-O2 -fplugin-arg-dragonegg-enable-gcc-optzns fails to link induct.f90

[llvmbot]

Bugzilla Link	9692
Resolution	FIXED
Resolved on	May 25, 2011 16:45
Version	trunk
OS	MacOS X
Reporter	LLVM Bugzilla Contributor

Extended Description

Dragonegg svn fails to link induct.f90 on x86_64 darwin when using -O2 -fplugin-arg-dragonegg-enable-gcc-optzns with the error...

[MacPro:~/de_bug4] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -O2 -fplugin-arg-dragonegg-enable-gcc-optzns induct.f90 -o induct
Undefined symbols for architecture x86_64:
"_cexpi", referenced from:
___mcc_m_MOD_mutual_ind_cir_cir_coils in ccmceubx.o
___mcr_m_MOD_mutual_ind_cir_rec_coils in ccmceubx.o
___mqc_m_MOD_mutual_ind_quad_cir_coil in ccmceubx.o
induct in ccmceubx.o
ld: symbol(s) not found for architecture x86_64
collect2: ld returned 1 exit status

which shows...

[MacPro:~/de_bug4] howarth% nm induct.* | grep cexpi
U _cexpi

when -fplugin-arg-dragonegg-enable-gcc-optzns is not used the induct.o contains no symbols
for _cexpi.

[llvmbot]

Error doesn't occur at -O0 but does at -01 ,-O2 and -Os with -fplugin-arg-dragonegg-enable-gcc-optzns.

[llvmbot]

reduced testcase of undefined cexp symbol generation

[llvmbot]

assembly from reduced testcase compiled with -O0 -fplugin-arg-dragonegg-enable-gcc-optzns

[llvmbot]

assembly from reduced testcase compiled with -O1 -fplugin-arg-dragonegg-enable-gcc-optzns

[llvmbot]

Reduced testcase containing mutual_ind_cir_cir_coils demonstrates the bug as follows...

[MacPro:pb05/lin/source] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -O0 -fplugin-arg-dragonegg-enable-gcc-optzns -c reduced.f90 --save-temps -fverbose-asm
[MacPro:pb05/lin/source] howarth% grep cexp reduced.s
[MacPro:pb05/lin/source] howarth%

vs

[MacPro:pb05/lin/source] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -O1 -fplugin-arg-dragonegg-enable-gcc-optzns -c reduced.f90 --save-temps -fverbose-asm
[MacPro:pb05/lin/source] howarth% grep cexp reduced.s callq _cexpi
callq _cexpi

Note that this undefined cexpi symbol isn't generated without -fplugin-arg-dragonegg-enable-gcc-optzns at -O1 and higher...

[MacPro:pb05/lin/source] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -O1 -c reduced.f90 --save-temps -fverbose-asm[MacPro:pb05/lin/source] howarth% grep cexp reduced.s [MacPro:pb05/lin/source] howarth%

[llvmbot]

At r129460, now that bug 9664 is fixed, additional Polyhedron 2005 benchmarks are exhibiting the undefined symbols with -O2 -fplugin-arg-dragonegg-enable-gcc-optzns...

[MacPro:pb05/lin/source] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -msse3 -O2 -fplugin-arg-dragonegg-enable-gcc-optzns fatigue.f90 -o fatigue
Undefined symbols for architecture x86_64:
"cexpi", referenced from:
MAIN in ccvKQYUi.o
ld: symbol(s) not found for architecture x86_64
collect2: ld returned 1 exit status

[MacPro:pb05/lin/source] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -msse3 -O2 -fplugin-arg-dragonegg-enable-gcc-optzns aermod.f90 -o aermod
Undefined symbols for architecture x86_64:
"_cexpif", referenced from:
pcalc in ccCLQBmM.o
evloc in ccCLQBmM.o
metini in ccCLQBmM.o
piteff in ccCLQBmM.o
deltah in ccCLQBmM.o
sblris in ccCLQBmM.o
setpol in ccCLQBmM.o
...
ld: symbol(s) not found for architecture x86_64
collect2: ld returned 1 exit status

This is using dragonegg trunk svn with llvm 2.9. I see the same failure at -m64 and -m32 on x86_64-apple-darwin10.

[llvmbot]

Oddly...

[MacPro:pb05/lin/source] howarth% /sw/lib/gcc4.5/bin/gfortran -fplugin=/sw/lib/gcc4.5/lib/dragonegg.so -O0 -fauto-inc-dec -fcprop-registers -fdce -fdefer-pop -fdse -fguess-branch-probability -fif-conversion2 -fif-conversion -fipa-pure-const -fipa-reference -fmerge-constants -fsplit-wide-types -ftree-builtin-call-dce -ftree-ccp -ftree-ch -ftree-copyrename -ftree-dce -ftree-dominator-opts -ftree-dse -ftree-forwprop -ftree-fre -ftree-phiprop -ftree-sra -ftree-pta -ftree-ter -funit-at-a-time -fplugin-arg-dragonegg-enable-gcc-optzns fatigue.f90 -o fatigue

doesn't trigger the problem so the offending optimization doesn't' appear to be under control of a specific flag (except -O1 or higher). Is this issue darwin specific?

[llvmbot]

This is not darwin specific. GCC turns sincos or cexp into the cexpi builtin in
some circumstances. This needs to be turned back into a call to sincos or cexp
by dragonegg.

[llvmbot]

*** Bug llvm/llvm-bugzilla-archive#9754 has been marked as a duplicate of this bug. ***

[llvmbot]

This is not darwin specific. GCC turns sincos or cexp into the cexpi builtin
in
some circumstances. This needs to be turned back into a call to sincos or cexp
by dragonegg.

I found this builtin is defined in gcc/builtins.def of gcc source:

DEF_GCC_BUILTIN (BUILT_IN_CEXPI, "cexpi", BT_FN_COMPLEX_DOUBLE_DOUBLE, ATTR_MATHFN_FPROUNDING)

There are many other builtins there, I suppose the similar issues will occur with all symbols prefixed DEF_GCC_BUILTIN. Maybe it would be better to switch off their generation on frontend side (if possible) ?

[llvmbot]

Most gcc builtins directly correspond to a library function, eg: builtin_sinf ->
sinf. These are all handled automagically, so the number of builtins that need
special support is not as great as it might first seem.

[llvmbot]

Any plans on how this will be handled? Since it impacts -O1 and above, this blocks the use of -fplugin-arg-dragonegg-enable-gcc-optzns entirely for parts of the pb05 test suite.

[llvmbot]

I have moved this higher on my list of things to do :)

[llvmbot]

Fixed here:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20110523/121435.html

[llvmbot]

mentioned in issue llvm/llvm-bugzilla-archive#9754