ftb 小测试
源代码说明
程序位置在
https://github.com/OpenXiangShan/nexus-am/blob/master/tests/frontendtest/tests/tage1.c
#include "common.h"
void __attribute__ ((noinline)) branch_test(int cnt) {
int tmp = 0;
// 设计交替的分支模式：
// 通过判断t2的奇偶性来实现T-NT交替
#define ONE \
"andi t4, t2, 1\n\t" /* 检查t2是否为奇数 */ \
"beqz t4, 2f\n\t" /* 如果是偶数则跳转 */ \
"nop\n\t" \
"2:\n\t"
asm volatile(
// 初始化计数器
"li t0, 0\n\t" // 循环计数器
"li t2, 0\n\t" // 用于产生交替模式的计数器
".align 4\n\t"
"1:\n\t"
// 重复执行分支指令
TEN // 每次循环执行10次分支指令
"addi t2, t2, 1\n\t" // t2每次加1，产生0,1,2,3...序列
"addi t0, t0, 1\n\t" // 增加循环计数器
"blt t0, %1, 1b\n\t" // 循环控制
: "+r"(tmp)
: "r"(cnt)
: "t0", "t2", "t3", "t4", "memory"
);
}
int main() {
branch_test(1000); // 执行1000次循环
return 0;
}
// 这种情况下，base 表几乎都完全命中，导致几乎不会分配新的tage 表项
主要逻辑：
1. 外层循环执行1000次(cnt=1000)
2. 每次循环中：
- 执行10次相同的分支指令(TEN = ONE * 10)
- t2计数器每次加1
- t0作为循环计数器
ONE宏的分支逻辑：
1. andi t4, t2, 1 // 取t2的最低位判断奇偶性
2. beqz t4, 2f // 如果t4为0(t2为偶数)则跳转到2标签
3. nop // 延迟槽指令
4. 2: // 跳转目标
t2的变化：0,1,2,3,4,5,6,7...
对应的跳转行为：
- t4=0 (偶数): 跳转(Taken)
- t4=1 (奇数): 不跳转(Not Taken)
- t4=2 (偶数): 跳转(Taken)
- t4=3 (奇数): 不跳转(Not Taken)
...
形成规律：T-NT-T-NT-T-NT...的交替模式
反汇编
/nfs/home/yanyue/tools/nexus-am-xs/tests/cputest/build/tage1-riscv64-xs.txt
80000130: 0013fe93 and t4,t2,1
80000134: 000e8363 beqz t4,8000013a <branch_test+0x1a>
80000138: 0001 nop
8000013a: 0013fe93 and t4,t2,1
8000013e: 000e8363 beqz t4,80000144 <branch_test+0x24>
80000142: 0001 nop
80000144: 0013fe93 and t4,t2,1
80000148: 000e8363 beqz t4,8000014e <branch_test+0x2e>
8000014c: 0001 nop
主要关注130 是起始地址
134 是第一个br, 其中为T-NT-T-NT 交错执行
如果T, target = 13a; NT target= 138
测试输出
首先看CommitTrace 查看小测试执行情况是否如预期， 这表示程序最终commit 顺序
./build/RISCV/gem5.debug --outdir=debug/tage1 --debug-flags=CommitTrace --debug-file=tage1.commit ./configs/example/kmhv3.py --generic-rv-cpt=/nfs/home/yanyue/tools/nexus-am-xs/tests/cputest/build/tage1-riscv64-xs.bin --raw-cpt
549783: global: [c: 1651] system.cpu [sn:1500 pc:0x80000130] enDqT: 547452, exDqT: 547785, readyT: 547785, CompleT:548784, andi t4, t2, 1, res: 0
549783: global: [c: 1651] system.cpu [sn:1501 pc:0x80000134] enDqT: 547452, exDqT: 547785, readyT: 548118, CompleT:549117, beq t4, zero, 6
555444: global: [c: 1668] system.cpu [sn:1600 pc:0x8000013a] enDqT: 553446, exDqT: 553779, readyT: 553779, CompleT:554778, andi t4, t2, 1, res: 0
...
559440: global: [c: 1680] system.cpu [sn:1621 pc:0x80000130] enDqT: 557442, exDqT: 557775, readyT: 557775, CompleT:558774, andi t4, t2, 1, res: 0x1
559773: global: [c: 1681] system.cpu [sn:1622 pc:0x80000134] enDqT: 557442, exDqT: 557775, readyT: 558108, CompleT:559107, beq t4, zero, 6
559773: global: [c: 1681] system.cpu [sn:1623 pc:0x80000138] enDqT: 557442, exDqT: 557775, readyT: 557775, CompleT:559107, c_addi zero, 0
559773: global: [c: 1681] system.cpu [sn:1624 pc:0x8000013a] enDqT: 557442, exDqT: 557775, readyT: 557775, CompleT:558774, andi t4, t2, 1, res: 0x1
能看出第一次 t4 = 0, 134 这个beq 发生taken, 跳到13a 的andi 指令执行
第二次 t4=1, 134 这个beq NT, 顺序执行138 这个nop指令（c_addi), 然后再执行13a 这个andi 指令
然后可以查看stats.txt 获取基本信息，一般主要关注如下
system.cpu.numCycles 10861
system.cpu.numInsts 28166
system.cpu.numBranches 11023
system.cpu.ipc 2.587239
system.cpu.commit.branchMispredicts 23 # commit 阶段的最终错误预测指令
# 来自哪个阶段的预测结果，基本都是0阶段
system.cpu.branchPred.predsOfEachStage::0 7246 # the number of preds of each stage that account for final pred (Count)
system.cpu.branchPred.predsOfEachStage::1 979 # the number of preds of each stage that account for final pred (Count)
system.cpu.branchPred.predsOfEachStage::2 0
# fsq 中包含的指令数目，基本是2条，9条，10条
system.cpu.branchPred.commitFsqEntryHasInsts::2 5002 66.53% 66.56% # number of insts that commit fsq entries have (Count)
system.cpu.branchPred.commitFsqEntryHasInsts::9 500 6.65% 86.70% # number of insts that commit fsq entries have (Count)
system.cpu.branchPred.commitFsqEntryHasInsts::10 1000 13.30% 100.00% # number of insts that commit fsq entries have (Count)
system.cpu.branchPred.ftbHit 6483 # ftb hits (in predict block) (Count)
system.cpu.branchPred.ftbMiss 25 # ftb misses (in predict block) (Count)
# 主要关注updateMiss, 这是实际正确路径上ftb miss 情况，predMiss 包括错误路径的
system.cpu.branchPred.ftb.predMiss 1601 # misses encountered on prediction (Count)
system.cpu.branchPred.ftb.predHit 6625 # hits encountered on prediction (Count)
system.cpu.branchPred.ftb.updateMiss 23 # misses encountered on update (Count)
system.cpu.branchPred.ftb.updateHit 6485 # hits encountered on update (Count)
还有tage, 等各类计数器，选择性观看
最后可以看看 各类debug-flags=DecoupleBP,FTB
./build/RISCV/gem5.debug --outdir=debug/tage1 --debug-flags=DecoupleBP,FTB --debug-file=tage1.bp2 --debug-end=1000000 ./configs/example/kmhv3.py --generic-rv-cpt=/nfs/home/yanyue/tools/nexus-am-xs/tests/cputest/build/tage1-riscv64-xs.bin --raw-cpt
输出文件位置在/nfs/home/yanyue/workspace/GEM5/debug/tage1/tage1.bp2
655011: system.cpu.branchPred: [c: 1967] looking up pc 0x80000134
655011: system.cpu.branchPred: [c: 1967] Supplying fetch with target ID 161
655011: system.cpu.branchPred: [c: 1967] Responsing fetch with FTQ:: 0x80000130 - [0, 0x80000150) --> 0, taken: 0, fsqID: 162, loop: 0, iter: 0, exit: 0
655011: system.cpu.branchPred: [c: 1967] FSQ: 0x80000130-[0, 0) --> 0, taken: 0, predEndPC: 0x80000150, isHit: 1, falseHit: 0
655011: system.cpu.branchPred: [c: 1967] FTB entry: valid 1, tag 0, fallThruAddr:0x80000150, slots:
655011: system.cpu.branchPred: [c: 1967] pc:0x80000134, size:4, target:0x8000013a, cond:1, indirect:0, call:0, return:0
655011: system.cpu.branchPred: [c: 1967] Predict it not taken to 0x80000138
656676: system.cpu.branchPred: [c: 1972] looking up pc 0x80000134
656676: system.cpu.branchPred: [c: 1972] Supplying fetch with target ID 165
656676: system.cpu.branchPred: [c: 1972] Responsing fetch with:: 0x80000130 - [0x80000134, 0x80000150) --> 0x8000013a, taken: 1, fsqID: 166, loop: 0, iter: 0, exit: 0
656676: system.cpu.branchPred: [c: 1972] 0x80000130-[0x80000134, 0x80000138) --> 0x8000013a, taken: 1, predEndPC: 0x80000150, isHit: 1, falseHit: 0
656676: system.cpu.branchPred: [c: 1972] FTB entry: valid 1, tag 0, fallThruAddr:0x80000150, slots:
656676: system.cpu.branchPred: [c: 1972] pc:0x80000134, size:4, target:0x8000013a, cond:1, indirect:0, call:0, return:0
656676: system.cpu.branchPred: [c: 1972] Predicted pc: 0x8000013a, upc: 0, npc(meaningless): 0x8000013e, instSeqNum: 2830
656676: system.cpu.branchPred: [c: 1972] Predict it taken to 0x8000013a
能看到第一次预测134 这个branch 时候，
ftq = 0x80000130 - [0, 0x80000150)
fsq = 0x80000130-[0, 0) --> 0, taken: 0, predEndPC: 0x80000150
ftb = 0x80000130 valid 1, tag 0, fallThruAddr:0x80000150, 有一个slot 但由于uftb预测NT, 也不跳
第二次预测134
ftq = 0x80000130 - [0x80000134, 0x80000150) --> 0x8000013a, 预测134 指令跳转到13a
fsq = 0x80000130-[0x80000134, 0x80000138) --> 0x8000013a， taken: 1, predEndPC: 0x80000150
内容和ftq基本一样
ftb = 0x80000130 valid 1, tag 0, fallThruAddr:0x80000150. 内部包含134 这个br 指令, 其预测taken了
如果观察第二条13e 这个branch 时候，比较有意思
655011: system.cpu.branchPred: [c: 1967] looking up pc 0x8000013e
655011: system.cpu.branchPred: [c: 1967] Supplying fetch with target ID 161
655011: system.cpu.branchPred: [c: 1967] Responsing fetch with:: 0x80000130 - [0, 0x80000150) --> 0, taken: 0, fsqID: 162, loop: 0, iter: 0, exit: 0
655011: system.cpu.branchPred: [c: 1967] 0x80000130-[0, 0) --> 0, taken: 0, predEndPC: 0x80000150, isHit: 1, falseHit: 0
655011: system.cpu.branchPred: [c: 1967] FTB entry: valid 1, tag 0, fallThruAddr:0x80000150, slots:
655011: system.cpu.branchPred: [c: 1967] pc:0x80000134, size:4, target:0x8000013a, cond:1, indirect:0, call:0, return:0
655011: system.cpu.branchPred: [c: 1967] Predict it not taken to 0x80000142
657009: system.cpu.branchPred: [c: 1973] looking up pc 0x8000013e
657009: system.cpu.branchPred: [c: 1973] Supplying fetch with target ID 166
657009: system.cpu.branchPred: [c: 1973] Responsing fetch with:: 0x8000013a - [0x8000013e, 0x8000015a) --> 0x80000144, taken: 1, fsqID: 167, loop: 0, iter: 0, exit: 0
657009: system.cpu.branchPred: [c: 1973] 0x8000013a-[0x8000013e, 0x80000142) --> 0x80000144, taken: 1, predEndPC: 0x8000015a, isHit: 1, falseHit: 0
657009: system.cpu.branchPred: [c: 1973] FTB entry: valid 1, tag 0, fallThruAddr:0x8000015a, slots:
657009: system.cpu.branchPred: [c: 1973] pc:0x8000013e, size:4, target:0x80000144, cond:1, indirect:0, call:0, return:0
657009: system.cpu.branchPred: [c: 1973] Predicted pc: 0x80000144, upc: 0, npc(meaningless): 0x80000148, instSeqNum: 2832
657009: system.cpu.branchPred: [c: 1973] Predict it taken to 0x80000144
第一次预测13e, 发现仍然使用130-150 的这个fsq, ftq, ftb, 甚至ftb 都不包含13e 这个br, 但是由于都是预测NT, 所以都没有预测其taken， 还是沿用老的ftb 项
第二次预测13e
ftq=0x8000013a - [0x8000013e, 0x8000015a) --> 0x80000144, fsq 类似
ftb= 0x8000013a FTB entry: valid 1, tag 0, fallThruAddr:0x8000015a, slots:
pc:0x8000013e, size:4, target:0x80000144, cond:1, indirect:0, call:0, return:0
会生成并使用一个新的ftb entry, 同时产生新的ftq, fsq 项！