https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86669
Bug ID: 86669
Summary: [5/6/7/8/9 regression] Complete object constructor
clone omits length for a c++11 braced initialiser
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: iains at gcc dot gnu.org
Target Milestone: ---
Created attachment 44435
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=44435&action=edit
pre-processed and reduced code
This applies to targets that do not allow (or have) alias support for
constructor clones.
c-reduce has obfuscated the variable names in the reduced code, of course ...
The original code looks like this:
====
AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
: LegalizerInfo() {
const LLT p0 = LLT::pointer(0, 64);
const LLT s1 = LLT::scalar(1);
const LLT s8 = LLT::scalar(8);
const LLT s16 = LLT::scalar(16);
const LLT s32 = LLT::scalar(32);
const LLT s64 = LLT::scalar(64);
const LLT s128 = LLT::scalar(128);
const LLT v2s32 = LLT::vector(2, 32);
const LLT v4s32 = LLT::vector(4, 32);
const LLT v2s64 = LLT::vector(2, 64);
for (auto Ty : {p0, s1, s8, s16, s32, s64})
setAction({/*G_IMPLICIT_DEF*/41, Ty}, Legal);
...
}
=====
The gimple for the base constructor contains:
"
try
{
D.2730 = {};
D.2730.i = 6; <<<<< OK _M_len is set
D.2729[0] = e;
D.2729[1] = f;
D.2729[2] = g;
D.2729[3] = h;
D.2729[4] = i;
D.2729[5] = j;
D.2730.a = &D.2729;
__for_range = &D.2730;
__for_begin = std::initializer_list<llvm::J>::begin
(__for_range);
__for_end = std::initializer_list<llvm::J>::end
(__for_range);
"
but for the complete constructor:
"
try
{
D.2745 = {}; <<< _M_len init missing
D.2744[0] = e;
D.2744[1] = f;
D.2744[2] = g;
D.2744[3] = h;
D.2744[4] = i;
D.2744[5] = j;
D.2745.a = &D.2744;
__for_range = &D.2745;
__for_begin = std::initializer_list<llvm::J>::begin
(__for_range);
__for_end = std::initializer_list<llvm::J>::end
(__for_range);
<D.2808>:
"
========
With clang
.../bin/clang -cc1 -triple x86_64-apple-macosx10.12.0 -x c++-cpp-output
AArch64LegalizerInfo-fmt.ii -O0 -S -std=c++11 -fvisibility-inlines-hidden
-fno-rtti -o tc.s
The complete constructor is rendered as a thunk calling the base constructor
(not sure why we conclude this is not viable in GCC)
======
With GCC (all versions I've tried 5...trunk) we get the effect above.
.../libexec/gcc/x86_64-apple-darwin14/7.3.0/cc1plus -fpreprocessed
AArch64LegalizerInfo-fmt.ii -fPIC -quiet -mmacosx-version-min=10.10
-mtune=core2 -g3 -O0 -Werror=date-time -Wall -Wextra -Wno-unused-parameter
-Wwrite-strings -Wcast-qual -Wno-missing-field-initializers -Wpedantic
-Wno-long-long -Wno-maybe-uninitialized -Wdelete-non-virtual-dtor -Wno-comment
-std=c++11 -fPIC -fvisibility-inlines-hidden -fno-exceptions -fno-rtti -o tg.s
======
I took a look at what was happening in cp/optimize.c:maybe_clone_body() but
couldn't see any special difference between the operations carried out to clone
the base and the complete constructors.
