOP_AGG_DESERIALIZE}, {opcode = 0x74ce8a <ExecInterpExpr+7943>,op = EEOP_AGG_STRICT_INPUT_CHECK_ARGS}, {opcode = 0x74cf18 <ExecInterpExpr+8085>,op = EEOP_AGG_STRICT_INPUT_CHECK_NULLS}, {opcode = 0x74cf9f <ExecInterpExpr+8220>,op = EEOP_AGG_PLAIN_PERGROUP_NULLCHECK}, {opcode = 0x74d02c <ExecInterpExpr+8361>,op = EEOP_AGG_PLAIN_TRANS_INIT_STRICT_BYVAL}, {opcode = 0x74d147 <ExecInterpExpr+8644>,op = EEOP_AGG_PLAIN_TRANS_STRICT_BYVAL}, {opcode = 0x74d22c <ExecInterpExpr+8873>,op = EEOP_AGG_PLAIN_TRANS_BYVAL}, {opcode = 0x74d2fb <ExecInterpExpr+9080>,op = EEOP_AGG_PLAIN_TRANS_INIT_STRICT_BYREF}, {opcode = 0x74d413 <ExecInterpExpr+9360>,op = EEOP_AGG_PLAIN_TRANS_STRICT_BYREF}, {opcode = 0x74d4f5 <ExecInterpExpr+9586>,op = EEOP_AGG_PLAIN_TRANS_BYREF}, {opcode = 0x74d5c1 <ExecInterpExpr+9790>,op = EEOP_AGG_PRESORTED_DISTINCT_SINGLE}, {opcode = 0x74d648 <ExecInterpExpr+9925>,op = EEOP_AGG_PRESORTED_DISTINCT_MULTI}, {opcode = 0x74d6cf <ExecInterpExpr+10060>,op = EEOP_AGG_ORDERED_TRANS_DATUM}, {opcode = 0x74d6fd <ExecInterpExpr+10106>,op = EEOP_AGG_ORDERED_TRANS_TUPLE}}
翻译后
(gdb) p/x state->steps[0]->opcode
$15 = 0x74b0ad EEOP_SCAN_FETCHSOME
(gdb) p/x state->steps[1]->opcode
$16 = 0x74b1ec EEOP_SCAN_VAR
(gdb) p/x state->steps[2]->opcode
$17 = 0x74b784 EEOP_FUNCEXPR_STRICT
(gdb) p/x state->steps[3]->opcode
$18 = 0x74b591 EEOP_ASSIGN_TMP
(gdb) p/x state->steps[4]->opcode
$19 = 0x74b1ec EEOP_SCAN_VAR
(gdb) p/x state->steps[5]->opcode
$20 = 0x74b784 EEOP_FUNCEXPR_STRICT
(gdb) p/x state->steps[6]->opcode
$21 = 0x74b591 EEOP_ASSIGN_TMP
(gdb) p/x state->steps[7]->opcode
$22 = 0x74b1ec EEOP_SCAN_VAR
...
...
(gdb) p/x state->steps[34]->opcode
$27 = 0x74b784 EEOP_FUNCEXPR_STRICT
(gdb) p/x state->steps[35]->opcode
$28 = 0x74b591 EEOP_ASSIGN_TMP
(gdb) p/x state->steps[36]->opcode
$29 = 0x74b01a EEOP_DONE
可以看到表达式计算的流程:
第一步:EEOP_SCAN_FETCHSOME
- 从econtext->ecxt_scantuple读取到scanslot(当前要处理的一行数据)
- slot_getsomeattrs函数确保这一行数据中,至少有op->d.fetch.last_var个列是可以直接访问的(这里是3,因为t1表就三列,后面的处理可能需要访问这三列的任意一列)。为什么说有时不能直接访问,因为列有可能指向toast表。
EEO_CASE(EEOP_SCAN_FETCHSOME){CheckOpSlotCompatibility(op, scanslot);slot_getsomeattrs(scanslot, op->d.fetch.last_var);EEO_NEXT();}
第二步:EEOP_SCAN_VAR
输入
p state->steps[1].d.fetch
$46 = {last_var = 2,fixed = 23,known_desc = 0x0,kind = 0x0
}
执行,从行中拿到第2列的值(0列、1列、2列)
EEO_CASE(EEOP_SCAN_VAR){int attnum = op->d.var.attnum; // attnum = 2/* See EEOP_INNER_VAR comments */Assert(attnum >= 0 && attnum < scanslot->tts_nvalid);*op->resvalue = scanslot->tts_values[attnum];*op->resnull = scanslot->tts_isnull[attnum];EEO_NEXT();}
结果
(gdb) p state->steps[1].resvalue
$41 = (Datum *) 0x151f918
(gdb) p *state->steps[1].resvalue
$39 = 1(gdb) p state->steps[1].resnull
$42 = (_Bool *) 0x151f920
(gdb) p *state->steps[1].resnull
$40 = false
第三步:EEOP_FUNCEXPR_STRICT
输入
(gdb) p state->steps[2].d.func
$47 = {finfo = 0x151f8a8,fcinfo_data = 0x151f8f8,fn_addr = 0xa8ea89 <int4abs>,nargs = 1
}
执行
EEO_CASE(EEOP_FUNCEXPR_STRICT){FunctionCallInfo fcinfo = op->d.func.fcinfo_data;NullableDatum *args = fcinfo->args;int nargs = op->d.func.nargs;Datum d;/* strict function, so check for NULL args */for (int argno = 0; argno < nargs; argno++){if (args[argno].isnull){*op->resnull = true;goto strictfail;}}fcinfo->isnull = false;d = op->d.func.fn_addr(fcinfo);*op->resvalue = d;*op->resnull = fcinfo->isnull;strictfail:EEO_NEXT();}
注意这里有一个地方很有意思,按理说EEOP_SCAN_VAR执行完才把值拿到,但从EEOP_FUNCEXPR_STRICT的执行来看,并没有发现把前一步的结果,放到函数args的步骤。
但是从GDB来看,函数的入参的地址和上一步取值后存放结果的地址是相同的,也就是上一步取值就是为了拿入参的args:
(gdb) p state->steps[4].resvalue
$59 = (Datum *) 0x151f9b8(gdb) p state->steps[5].d.func->fcinfo_data->args
$58 = 0x151f9b8
具体是怎么做到的呢?在构造steps时:
对于函数入参value会调用ExecInitExprRec去取值,在这个过程中,把参数的value指向新step的resvalue:
而这个新的step就是EEOP_SCAN_VAR:
第四步:暂存结果集
暂存结果集到resultslot中:
EEO_CASE(EEOP_ASSIGN_TMP){int resultnum = op->d.assign_tmp.resultnum;Assert(resultnum >= 0 && resultnum < resultslot->tts_tupleDescriptor->natts);resultslot->tts_values[resultnum] = state->resvalue;resultslot->tts_isnull[resultnum] = state->resnull;EEO_NEXT();}
第五步:继续上述流程直到执行完成
(gdb) p/x state->steps[0]->opcode
$15 = 0x74b0ad EEOP_SCAN_FETCHSOME 第一步
(gdb) p/x state->steps[1]->opcode
$16 = 0x74b1ec EEOP_SCAN_VAR 第二步
(gdb) p/x state->steps[2]->opcode
$17 = 0x74b784 EEOP_FUNCEXPR_STRICT 第三步
(gdb) p/x state->steps[3]->opcode
$18 = 0x74b591 EEOP_ASSIGN_TMP 第四步
(gdb) p/x state->steps[4]->opcode
$19 = 0x74b1ec EEOP_SCAN_VAR 第五步 和上述流程相同,每个函数的执行流程都是相似的
(gdb) p/x state->steps[5]->opcode
$20 = 0x74b784 EEOP_FUNCEXPR_STRICT
(gdb) p/x state->steps[6]->opcode
$21 = 0x74b591 EEOP_ASSIGN_TMP
(gdb) p/x state->steps[7]->opcode
$22 = 0x74b1ec EEOP_SCAN_VAR
...
...
(gdb) p/x state->steps[34]->opcode
$27 = 0x74b784 EEOP_FUNCEXPR_STRICT
(gdb) p/x state->steps[35]->opcode
$28 = 0x74b591 EEOP_ASSIGN_TMP
(gdb) p/x state->steps[36]->opcode
$29 = 0x74b01a EEOP_DONE