http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49279
Summary: Optimization incorrectly presuming constant variable
inside loop in g++ 4.5 and 4.6 with -O2 and -O3 for
x86_64 targets
Product: gcc
Version: 4.6.0
Status: UNCONFIRMED
Severity: major
Priority: P3
Component: c++
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: tcmart...@gmail.com
Created attachment 24427
--> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24427
Testcase (reduced automatically using multidelta)
GCC is apparently producing the wrong code with Eigen2 (a template-based linear
algebra library) with optimization levels -O3 and -O2 for
x86_64-unknown-linux-gnu targets. A reduced test case is provided that
reproduces the error.
As I understand, the core of the problem is this loop (line 1132 of
the submitted test case):
for (; (ProcessFirstHalf ? i && i.index () < j : i); ++i) {
if (LhsIsSelfAdjoint) {
int a = LhsIsRowMajor ? j : i.index ();
int b = LhsIsRowMajor ? i.index () : j;
Scalar v = i.value ();
derived ().row (b) += ei_conj (v) * product.rhs ().row (a);
}
}
which is being translated into:
movq -8(%rsp), %rsi
movq (%rsi), %rbp
addq %rdx, %rbp
movsd 0(%rbp), %xmm1 # <- %xmm1 is initialized here and
.L5: # no longer touched!
leaq 0(,%rcx,8), %rsi
leaq 4(%r8,%rcx,4), %r8
movl %r9d, %ecx
jmp .L8
.L13: # <-Loop here!!!
movl (%r8), %r10d
addq $4, %r8
.L8:
movsd 0(%r13,%rsi), %xmm0
movslq %r10d, %r10
addl $1, %ecx
mulsd (%r12,%r10,8), %xmm0
cmpl %r11d, %ecx
addsd %xmm1, %xmm0
movsd %xmm0, 0(%rbp) # <- % shouldn't %xmm1 be updated here?
je .L3
addq $8, %rsi
cmpl %ecx, %r9d
jle .L13 # <- Loop ends
the sum operation on line
derived ().row (b) += ei_conj (v) * product.rhs ().row (a);
is apparently being performed by the instruction
addsd %xmm1, %xmm0
but the value of %xmm1 isn't being updated inside the loop!! Apparently the
compiler is presuming derived ().row (b) is constant inside the loop, which is
evidently *not* true. Since the value of %xmm1 is never updated, the
value of derived ().row (b) at the end of the loop is equal to the last
ei_conj (v) * product.rhs ().row (a) result.
The bug was verified on gcc versions 4.5.2 and 4.6.0 with -O2 and -O3 switches.
The compiler produces the correct code with -O0 and -O switches.
It is *NOT* present on the 4.4 branch (that is, 4.4 compiles the code
correctly) for -O0, -0, -02 and -O3 switches.
I suppose it is a regression of the 4.5 branch.
Command line (for gcc 4.6.0):
/opt/gnu/gcc-4.6/bin/g++ -v -save-temps -nostdinc -O3 testcase.a.cpp
Compiler output:
Using built-in specs.
COLLECT_GCC=/opt/gnu/gcc-4.6/bin/g++
COLLECT_LTO_WRAPPER=/opt/gnu/gcc-4.6/bin/../libexec/gcc/x86_64-unknown-linux-gnu/4.6.0/lto-wrapper
Target: x86_64-unknown-linux-gnu
Configured with: ../configure : (reconfigured) ../configure : (reconfigured)
../configure
Thread model: posix
gcc version 4.6.0 (GCC)
COLLECT_GCC_OPTIONS='-v' '-save-temps' '-nostdinc' '-O3' '-shared-libgcc'
'-mtune=generic' '-march=x86-64'
/opt/gnu/gcc-4.6/bin/../libexec/gcc/x86_64-unknown-linux-gnu/4.6.0/cc1plus -E
-quiet -nostdinc -v -iprefix
/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/ -D_GNU_SOURCE
testcase.a.cpp -mtune=generic -march=x86-64 -O3 -fpch-preprocess -o
testcase.a.ii
#include "..." search starts here:
#include <...> search starts here:
End of search list.
COLLECT_GCC_OPTIONS='-v' '-save-temps' '-nostdinc' '-O3' '-shared-libgcc'
'-mtune=generic' '-march=x86-64'
/opt/gnu/gcc-4.6/bin/../libexec/gcc/x86_64-unknown-linux-gnu/4.6.0/cc1plus
-fpreprocessed testcase.a.ii -quiet -dumpbase testcase.a.cpp -mtune=generic
-march=x86-64 -auxbase testcase.a -O3 -version -o testcase.a.s
GNU C++ (GCC) version 4.6.0 (x86_64-unknown-linux-gnu)
compiled by GNU C version 4.6.0, GMP version 4.3.2, MPFR version
3.0.0-p8, MPC version 0.9
GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
GNU C++ (GCC) version 4.6.0 (x86_64-unknown-linux-gnu)
compiled by GNU C version 4.6.0, GMP version 4.3.2, MPFR version
3.0.0-p8, MPC version 0.9
GGC heuristics: --param ggc-min-expand=100 --param ggc-min-heapsize=131072
Compiler executable checksum: 3f3899c46d47b31a2bc0cb7f3d1408a6
COLLECT_GCC_OPTIONS='-v' '-save-temps' '-nostdinc' '-O3' '-shared-libgcc'
'-mtune=generic' '-march=x86-64'
as --64 -o testcase.a.o testcase.a.s
COMPILER_PATH=/opt/gnu/gcc-4.6/bin/../libexec/gcc/x86_64-unknown-linux-gnu/4.6.0/:/opt/gnu/gcc-4.6/bin/../libexec/gcc/
LIBRARY_PATH=/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/:/opt/gnu/gcc-4.6/bin/../lib/gcc/:/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/../../../../lib64/:/lib/../lib64/:/usr/lib/../lib64/:/usr/lib/nvidia-current/:/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/../../../:/lib/:/usr/lib/
COLLECT_GCC_OPTIONS='-v' '-save-temps' '-nostdinc' '-O3' '-shared-libgcc'
'-mtune=generic' '-march=x86-64'
/opt/gnu/gcc-4.6/bin/../libexec/gcc/x86_64-unknown-linux-gnu/4.6.0/collect2
--eh-frame-hdr -m elf_x86_64 -dynamic-linker /lib64/ld-linux-x86-64.so.2
/usr/lib/../lib64/crt1.o /usr/lib/../lib64/crti.o
/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/crtbegin.o
-L/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0
-L/opt/gnu/gcc-4.6/bin/../lib/gcc
-L/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/../../../../lib64
-L/lib/../lib64 -L/usr/lib/../lib64 -L/usr/lib/nvidia-current
-L/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/../../..
testcase.a.o -lstdc++ -lm -lgcc_s -lgcc -lc -lgcc_s -lgcc
/opt/gnu/gcc-4.6/bin/../lib/gcc/x86_64-unknown-linux-gnu/4.6.0/crtend.o
/usr/lib/../lib64/crtn.o
A test case was produced with the preprocessed output (generated from Eigen
version 2.0.15) and automatically reduced using the multidelta tool. Testcase
included.
