owenpan added inline comments.
================
Comment at: clang/lib/Format/UnwrappedLineFormatter.cpp:215
+ const auto &NextLine = *I[1];
+ const auto *PreviousLine = I != AnnotatedLines.begin() ? I[-1] : nullptr;
+ if (NextLine.Type == LT_Invalid || NextLine.First->MustBreakBefore)
----------------
owenpan wrote:
> HazardyKnusperkeks wrote:
> > owenpan wrote:
> > > HazardyKnusperkeks wrote:
> > > > owenpan wrote:
> > > > > I would move this line to just before handling empty record blocks
> > > > > below.
> > > > I'd rather keep the definitions close together.
> > > If it were just a simple initialization, it wouldn't matter much.
> > > However, it would be a bit wasteful as `PreviousLine` always gets
> > > computed here even if the function may return before the pointer would
> > > get chance to be used. If you really want to keep all related definitions
> > > together, wouldn't you want to move lines 214-215 up to right after line
> > > 211?
> > In a release build I'm betting that the compiler is smart enough to never
> > calculate `PreviousLine` and that performance doesn't really matter was
> > shown in D116318.
> >
> > But yeah moving it up to `TheLine` is consistent and will do.
> > In a release build I'm betting that the compiler is smart enough to never
> > calculate `PreviousLine`
>
> No compiler will (or should) skip generating code that calculates
> `PreviousLine`, which may or may not get used at runtime. I'm not aware of
> any compiler that is smart enough to move the initialization code to just
> before where the variable is first used.
>
> To illustrate the difference, I compiled the code below with `clang++ -Wall
> -std=c++11 -O3 -S foo.cpp`:
> ```
> #include <vector>
>
> using namespace std;
> using Type = vector<int *>;
>
> int f(const Type &V, Type::const_iterator I) {
> const auto *P = I != V.begin() ? I[-1] : nullptr;
> if (I == V.end())
> return 0;
> return *P;
> }
> ```
> The generated code in foo.s includes the following:
> ```
> .cfi_startproc
> ; %bb.0:
> ldr x8, [x0]
> cmp x8, x1
> b.eq LBB0_3
> ; %bb.1:
> ldur x8, [x1, #-8]
> ldr x9, [x0, #8]
> cmp x9, x1
> b.eq LBB0_4
> LBB0_2:
> ldr w0, [x8]
> ret
> LBB0_3:
> mov x8, #0
> ldr x9, [x0, #8]
> cmp x9, x1
> b.ne LBB0_2
> LBB0_4:
> mov w0, #0
> ret
> .cfi_endproc
> ```
> As you can see, the initialization code was neither optimized away nor moved.
>
> When I changed the function body to:
> ```
> if (I == V.end())
> return 0;
> const auto *P = I != V.begin() ? I[-1] : nullptr;
> return *P;
> ```
> The generated code became much simpler:
> ```
> .cfi_startproc
> ; %bb.0:
> ldr x8, [x0, #8]
> cmp x8, x1
> b.eq LBB0_2
> ; %bb.1:
> ldur x8, [x1, #-8]
> ldr w0, [x8]
> ret
> LBB0_2:
> mov w0, #0
> ret
> .cfi_endproc
> ```
> and that performance doesn't really matter was shown in D116318.
It wasn't quite the same there as we must pay either in the caller or in the
callee. That being said, I agree that the performance hit here would be
negligible.
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D115060/new/
https://reviews.llvm.org/D115060
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
