Subversion Repositories Open64

[/] [trunk/] - Rev 3353

Rev

Go to most recent revision

Filtering Options

Clear current filter

Rev Log message Author Age Path
3353 Fix problem exposed by r3284, libhugetlbfs 64 bit linker scripts
should be installed the new 64 bit library directory.

Also included is a cleanup change that removes an error message
that is innocuously generated when the install script is run more
than once.

Reviewed by Mei.
dgilmore 970d 04h /trunk/
3352 Add myself to maintainers. lidr 970d 12h /trunk/
3351 Fix the data race bug of __ompc_critical() in osprey/libopenmp/omp_lock.c.

Approved by: Tianwei

BTW, revert a minor portion of my previous checkin (rev 3345). Some code in that checkin is to fix the bug of local variable defined inside OpenMP paralllel region, but I find that bug has been already fixed in rev 3211 in open64-booster branch, and that change has been merged into trunk in rev 3314.
hjz 970d 19h /trunk/
3348 Fix for bug 600:

The gspin tree of the failing example has the following pattern:
...
@gs_constructor_elts_value CONS<717520>
GS_NOP_EXPR<715928> { ...
@gs_tree_type LINK 34236
@operand 0 GS_CONSTRUCTOR<716012>
...

Here, GS_ CONSTRUCTOR is the operand of a GS_NOP_EXPR. In wgen
Traverse_Aggregate_Struct, it processes GS_CONSTRUCTOR,
pointer-to-member-functions and others cases, while it does not skip
the GS_NOP_EXPR.

The fix is to skip GS_NOP_EXPR before processing GS_CONSTRUCTOR, or others.

Code Review: Jian-Xin Lai
minz 971d 10h /trunk/
3347 Fix for bug 611:

The reason for the assertion is that gcc FE treats returning "complex double"
as return_in_memory and performs nrv (named return value) opt in -m32 mode.
While, wgen not treat any complex type as return_in_memroy.

The fix is to set complex type whose tsize larger than 12 to be
return_in_memory in -m32 mode (according to gcc FE convention).
Two assertions are also changed to include complex type as one possibile
return_in_memory type.

Code Review: Jian-Xin Lai
minz 971d 10h /trunk/
3346 Fix bug in VHO_Enable_Simple_If_Conv optimization.

Reviewed by Fred, Sun & Mei
shenruifen 971d 23h /trunk/
3345 Fixed several OpenMP bugs.

1. Fixed the bug of OpenMP driver. Now -fopenmp option works as well as -mp.
2. Fixed the bug of OpenMP atomic directive. Previously, some operation inside OpenMP ATOMIC was lowered to INTRN_COMPARE_AND_SWAP_XXX, but in fact they should be lowered to INTRN_BOOL_COMPARE_AND_SWAP_XXX.
3. Previously, local variables defined in parallel region was treated as shared variable. I added new code in wgen to generate PRIVATE clause in such parallel region for such variables implicitly, to solve the problem.
4. Previously, some nested OpenMP directives (such as CRITICAL inside another CRITICAL) may trigger an assertion failure. I fixed this problem by changing the place to call the verifier.

Approved by: Sun Chan
hjz 972d 15h /trunk/
3343 Avoid bash-specific redirection operators in make rules.

For the top-level clean action, clean both 'lib' and 'lib2' targets if
they are in use. Do not clean GNU3 components if not building them.

For libhugetlbfs, ensure that make does not try to build the 'version'
file in the source tree using an implicit rule for creating executables.

Approved by: Suneel Jain
dcoakley 974d 11h /trunk/
3342 Fixed g++ builtin regression tests fails. (Bug 627)

Approved by Lai JianXin.
chenr 978d 02h /trunk/
3340 Bug 592: Assertion failure in file: osprey/../libspin/gspin-tel.h,
function: gs_tree_value, line: 947

It happens when the initializer list for vector typed variable is empty.
The code used to get the element type from initializer, when it is empty,
the code gets a null pointer access. Now it is changed to use the vector
MTYPE to get its element type.

__attribute__ ((vector_size (sizeof (double) * 4))) double d1 = { 1., 2 };
__attribute__ ((vector_size (sizeof (double) * 4))) double d2 = { };
__attribute__ ((vector_size (sizeof (float) * 4))) float f1 = { 1., 2 };
__attribute__ ((vector_size (sizeof (float) * 4))) float f2 = { };
__attribute__ ((vector_size (sizeof (int) * 4))) int a = { 1, 2 };
__attribute__ ((vector_size (sizeof (int) * 4))) int c = { };

Code Review: JianXin Lai
xiaohua_zhang 978d 07h /trunk/
3338 M MAINTAINERS jaewook 982d 09h /trunk/
3337 Fix for bug 593:

The IR generated for __builtin_va_arg (ap, _Complex long double) is as:
U8LDA 0 <2,2,ap> T<69,anon_ptr.,8>
U8U8ILOAD 8 T<9,.predef_U8,8> T<67,anon_ptr.,8>
MMILOAD -32 T<23,.predef_C10,16> T<98,anon_ptr.,8>
This is not appropriate. We should generate C10C10ILOAD instead.

The problem is that the GS_VA_ARG_EXPR handling only considers MTYPE_C4,
not other complex types (e.g., MTYPE_C8 or MTYPE_C10).
The fix is to include them.

Code Review: Sun Chan
minz 982d 10h /trunk/
3336 fixed IA64 compilation bug since revision 3314
this check in enclose some of the x8664 only code with #ifdef macros.

Approved by Jianxin
ycwu 984d 00h /trunk/
3335 Add option entry for Wextra,Wno-extra,Wmissing-include-dirs zhuqing 984d 20h /trunk/
3334 add target constraint for _mm_extract_pi16, fix bug602 zhuqing 989d 20h /trunk/
3333 modify comment format zhuqing 990d 20h /trunk/
3332 add intrinsic for _mm_extract_pi16 zhuqing 990d 21h /trunk/
3330 add new maintainers. zhuqing 990d 21h /trunk/
3327 Fix for bug 582.

Approved by Lai JianXin.
chenr 992d 01h /trunk/
3326 Avoid writing build artifacts into the source tree.

Approved by: Jian-Xin Lai
dcoakley 992d 21h /trunk/

1 2 Next >

Show All