Basic block in LLVM IR seems broken in assembly codes - llvm

Question

I add debug info to the first and last instruction of the basic blocks by llvm pass, and then I successfully find the info I add in the assembly codes. But the numbers of info of first instruction and last instruction are different.
I want to know if the result is correct, and if it is correct, why the numbers are different and how to correctly get the boundary of block in assembly code?
The code of my pass
bool runOnFunction(Function &F) override {
unsigned start = 2333;
unsigned end = 23333;
MDNode *N = F.getMetadata("dbg");
for (BasicBlock &B : F) {
errs() << "Hello: ";
Instruction *I = B.getFirstNonPHI();
DebugLoc loc = DebugLoc::get(start, 0, N);
if (loc && I!=NULL) {
I->setDebugLoc(loc);
} else {
errs() << "start error";
}
I = B.getTerminator();
loc = DebugLoc::get(end, 1, N);
if (loc && I!= NULL) {
I->setDebugLoc(loc);
} else {
errs() << "end error";
}
errs() << "\n";
}
return true;
}
};
}
It runs without any errors. Some of the result:
# %bb.2: # %for.body
# in Loop: Header=BB0_1 Depth=1
.loc 1 2333 0 # myls.c:2333:0
movl -4(%ebp), %eax
.Ltmp4:
.loc 1 71 14 # myls.c:71:14
movl %eax, -8(%ebp)
.Ltmp5:
.LBB0_3: # %for.cond1
# Parent Loop BB0_1 Depth=1
# => This Inner Loop Header: Depth=2
.loc 1 2333 0 # myls.c:2333:0
movl -8(%ebp), %eax
.Ltmp6:
.loc 1 71 18 # myls.c:71:18
cmpl n, %eax
.Ltmp7:
.loc 1 23333 1 # myls.c:23333:1
jge .LBB0_13
......
LBB2_10: # %sw.epilog
.loc 1 23333 1 # myls.c:23333:1
jmp .LBB2_11
.LBB2_11: # %while.cond
# =>This Inner Loop Header: Depth=1
.loc 1 162 13 # myls.c:162:13
cmpl $0, -164(%ebp)
.loc 1 23333 1 # myls.c:23333:1
jl .LBB2_20
# %bb.12: # %while.body
# in Loop: Header=BB2_11 Depth=1
.Ltmp75:
.loc 1 164 14 # myls.c:164:14
movl -136(%ebp), %eax
.loc 1 164 28 is_stmt 0 # myls.c:164:28
movl -164(%ebp), %ecx
# kill: def $cl killed $ecx
.loc 1 164 25 # myls.c:164:25
movl $1, %edx
shll %cl, %edx
.loc 1 164 22 # myls.c:164:22
andl %edx, %eax
cmpl $0, %eax
.Ltmp76:
.loc 1 23333 1 is_stmt 1 # myls.c:23333:1
je .LBB2_18
I find the 2333 and 23333 don't match, and the numbers of 2333 and 23333 are different in assembly code for different architectures. I use opt to run my pass and llc to get the assembly code.
I appreciate every help.

Answer 1

I think the problem is with optimization. After your pass, the IR goes through other rounds of optimization that changed the basic blocks and eliminated some instructions.
You probably have a register function like the following from LLVM page
static llvm::RegisterStandardPasses Y(
llvm::PassManagerBuilder::EP_EarlyAsPossible,
[](const llvm::PassManagerBuilder &Builder,
llvm::legacy::PassManagerBase &PM) { PM.add(new Hello()); });
Try use EP_OptimizerLast rather than EP_EarlyAsPossible. That will run your pass as the last optimizer.
The other option is to use EP_EnabledOnOptLevel0 and use opt with -O0.
You can look up the flags in this page
It might be helpful to use -emit-llvm to generate LLVM IR in *.ll. It is more visible what your code has done to the IR