Quuxplusone / LLVMBugzillaTest

0 stars 0 forks source link

Worse zero struct init with = T{} than with memset #46037

Open Quuxplusone opened 4 years ago

Quuxplusone commented 4 years ago
Bugzilla Link PR47068
Status NEW
Importance P enhancement
Reported by David Bolvansky (david.bolvansky@gmail.com)
Reported on 2020-08-09 18:28:30 -0700
Last modified on 2020-08-10 01:51:40 -0700
Version trunk
Hardware PC Linux
CC efriedma@quicinc.com, llvm-bugs@lists.llvm.org, richard-llvm@metafoo.co.uk, spatel+llvm@rotateright.com
Fixed by commit(s)
Attachments
Blocks
Blocked by
See also
class pt {
  int x;
  int y;
};

class pt2 {
  int x;
  char y;
};

void foo(pt* s) {
  *s = {};
}

void bar(pt2* s) {
  *s = {};
}

For foo case, codegen is fine. The padding is problematic here.

Clang:
        ret
bar(pt2*):
        mov     DWORD PTR [rdi], 0
        mov     BYTE PTR [rdi+4], 0
        ret

ICC:

bar(pt2*):
        xor       eax, eax                                      #18.4
        mov       QWORD PTR [rdi], rax                          #18.8
        ret

So ideally we should have:
        mov       QWORD PTR [rdi], 0

define dso_local void @_Z3fooP2pt(%class.pt* nocapture %0) local_unnamed_addr
#0 {
  %2 = bitcast %class.pt* %0 to i64*
  store i64 0, i64* %2, align 4
  ret void
}

define dso_local void @_Z3barP3pt2(%class.pt2* nocapture %0) local_unnamed_addr
#1 {
  %2 = bitcast %class.pt2* %0 to i40*
  store i40 0, i40* %2, align 4, !tbaa.struct !2
  ret void
}

Looking at dumps, SROA to blame?

--------------------------------------------------

With
class pt3 {
  int x;
  int y;
  char z;
};

Unoptimized:
define dso_local void @_Z3bazP3pt3(%class.pt3* %0) #0 {
  %2 = alloca %class.pt3*, align 8
  %3 = alloca %class.pt3, align 4
  store %class.pt3* %0, %class.pt3** %2, align 8
  %4 = bitcast %class.pt3* %3 to i8*
  call void @llvm.memset.p0i8.i64(i8* align 4 %4, i8 0, i64 12, i1 false)
  %5 = load %class.pt3*, %class.pt3** %2, align 8
  %6 = bitcast %class.pt3* %5 to i8*
  %7 = bitcast %class.pt3* %3 to i8*
  call void @llvm.memcpy.p0i8.p0i8.i64(i8* align 4 %6, i8* align 4 %7, i64 9, i1 false)
  ret void
}

Optimized:
define dso_local void @_Z3bazP3pt3(%class.pt3* nocapture %0) local_unnamed_addr
#2 {
  %2 = bitcast %class.pt3* %0 to i8*
  call void @llvm.memset.p0i8.i64(i8* nonnull align 4 dereferenceable(9) %2, i8 0, i64 9, i1 false)
  ret void
}

In this case, we are doing bad job when combing memset and memcpy in
MemCpyOptimizer. It should be:
call void @llvm.memset.p0i8.i64(i8* nonnull align 4 dereferenceable(12) %2, i8
0, i64 12, i1 false)
Quuxplusone commented 4 years ago

It isn't obvious to me that this transform is valid, at first glance. pt2 could have subclasses that store data in the padding that's overwritten by icc.