Reading constants into an SSE/AVX registers in inline assembly

Question

I'm trying to import a .o file from Lazarus into Delphi.

function Test: boolean;
const
  TestData: array[0..15] of byte = (0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0);
asm
  movdqu xmm0,[rip+testData]   //xmm0 should be all zeros
  pcmpeqq xmm1,xmm1        //xmm1 is all ones
  ptest xmm0,xmm1          //ZF = (xmm0 and xmm1) = 0
  setz al                  //return true if test works, false if not
end;

The above routine works in Delphi, but fails in Lazarus.

I'm importing the .o file using

   {$L 'C:\pathname\TestUnit.o'}

function TESTUNIT_TEST: boolean; external name 'TESTUNIT_$$_TEST$$BOOLEAN';

Lazarus replaces the reference to TestData with a zero offset.

AVXGENERATE_TEST:
//never mind the stack frame
0000000000A0F3B0 488D6424F8       lea rsp,[rsp-$08]
//                                                        rel 0? why?
0000000000A0F3B5 F30F6F0500000000 movdqu xmm0,dqword ptr [rel $00000000]
0000000000A0F3BD 660F3829C9       pcmpeqq xmm1,xmm1
0000000000A0F3C2 660F3817C1       ptest xmm0,xmm1
0000000000A0F3C7 0F94C0           setz al
0000000000A0F3CA 488D642408       lea rsp,[rsp+$08]
0000000000A0F3CF C3               ret

Here is the output from `objdump -dr "C:\pathname\TestUnit.o"

Disassembly of section .text.n_avxgenerate_$$_test$$boolean:

0000000000000000 <AVXGENERATE_$$_TEST$$BOOLEAN>:
   0:   48 8d 64 24 f8          lea    -0x8(%rsp),%rsp
   5:   f3 0f 6f 05 00 00 00    movdqu 0x0(%rip),%xmm0        # d <AVXGENERATE_$$_TEST$$BOOLEAN+0xd>
   c:   00
                        9: R_X86_64_PC32        .data.n_tc_$avxgenerate_$$_testdata
   d:   66 0f 38 29 c9          pcmpeqq %xmm1,%xmm1
  12:   66 0f 38 17 c1          ptest  %xmm1,%xmm0
  17:   0f 94 c0                sete   %al
  1a:   48 8d 64 24 08          lea    0x8(%rsp),%rsp
  1f:   c3

How do I get lazarus to read the constant correctly into an xmm register?
Is there a workaround for this issue?

I plan to include some AVX assembly written in lazarus in a Delphi program.
For testing purposes I'm just using SSE code here.

Version info
FPC 3.0.0
Lazarus version 1.6.2

Answer 1

A workaround is to put the routines to be imported into a DLL.
This means that you no longer have all your stuff in a standalone .exe file.
But at least all code links correctly.

In Lazarus

library Test_lib;

{$mode delphi}{$H+}

uses
  Classes,
  TestUnit;

{$R *.res}

exports
  Test;

begin
end.    

in Delphi

const
  Testlib = 'C:\pathname\Lazarus\Test_lib.dll';

function LazarusTest: boolean; external Testlib name 'Test';

Now the output is:

0000000110020570 488D6424F8       lea rsp,[rsp-$08]
{                                 links to the correct location VVVVVV}
0000000110020575 F30F6F0553930000 movdqu xmm0,dqword ptr [rel $00009353]
000000011002057D 660F3829C9       pcmpeqq xmm1,xmm1
0000000110020582 660F3817C1       ptest xmm0,xmm1
0000000110020587 0F94C0           setz al
000000011002058A 488D642408       lea rsp,[rsp+$08]
000000011002058F C3               ret