Bug 1442

Summary: openGL application seems not working on qemux86
Product: [Build System, Metadata & Runtime] OE-Core Reporter: Lianhao Lu <lianhao.lu>
Component: graphicsAssignee: Edwin Zhai <edwin.zhai>
Status: RESOLVED FIXED QA Contact:
Severity: normal    
Priority: High CC: edwin.zhai, meta.mr.watcher, meta.watcher, sgw
Version: unspecified   
Target Milestone: 1.1   
Hardware: x86   
OS: x86   
Whiteboard:
OS type for building Yocto: --- Type of Regression: ---
Verified: Documentation change: ---

Description Lianhao Lu 2011-09-01 19:59:53 UTC
Tree/Branch: Poky/Master
Commit: dfce2695b96ab587b8235cfab2859a874e9d2eef

Using the core-image-sato-sdk image for qemux86, when running openGL applications in the target, the applications would report the following errors in the target:

sh-4.1# ./glxgears
go on...
You maybe forgot to launch QEMU with -enable-gl argument.
Trying TCP/IP communication instead
impossible to connect to server: Connection refused

Checking the QEMU launch options confirmed that "-enable-gl" argument was used. And the qemu printed the following errors on the host:

func_number >= 0 && func_number < GL_N_CALLS failed
cpu_get_phys_page_debug(env, 00000000) == ffffffff
not in phys mem 00000000(ffffffff 00000000)
cpu_x86_handle_mmu_fault = 1
call _init32 pid=1159
cannot get call parameters
Comment 1 Edwin Zhai 2011-09-21 23:56:00 UTC
Root cause is new gcc's optimization plus our small bug for saving register.
opengl calling (in call_opengl_qemu) includes 4 steps:
1. prepare opengl parameters on stack
2. save some "input" register by push
3. load "input" register with parameters on stack via same index as step 1
4. issue "int 0x99" to trap into qemu, who will get parameter in the registers

Attaching objdump of different gcc output for same qemugl code, you can see new gcc uses "%esp" rather than "%ebp" to index local variable in stack, which leads wrong index in step 3, as push decrease "%esp" automatically. So when trapping into qemu, wrong parameters are accessed.

Saving registers via local variables will fix it.


** compiled by new gcc **
mov    %eax,0x38(%esp)
mov    %edx,0x34(%esp)
mov    0x60(%esp),%eax
mov    0x68(%esp),%edx
mov    %eax,0x30(%esp)
mov    %ebp,0x2c(%esp)
mov    %edx,0x28(%esp)
push   %ebx
push   %ecx
push   %edx
push   %esi
mov    0x38(%esp),%eax
mov    0x34(%esp),%ebx
mov    0x30(%esp),%ecx
mov    0x2c(%esp),%edx
mov    0x28(%esp),%esi
int    $0x99
pop    %esi
pop    %edx
pop    %ecx
pop    %ebx


** compiled by old gcc **
mov    %eax,-0x1c(%ebp)
mov    %edx,-0x20(%ebp)
mov    0x8(%ebp),%eax
mov    0xc(%ebp),%edx
mov    0x10(%ebp),%ecx
mov    %eax,-0x24(%ebp)
mov    %edx,-0x28(%ebp)
mov    %ecx,-0x2c(%ebp)
push   %ebx
push   %ecx
push   %edx
push   %esi
mov    -0x1c(%ebp),%eax
mov    -0x20(%ebp),%ebx
mov    -0x24(%ebp),%ecx
mov    -0x28(%ebp),%edx
mov    -0x2c(%ebp),%esi
int    $0x99
pop    %esi
pop    %edx
pop    %ecx
pop    %ebx