Open MPI logo

Open MPI Development Mailing List Archives

  |   Home   |   Support   |   FAQ   |   all Development mailing list

Subject: Re: [OMPI devel] amd64 atomic.h warnings
From: Jeff Squyres (jsquyres_at_[hidden])
Date: 2010-06-08 12:01:53


We talked about this on the phone on the call today. I filed https://svn.open-mpi.org/trac/ompi/ticket/2437 about what we plan to do.

On Jun 8, 2010, at 11:40 AM, George Bosilca wrote:

>
> On Jun 8, 2010, at 14:49 , Jeff Squyres wrote:
>
> > ## -cmdline +pgcc george2.c -O3 -c -S -x 123 4 -x 123 0x80000000 -alwaysinline /opt/pgi/7.0.7/linux86-64/7.0-7/lib/libintrinsics.il 4
> > ## -asm george2.s
> > ## lineno: 7
> > .text
> > .align 16
> > opal_atomic_cmpset_32:
> > ..Dcfb0:
> > pushq %rbp
> > ..Dcfi0:
> > movq %rsp, %rbp
> > ..Dcfi1:
> > movl %esi, -4(%rbp)
> > movl %edx, -8(%rbp)
> > ..EN1:
> > ## lineno: 16
> > movl -8(%rbp), %edx
> > movl -4(%rbp), %eax
>
> oldval is moved into the %eax ... once
>
> > lock;cmpxchgl %edx,(%rdi)
> > sete %cl
>
> The CCR is retrieved
>
> > movb %cl, -9(%rbp)
>
> And stored.
>
> But the loop disappeared. I really have a doubt about the correctness of this assembly code.
>
> george.
>
> > ## lineno: 17
> > movzbl -9(%rbp), %eax
> > ## lineno: 0
> > popq %rbp
> > ret
> > .type opal_atomic_cmpset_32,@function
> > .size opal_atomic_cmpset_32,.-opal_atomic_cmpset_32
> > ..Dcfe0:
> > __opal_atomic_cmpset_32END:
> > .section .pgi_trace
> > .align 8
> > .quad opal_atomic_cmpset_32 ## address of routine
> > .quad __opal_atomic_cmpset_32END - opal_atomic_cmpset_32 ## size of routine
> > .2byte 0 ## flags for future use
> > .2byte 21 ## length of following string
> > ## name:opal_atomic_cmpset_32:
> > .byte 0x6f,0x70,0x61,0x6c,0x5f,0x61,0x74,0x6f,0x6d,0x69,0x63
> > .byte 0x5f,0x63,0x6d,0x70,0x73,0x65,0x74,0x5f,0x33,0x32,0x00
> > ## lineno: 20
> > .text
> > .align 16
> > .globl main
> > main:
> > ..Dcfb1:
> > pushq %rbp
> > ..Dcfi2:
> > movq %rsp, %rbp
> > ..Dcfi3:
> > subq $16, %rsp
> > movq %rbx, -16(%rbp)
> > pushq %rax
> > pushq %rax
> > stmxcsr (%rsp)
> > popq %rax
> > orq $64, %rax
> > pushq %rax
> > ldmxcsr (%rsp)
> > popq %rax
> > popq %rax
> > ## lineno: 0
> > ..EN2:
> > ## lineno: 21
> > movl $0, -4(%rbp)
> > .align 8
> > .LB191:
> > ## lineno: 24
> > movl -4(%rbp), %ebx
> > leaq -4(%rbp), %rdi
> > leal 1(%rbx), %edx
> > movl %ebx, %esi
> > call opal_atomic_cmpset_32
> > testl %eax, %eax
> > je .LB191
> > ## lineno: 27
> > addl $1, %ebx
> > movl %ebx, %eax
> > ## lineno: 28
> > movq -16(%rbp), %rbx
> > leave
> > ret
> > .type main,@function
> > .size main,.-main
> > ..Dcfe1:
> > __mainEND:
> > .section .pgi_trace
> > .align 8
> > .quad main ## address of routine
> > .quad __mainEND - main ## size of routine
> > .2byte 0 ## flags for future use
> > .2byte 4 ## length of following string
> > ## name:main:
> > .byte 0x6d,0x61,0x69,0x6e,0x00
> > .data
> > .section .debug_frame
> > ..Dcieb0:
> > .4byte ..Dciee0-..Dcieb0-4 ## CIE length
> > .4byte 0xffffffff ## CIE ID
> > .byte 0x1 ## CIE version
> > .byte 0x0 ## no augmentation
> > .byte 0x1 ## ULEB128 1, code alignment factor
> > .byte 0x78 ## SLEB128 -8, data alignment factor
> > .byte 0x10 ## return address column
> > .byte 0xc ## DW_CFA_def_cfa (col 7)
> > .byte 0x7 ## ULEB128 7
> > .byte 0x8 ## ULEB128 8
> > .byte 0x90 ## DW_CFA_offset (col 16)
> > .byte 0x1 ## ULEB128 1
> > .align 8
> > ..Dciee0:
> > .4byte ..Dfdee0-..Dfdeb0 ## FDE length
> > ..Dfdeb0:
> > .4byte ..Dcieb0 ## CIE pointer
> > .quad ..Dcfb0 ## initial location
> > .quad ..Dcfe0-..Dcfb0 ## address range
> > .byte 0x4 ## DW_CFA_advance_loc4
> > .4byte ..Dcfi0-..Dcfb0
> > .byte 0xe ## DW_CFA_def_cfa_offset
> > .byte 0x10 ## ULEB128 16
> > .byte 0x86 ## DW_CFA_offset (col 6)
> > .byte 0x2 ## ULEB128 2
> > .byte 0x4 ## DW_CFA_advance_loc4
> > .4byte ..Dcfi1-..Dcfi0
> > .byte 0xd ## DW_CFA_def_cfa_register (col 6)
> > .byte 0x6 ## ULEB128 6
> > .align 8
> > ..Dfdee0:
> > .4byte ..Dfdee1-..Dfdeb1 ## FDE length
> > ..Dfdeb1:
> > .4byte ..Dcieb0 ## CIE pointer
> > .quad ..Dcfb1 ## initial location
> > .quad ..Dcfe1-..Dcfb1 ## address range
> > .byte 0x4 ## DW_CFA_advance_loc4
> > .4byte ..Dcfi2-..Dcfb1
> > .byte 0xe ## DW_CFA_def_cfa_offset
> > .byte 0x10 ## ULEB128 16
> > .byte 0x86 ## DW_CFA_offset (col 6)
> > .byte 0x2 ## ULEB128 2
> > .byte 0x4 ## DW_CFA_advance_loc4
> > .4byte ..Dcfi3-..Dcfi2
> > .byte 0xd ## DW_CFA_def_cfa_register (col 6)
> > .byte 0x6 ## ULEB128 6
> > .align 8
> > ..Dfdee1:
> > .ident "PGC 7.0-7"
> > [7:49] svbu-mpi:~/tmp %
> >
> > -----
> >
> >
> >
> > On Jun 8, 2010, at 10:46 AM, George Bosilca wrote:
> >
> >> It didn't work. Let's try with this small complete application:
> >>
> >> #include <stdint.h>
> >>
> >> #define SMPLOCK "lock;"
> >>
> >> static inline int opal_atomic_cmpset_32( volatile int32_t *addr,
> >> int32_t oldval, int32_t newval)
> >> {
> >> unsigned char ret;
> >> __asm__ __volatile__ (
> >> SMPLOCK "cmpxchgl %1,%2 \n\t"
> >> "sete %0 \n\t"
> >> : "=qm" (ret)
> >> : "q"(newval), "m"(*addr), "a"(oldval)
> >> : "memory");
> >>
> >> return (int)ret;
> >> }
> >>
> >> int main(int argc, char* argv[] )
> >> {
> >> int32_t value = 0, oldval = 0, delta = 1;
> >> int32_t* addr = &value;
> >>
> >> do {
> >> oldval = *addr;
> >> } while (0 == opal_atomic_cmpset_32(addr, oldval, oldval + delta));
> >> return (oldval + delta);
> >> }
> >>
> >>
> >>
> >> Thanks,
> >> george.
> >>
> >>
> >> On Jun 8, 2010, at 14:42 , Jeff Squyres wrote:
> >>
> >>> Look at my output -- I did...
> >>>
> >>> On Jun 8, 2010, at 10:40 AM, George Bosilca wrote:
> >>>
> >>>> Still no good, the opal_atomic_cmpset_32 is not inlined. Try to add -O3 to your command line, this helped for gcc.
> >>>>
> >>>> Thanks,
> >>>> george.
> >>>>
> >>>> On Jun 8, 2010, at 14:14 , Jeff Squyres wrote:
> >>>>
> >>>>> On Jun 8, 2010, at 9:53 AM, George Bosilca wrote:
> >>>>>
> >>>>>> As you can see there is no explicit call, the opal_atomic_cmpset_32 is really inlined. I think the problem is that you didn't specify the -O3 flag on your command line.
> >>>>>
> >>>>> Ah, you wanted me to compile the OMPI code itself and send you the assembly. That's not what you asked for. :-)
> >>>>>
> >>>>> (I just took the code you sent in the mail, stuffed it into george.c, and compiled that with -s -- outside of the context of the Open MPI code tree)
> >>>>>
> >>>>> Here's the new output. It still didn't inline, but you can see the code for the _cmpset function:
> >>>>>
> >>>>> -----
> >>>>> [7:13] svbu-mpi:~/tmp % cat george.c
> >>>>> #include <stdint.h>
> >>>>>
> >>>>> #include "opal/sys/atomic.h"
> >>>>>
> >>>>> int foo(void) {
> >>>>> int32_t oldval, delta;
> >>>>> int32_t *addr = 0;
> >>>>> do {
> >>>>> oldval = *addr;
> >>>>> } while (0 == opal_atomic_cmpset_32(addr, oldval, oldval + delta));
> >>>>> return (oldval + delta);
> >>>>> }
> >>>>>
> >>>>> [7:13] svbu-mpi:~/tmp % pgcc -O3 -I /home/jsquyres/svn/ompi4 -I/home/jsquyres/svn/ompi4/opal/include -c -s george.c
> >>>>> [7:13] svbu-mpi:~/tmp % cat george.s .file "george.c"
> >>>>> .version "01.01"
> >>>>> ## PGC 7.0 -opt 1
> >>>>> ## PGC 06/08/2010 05:10:04
> >>>>> ## pgcc george.c -c -S
> >>>>> ## /opt/pgi/7.0.7/linux86-64/7.0-7/bin/pgc
> >>>>> ## george.c -opt 1 -terse 1 -inform warn -x 119 0xa10000 -x 122 0x40 -x 123 0x1000
> >>>>> ## -x 127 4 -x 127 16 -x 19 0x400000 -x 28 0x40000 -x 70 0x8000 -x 122 1 -quad
> >>>>> ## -x 59 4 -x 59 4 -tp p7-64 -astype 0 -stdinc /opt/pgi/7.0.7/linux86-64/7.0-7/include:/usr/local/include:/usr/lib/gcc/x86_64-redhat-linux/4.1.2/include:/usr/lib/gcc/x86_64-redhat-linux/4.1.2/include:/usr/include
> >>>>> ## -def unix -def __unix -def __unix__ -def linux -def __linux -def __linux__
> >>>>> ## -def __NO_MATH_INLINES -def __x86_64__ -def __LONG_MAX__=9223372036854775807L
> >>>>> ## -def __SIZE_TYPE__=unsigned long int -def __PTRDIFF_TYPE__=long int -def __THROW=
> >>>>> ## -def __extension__= -def __amd64__ -def __SSE__ -def __MMX__ -def __SSE2__
> >>>>> ## -def __SSE3__ -predicate #machine(x86_64) #lint(off) #system(posix) #cpu(x86_64)
> >>>>> ## -cmdline +pgcc george.c -c -S -x 123 4 -x 123 0x80000000 -alwaysinline /opt/pgi/7.0.7/linux86-64/7.0-7/lib/libintrinsics.il 4
> >>>>> ## -asm george.s
> >>>>> ## lineno: 3
> >>>>> .text
> >>>>> .align 16
> >>>>> .globl foo
> >>>>> foo:
> >>>>> ..Dcfb0:
> >>>>> pushq %rbp
> >>>>> ..Dcfi0:
> >>>>> movq %rsp, %rbp
> >>>>> ..Dcfi1:
> >>>>> subq $16, %rsp
> >>>>> ..EN1:
> >>>>> ## lineno: 5
> >>>>> movq $0, -8(%rbp)
> >>>>> .p2align 4,,3
> >>>>> .LB157:
> >>>>> ## lineno: 6
> >>>>> movq -8(%rbp), %rdi
> >>>>> movl (%rdi), %esi
> >>>>> movl %esi, -12(%rbp)
> >>>>> movl -16(%rbp), %edx
> >>>>> addl %esi, %edx
> >>>>> xorl %eax, %eax
> >>>>> call opal_atomic_cmpset_32
> >>>>> testl %eax, %eax
> >>>>> je .LB157
> >>>>> movl -16(%rbp), %eax
> >>>>> addl -12(%rbp), %eax
> >>>>> ## lineno: 10
> >>>>> leave
> >>>>> ret
> >>>>> .type foo,@function
> >>>>> .size foo,.-foo
> >>>>> ..Dcfe0:
> >>>>> __fooEND:
> >>>>> .section .pgi_trace
> >>>>> .align 8
> >>>>> .quad foo ## address of routine
> >>>>> .quad __fooEND - foo ## size of routine
> >>>>> .2byte 0 ## flags for future use
> >>>>> .2byte 3 ## length of following string
> >>>>> ## name:foo:
> >>>>> .byte 0x66,0x6f,0x6f,0x00
> >>>>> .data
> >>>>> .globl opal_atomic_cmpset_32
> >>>>> .section .debug_frame
> >>>>> ..Dcieb0:
> >>>>> .4byte ..Dciee0-..Dcieb0-4 ## CIE length
> >>>>> .4byte 0xffffffff ## CIE ID
> >>>>> .byte 0x1 ## CIE version
> >>>>> .byte 0x0 ## no augmentation
> >>>>> .byte 0x1 ## ULEB128 1, code alignment factor
> >>>>> .byte 0x78 ## SLEB128 -8, data alignment factor
> >>>>> .byte 0x10 ## return address column
> >>>>> .byte 0xc ## DW_CFA_def_cfa (col 7)
> >>>>> .byte 0x7 ## ULEB128 7
> >>>>> .byte 0x8 ## ULEB128 8
> >>>>> .byte 0x90 ## DW_CFA_offset (col 16)
> >>>>> .byte 0x1 ## ULEB128 1
> >>>>> .align 8
> >>>>> ..Dciee0:
> >>>>> .4byte ..Dfdee0-..Dfdeb0 ## FDE length
> >>>>> ..Dfdeb0:
> >>>>> .4byte ..Dcieb0 ## CIE pointer
> >>>>> .quad ..Dcfb0 ## initial location
> >>>>> .quad ..Dcfe0-..Dcfb0 ## address range
> >>>>> .byte 0x4 ## DW_CFA_advance_loc4
> >>>>> .4byte ..Dcfi0-..Dcfb0
> >>>>> .byte 0xe ## DW_CFA_def_cfa_offset
> >>>>> .byte 0x10 ## ULEB128 16
> >>>>> .byte 0x86 ## DW_CFA_offset (col 6)
> >>>>> .byte 0x2 ## ULEB128 2
> >>>>> .byte 0x4 ## DW_CFA_advance_loc4
> >>>>> .4byte ..Dcfi1-..Dcfi0
> >>>>> .byte 0xd ## DW_CFA_def_cfa_register (col 6)
> >>>>> .byte 0x6 ## ULEB128 6
> >>>>> .align 8
> >>>>> ..Dfdee0:
> >>>>> .ident "PGC 7.0-7"
> >>>>> [7:13] svbu-mpi:~/tmp %
> >>>>> -----
> >>>>>
> >>>>> --
> >>>>> Jeff Squyres
> >>>>> jsquyres_at_[hidden]
> >>>>> For corporate legal information go to:
> >>>>> http://www.cisco.com/web/about/doing_business/legal/cri/
> >>>>>
> >>>>>
> >>>>> _______________________________________________
> >>>>> devel mailing list
> >>>>> devel_at_[hidden]
> >>>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >>>>
> >>>>
> >>>> _______________________________________________
> >>>> devel mailing list
> >>>> devel_at_[hidden]
> >>>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >>>>
> >>>
> >>>
> >>> --
> >>> Jeff Squyres
> >>> jsquyres_at_[hidden]
> >>> For corporate legal information go to:
> >>> http://www.cisco.com/web/about/doing_business/legal/cri/
> >>>
> >>>
> >>> _______________________________________________
> >>> devel mailing list
> >>> devel_at_[hidden]
> >>> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >>
> >>
> >> _______________________________________________
> >> devel mailing list
> >> devel_at_[hidden]
> >> http://www.open-mpi.org/mailman/listinfo.cgi/devel
> >
> >
> > --
> > Jeff Squyres
> > jsquyres_at_[hidden]
> > For corporate legal information go to:
> > http://www.cisco.com/web/about/doing_business/legal/cri/
> >
> >
> > _______________________________________________
> > devel mailing list
> > devel_at_[hidden]
> > http://www.open-mpi.org/mailman/listinfo.cgi/devel
>
>
> _______________________________________________
> devel mailing list
> devel_at_[hidden]
> http://www.open-mpi.org/mailman/listinfo.cgi/devel
>

-- 
Jeff Squyres
jsquyres_at_[hidden]
For corporate legal information go to:
http://www.cisco.com/web/about/doing_business/legal/cri/