Intel Software Network - <![CDATA[ Why "subq" as allocate by ICC-v10.0 but not as prologue, but ICC-v11.0 uses "pushq" as prologue? ]]> feed http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072 en-us Re: Why &#34;subq&#34; as allocate by ICC-v10.0 but not as prologue, but ICC-v11.0 uses &#34;pushq&#34; as prologue?


(1)  In 11.0 version of ICC generates an aligned (to 128) frame for main. This is the intentional behavior and is done for performance reasons though not for this specific example. This where all the uses of %rbp arise (it is used to save the value of %rsp before alignment). If you are confused with this extra alignment, please put your code into a function other than “main”. Consider the prologue for “main” with 10.1:

main:

        subq      $4198408, %rsp                          

vs 11.0 prologue:

 main:

        pushq     %rbp                                       

        movq      %rsp, %rbp                             

        andq      $-128, %rsp                          

        subq      $4198400, %rsp                     

 

(2)  The “-fno-builtin” option causes the compiler to not expand intrinsics code inline. The code in the example doesn’t make use of any intrinsics which might suggest that this option should have no effect here. Interestingly but it does affect the way the pattern of setting memory to zero is recognized. It may be considered a "bug" in the sense that the behavior is not the one as expected, but I don't fell too strong about it. It is important that both variants are correct to the matter of what the semantics of the option is and they both look reasonably adequate in terms of performance. Generally it is true that the use of “-fno-builtin” would result in a smaller though slower code, but it is incorrect assumption that the code will indeed be smaller. If you are really interested in code size vs. performance, please use option “-Os”.

 

(3)  The “pushq %rsi” code that you refer in (d) has nothing to do to parameter passing for routine “__sti__$E”. It is just an easy way to adjust the stack pointer by 8 bytes to make sure that it is properly aligned to 16-byte boundary at the subsequent call, to conform to x86-64 ABI. The “popq %rcx” is its counterpart in function epilog.

34) I cannot reproduce the behavior you describe with regards to MOVNTDQ and also your asm snippets don't contain  any MOVNTDQ. Maybe you have more details that you haven't shared?

Regards,

-Sergey

 ]]> http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072/ Tue, 27 Jan 2009 04:33:53 -0800 http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072/ Parallel Programming ISN General Re: Why &#34;subq&#34; as allocate by ICC-v10.0 but not as prologue, but ICC-v11.0 uses &#34;pushq&#34; as prologue? Hello,

Thanks for your valuable input.

I have new query similar to above as below -

(a) For multiple C++ package file, when I do vectorizations (calling of pragma's) within that file within section of code, I get starting and ending asm as -
{
44d960:            55                                  push   %rbp
44d961:            48 83 ec 50                   sub    $0x50,%rsp
44d965:            49 89 f0                         mov    %rsi,%r8
44d968:            4c 63 c9                        movslq %ecx,%r9
...

...
44dc84:            48 83 c4 50                  add    $0x50,%rsp
44dc88:            5d                                 pop    %rbp
44dc89:            c3                                 retq
44dc8a:            90                                 nop
44dc8b:            48 8d 74 26 00             lea    0x0(%rsi),%rsi
}

(b) But the same code w/o using any pragma's call, the starting & ending asm are as -
{
44d960:           48 83 ec 68                  sub    $0x68,%rsp
44d964:           49 89 f9                        mov    %rdi,%r9
44d967:           49 89 d0                       mov    %rdx,%r8
44d96a:           4c 63 d1                       movslq %ecx,%r10
..
..
..
44dc4e:           48 83 c4 68                   add    $0x68,%rsp
44dc52:           c3                                  retq
44dc53:           90                                  nop
44dc54:           48 8d 74 26 00             lea    0x0(%rsi),%rsi
44dc59:           48 8d bf 00 00 00 00     lea    0x0(%rdi),%rdi
}
---

Query:
(1) Could the difference between having PUSH/POP call with pragma vectorization calls and not having w/o it be differentiated?

(2) W/o pragma calls, the asm in (b) has "lea" calls twice and also the during starting it has - sub, mov, mov & movslq than with pragma calls, why pragma calls bring such a difference?

Sorry, I didn't thought of creating a new thread.

~BR ]]> http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072/ Fri, 24 Apr 2009 02:29:46 -0700 http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072/ Parallel Programming ISN General Re: Why &#34;subq&#34; as allocate by ICC-v10.0 but not as prologue, but ICC-v11.0 uses &#34;pushq&#34; as prologue?
Quoting - srimks
Hello,

Thanks for your valuable input.

I have new query similar to above as below -

(a) For multiple C++ package file, when I do vectorizations (calling of pragma's) within that file within section of code, I get starting and ending asm as -
{
44d960:            55                                  push   %rbp
44d961:            48 83 ec 50                   sub    $0x50,%rsp
44d965:            49 89 f0                         mov    %rsi,%r8
44d968:            4c 63 c9                        movslq %ecx,%r9
...

...
44dc84:            48 83 c4 50                  add    $0x50,%rsp
44dc88:            5d                                 pop    %rbp
44dc89:            c3                                 retq
44dc8a:            90                                 nop
44dc8b:            48 8d 74 26 00             lea    0x0(%rsi),%rsi
}

(b) But the same code w/o using any pragma's call, the starting & ending asm are as -
{
44d960:           48 83 ec 68                  sub    $0x68,%rsp
44d964:           49 89 f9                        mov    %rdi,%r9
44d967:           49 89 d0                       mov    %rdx,%r8
44d96a:           4c 63 d1                       movslq %ecx,%r10
..
..
..
44dc4e:           48 83 c4 68                   add    $0x68,%rsp
44dc52:           c3                                  retq
44dc53:           90                                  nop
44dc54:           48 8d 74 26 00             lea    0x0(%rsi),%rsi
44dc59:           48 8d bf 00 00 00 00     lea    0x0(%rdi),%rdi
}
---

Query:
(1) Could the difference between having PUSH/POP call with pragma vectorization calls and not having w/o it be differentiated?

(2) W/o pragma calls, the asm in (b) has "lea" calls twice and also the during starting it has - sub, mov, mov & movslq than with pragma calls, why pragma calls bring such a difference?

Sorry, I didn't thought of creating a new thread.

~BR

The prolog/epilog sequences (i.e. what you refer to as "starting and ending asm") can indeed be very sensitive to the other code in the routine & to the level of optimizations applied. Several simple examples can be different register pressure and/or different alignment constraints for the local variables. You shouldn't however be ever able to reason on one sequence or another - it is completely implementation dependent. Well, you may always make a nice guess, but that would be perfectly incorrect to make any assumptions based on such a guess.

BTW, the "lea" instructions you are seeing at the end are just NOPs that are never executed and are their for code alignment purposes only.

]]> http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072/ Mon, 02 Nov 2009 22:18:24 -0800 http://software.intel.com/en-us/forums/intel-avx-and-cpu-instructions/topic/63072/ Parallel Programming ISN General