problem using -ipo linking with ICC object files

[maverick6664]

I've been using ICC to build seti@home for a while and think I know -ipo with ICC well.

Now I'm building scilab (available www.scilab.org) with Intel Fortran and C++ compilers. I can build it with -ip flag with IFORT/ICC, but when I change -ip flag(s) to -ipo, linking fails with

ld: /tmp/ipo_ifortUx2MaC.o: No such file: No such file or directory

error. Does it mean IPO with mixed environment of IFORT/ICC is impossible? My options for ICC/IFORT are almost the same

FC_OPTIONS = -O3 -xN -ipo -parallel -cm -w90 -w95 -mp -pc64 -ffnalias -falias

and I use xiar instead of ar. During linking, I can see 5 assembler source files are created in /tmp, but no object files are found.

In /tmp/ifort*arg file in another session, I see

maverick:~% cat /tmp/ifort*arg
-mP1OPT_version=900
-mGLOB_source_language=GLOB_SOURCE_LANGUAGE_F90
-mGLOB_tune_for_fort
-mGLOB_use_fort_dope_vector
-mP2OPT_static_promotion
-mP1OPT_print_version=FALSE
-mP3OPT_use_mspp_call_convention
-mCG_use_gas_got_workaround=F
-mP2OPT_align_option_used=TRUE
-mGLOB_options_string=-O3 -xN -ipo -parallel -cm -w90 -w95 -mp -pc64 -ffnalias -falias -o bin/scilex -lieee -ldl -i-static -L/usr/lib -ltk8.4 -L/usr/lib -ltcl8.4 -ldl -lXaw3d -lXmu -lXt -lXext -lSM -lICE -lX11 -lncurses
-mGLOB_cxx_limited_range=FALSE
-mGLOB_as_output_backup_file_name=/tmp/ifortcha2YIas_.s
-mP1OPT_fpsetmask_arg=0xffff127f
-mGLOB_machine_model=GLOB_MACHINE_MODEL_PW
-mGLOB_maintain_precision
-mGLOB_precision_mask=0xffffffff
-mGLOB_extended_instructions=0x40
-mP2OPT_subs_out_of_bound=FALSE
-mIPOPT_ninl_user_level=2
-mGLOB_opt_level=3
-mIPOPT_activate
-mP2OPT_hlo_level=3
-mP2OPT_hlo
-mPAROPT_concurrentize
-mIPOPT_link
-mIPOPT_ipo_activate
-mIPOPT_ipo_mo_activate
-mIPOPT_ipo_mo_nfiles=1
-mIPOPT_source_files_list=/tmp/ifortNybIDRlst
-mIPOPT_short_data_info=/tmp/ifort4u3RrPsdata
-mIPOPT_link_script_file=/tmp/ifortX4P3fNscript
-mIPOPT_global_data
-mIPOPT_link_version=2.16.1
...snip....

I have no idea what to do... Is it too large? I don't think so.

BTW I found -xP flag for Intel Fortran seemed to generate faulty code with scilab (while -xP with ICC has no problem) and reported to Premier Support, but it's another story.