Bug 11643 – Loop-invariant code motion optimization for associative array literals

A common idiom in scripting languages like Python is to define and use in-place associative arrays. Thanks to the handy D associative array literals this idiom is usable in D too: void main() { import core.stdc.stdio: printf; foreach (immutable i; 0 .. 10) printf("%d ", [1: 10, 2: 20].get(i, -1)); } Output: -1 10 20 -1 -1 -1 -1 -1 -1 -1 Unfortunately the associative array is rebuilt at every loop (dmd 2.065alpa not optimized built): __Dmain: L0: enter 01Ch,0 push EBX push ESI mov dword ptr -01Ch[EBP],0 LD: cmp dword ptr -01Ch[EBP],0Ah jge L7F mov EAX,-01Ch[EBP] mov -018h[EBP],EAX push dword ptr -018h[EBP] mov ECX,offset FLAT:_D4temp4mainFZv12__dgliteral1MFNaNbNfZi xor EBX,EBX push ECX push EBX mov EDX,1 mov -014h[EBP],EDX mov dword ptr -010h[EBP],2 mov dword ptr -0Ch[EBP],0Ah mov dword ptr -8[EBP],014h lea ESI,-0Ch[EBP] push ESI push 2 lea ECX,-014h[EBP] mov EAX,2 push ECX push EAX mov EBX,offset FLAT:_D12TypeInfo_Hii6__initZ push EBX call near ptr __d_assocarrayliteralTX add ESP,014h mov -4[EBP],EAX lea EAX,-4[EBP] call near ptr _D6object26__T16AssociativeArrayTiTiZ16AssociativeArray3getMFNaiLiZi push EAX mov EDX,offset FLAT:_DATA push EDX call near ptr _printf add ESP,8 inc dword ptr -01Ch[EBP] jmp short LD L7F: xor EAX,EAX pop ESI pop EBX leave ret Even compiling that code with ldc2 0.12.0, with -O (optimized build), the problem is still present: __Dmain: pushl %edi pushl %esi subl $24, %esp xorl %esi, %esi leal 20(%esp), %edi .align 16, 0x90 LBB0_1: movl $_.aaValuesStorage, 16(%esp) movl $2, 12(%esp) movl $_.aaKeysStorage, 8(%esp) movl $2, 4(%esp) movl $__D12TypeInfo_Hii6__initZ, (%esp) calll __d_assocarrayliteralTX movl %esi, 20(%esp) movl %edi, 8(%esp) movl %eax, (%esp) movl $__D10TypeInfo_i6__initZ, 4(%esp) calll __aaInX movl $-1, %ecx testl %eax, %eax je LBB0_3 movl (%eax), %ecx LBB0_3: movl %ecx, 4(%esp) movl $_.str, (%esp) calll ___mingw_printf incl %esi cmpl $10, %esi jne LBB0_1 xorl %eax, %eax addl $24, %esp popl %esi popl %edi ret So I suggest to introduce some loop-invariant code motion optimization, and initialize the associative array only once out of the loop (dmd asm coming from modified code): __Dmain: L0: enter 020h,0 push EBX push ESI mov dword ptr -018h[EBP],1 mov dword ptr -014h[EBP],2 mov dword ptr -010h[EBP],0Ah mov dword ptr -0Ch[EBP],014h lea EAX,-010h[EBP] push EAX push 2 lea ECX,-018h[EBP] mov EBX,2 push ECX push EBX mov EDX,offset FLAT:_D12TypeInfo_Hii6__initZ push EDX call near ptr __d_assocarrayliteralTX mov -020h[EBP],EAX mov dword ptr -8[EBP],0 add ESP,014h L4A: cmp dword ptr -8[EBP],0Ah jge L7E mov ESI,-8[EBP] mov -4[EBP],ESI push dword ptr -4[EBP] mov ECX,offset FLAT:_D4temp4mainFZv12__dgliteral1MFNaNbNfZi xor EAX,EAX push ECX push EAX lea EAX,-020h[EBP] call near ptr _D6object26__T16AssociativeArrayTiTiZ16AssociativeArray3getMFNaiLiZi push EAX mov EDX,offset FLAT:_DATA push EDX call near ptr _printf add ESP,8 inc dword ptr -8[EBP] jmp short L4A L7E: xor EAX,EAX pop ESI pop EBX leave ret

*** This issue has been marked as a duplicate of issue 4453 ***