Bug 5100 – [Intel Atom] -O Degrades performance of loop statements

Two example cases: loop1.d --------- void main() { for (int i = 0; i < int.max; i++) { } } loop2.d --------- void main() { int i = 0; while(i < int.max) { i++; } } $ dmd loop1.d $ time ./loop1 real 0m2.914s user 0m2.884s sys 0m0.012s $ ./dmd loop1.d -O $ time ./loop1 real 0m5.695s user 0m5.684s sys 0m0.004s $ ./dmd loop2.d $ time ./loop2 real 0m2.912s user 0m2.892s sys 0m0.004s $ ./dmd loop2.d -O $ time ./loop2 real 0m5.703s user 0m5.688s sys 0m0.004s The speed of execution slows by almost double when optimisations are turned on. Something isn't right here...

for the lazy #5100 is

(In reply to comment #1) > for the lazy #5100 is ooops very sorry this comment was meant to be for bug #5294 cause i think it is related.

Cannot reproduce. On Windows, for both test cases, without -O it's about 5 seconds (does an INC and CMP of a stack variable). With -O it is about 1 second (just does INC and CMP of EAX).

objdump without -O on Linux: push %ebp mov %esp,%ebp sub $0x4,%esp movl $0x0,-0x4(%ebp) cmpl $0x7fffffff,-0x4(%ebp) jge 1c <_Dmain+0x1c> addl $0x1,-0x4(%ebp) jmp d <_Dmain+0xd> xor %eax,%eax leave ret objdump with -O on Linux push %ebp mov %esp,%ebp xor %eax,%eax add $0x1,%eax cmp $0x7fffffff,%eax jb 5 <_Dmain+0x5> pop %ebp xor %eax,%eax ret Looks to be same as what Don said that was on his Windows box. Wonder why Linux is slower... (must be a quirk, that or my Intel Atom CPU is to blame).

Been playing about with GCC, this seems to be a better performant: Objdump: push %ebp mov %esp,%ebp and $0xfffffff0,%esp push %eax sub $0xc,%esp lea 0x0(%esi),%esi add $0x1,%eax cmp $0x7fffffff,%eax jne 30 <_Dmain+0x10> add $0xc,%esp mov %ebp,%esp pop %ebp ret GCC assembly: .globl _Dmain .type _Dmain, @function _Dmain: .LFB0: .cfi_startproc .cfi_personality 0x0,__gdc_personality_v0 pushl %ebp .cfi_def_cfa_offset 8 movl %esp, %ebp .cfi_offset 5, -8 .cfi_def_cfa_register 5 andl $-16, %esp pushl %eax .cfi_escape 0x10,0x3,0x7,0x55,0x9,0xf0,0x1a,0x9,0xfc,0x22 subl $12, %esp .p2align 4,,7 .p2align 3 .L4: addl $1, %eax cmpl $2147483647, %eax jne .L4 addl $12, %esp movl %ebp, %esp popl %ebp ret .cfi_endproc .LFE0: .size _Dmain, .-_Dmain Can attach the full .s file if needed. Regards

Perhaps it's because gcc is doing: ADD EAX,1 instead of: INC EAX

Maybe not... I actually get the reverse on my new laptop with 2.057, $ dmd loop2.d $ objdump loop2.o -d push %ebp mov %esp,%ebp sub $0x4,%esp movl $0x0,-0x4(%ebp) cmpl $0x7fffffff,-0x4(%ebp) jge 1b <_Dmain+0x1b> incl -0x4(%ebp) jmp d <_Dmain+0xd> xor %eax,%eax leave ret $ time ./loop2 real 0m11.780s user 0m11.769s sys 0m0.004s $ dmd loop2.d -O $ objdump loop2.o -d push %ebp mov %esp,%ebp xor %eax,%eax inc %eax cmp $0x7fffffff,%eax jb 5 <_Dmain+0x5> pop %ebp xor %eax,%eax ret $ time ./loop2 real 0m3.936s user 0m3.924s sys 0m0.008s

And on my netbook: $ dmd loop2.d $ time ./loop2 real 0m2.948s user 0m2.924s sys 0m0.012s $ dmd loop2.d -O $ time ./loop2 real 0m5.725s user 0m5.688s sys 0m0.012s Specs of Netbook: processor : 0 vendor_id : GenuineIntel cpu family : 6 model : 28 model name : Intel(R) Atom(TM) CPU N270 @ 1.60GHz stepping : 2 cpu MHz : 800.000 cache size : 512 KB cpu cores : 1 Specs of Laptop: processor : 0 vendor_id : AuthenticAMD cpu family : 20 model : 2 model name : AMD E-450 APU with Radeon(tm) HD Graphics stepping : 0 cpu MHz : 825.000 cache size : 512 KB cpu cores : 2 Regards

My gut feeling is that the main source of it slowing down is the needless push and pop of the frame pointer.

For a while now I've been thinking that the bottleneck is probably to do with alignment, but I'd have to get out my (now two generations old) atom netbook to investigate further.

THIS ISSUE HAS BEEN MOVED TO GITHUB https://github.com/dlang/dmd/issues/18313 DO NOT COMMENT HERE ANYMORE, NOBODY WILL SEE IT, THIS ISSUE HAS BEEN MOVED TO GITHUB