Hi Prof. Dong,
Since your team maintains the PPC port, could you please review my
change? Thank you very much.
2012/8/18 Sun Chan <sun.c...@gmail.com>:
> your changed touched upon ppc target, did you check with ppc people
> and get their ok?
> Sun
>
> On Fri, Aug 17, 2012 at 6:55 PM, Jian-Xin Lai <laij...@gmail.com> wrote:
>> Hi,
>>
>> Could someone review this patch? Thank you very much.
>>
>> 2012/8/1 Jian-Xin Lai <laij...@gmail.com>:
>>> Hi Sun,
>>>
>>> Could you please review the new patch? Thank you very much.
>>> A new function adjust_type_for_return_info is added to adjust the
>>> rtype and desc type according to the size left for copy.
>>>
>>> 2012/7/31 Jian-Xin Lai <laij...@gmail.com>:
>>>> I plan to change my patch. I think the change in targ_sim.cxx isn't
>>>> necessary since not only X86_64 has this issue. It looks like it's better
>>>> to fix this issue in the lowering phase. For the last piece of return
>>>> value, we should't use the mtype from the return info. A new mtype should
>>>> be calculated based on the size of the last piece.
>>>> I'll send a new patch later.
>>>>
>>>> Regards,
>>>> Jian-Xin
>>>>
>>>> 在 2012-7-31，19:17，Sun Chan <sun.c...@gmail.com> 写道：
>>>>
>>>> Mips used to pass float in integer registers, I don't know if they
>>>> continue to doing that. Longson people?
>>>> Sun
>>>>
>>>> On Tue, Jul 31, 2012 at 6:56 PM, Jian-Xin Lai <laij...@gmail.com> wrote:
>>>>> IA-64 has the similar problem.
>>>>>
>>>>> 2012/7/31 Sun Chan <sun.c...@gmail.com>:
>>>>>> And IA64?
>>>>>> Sun
>>>>>>
>>>>>> On Tue, Jul 31, 2012 at 6:05 PM, Jian-Xin Lai <laij...@gmail.com> wrote:
>>>>>>> On x86_64, the aggregate shown in the test case is partitioned into
>>>>>>> two eightbyte pieces. Both of them are classified to INTEGER class and
>>>>>>> returned by two 64-bit registers (RAX and RDX).
>>>>>>> If the size of the aggregate is larger than 12, both of the two parts
>>>>>>> should be MTYPE_I8 and use I8 for LDID/STID. Otherwise, the first part
>>>>>>> is MTYE_I8 and the second part should be MTYPE_I4.
>>>>>>>
>>>>>>> 2012/7/31 Sun Chan <sun.c...@gmail.com>:
>>>>>>>> I think this problem is target dependent. For x86, is 12 bytes wrong?
>>>>>>>> I remember it is correct for SGI (which does not matter anymore), but
>>>>>>>> the point really is, this is ABI dependent
>>>>>>>> Sun
>>>>>>>>
>>>>>>>> On Tue, Jul 31, 2012 at 4:15 PM, Jian-Xin Lai <laij...@gmail.com>
>>>>>>>> wrote:
>>>>>>>>> Hi,
>>>>>>>>>
>>>>>>>>> Could someone review the patch for open64 bug #969, #970? Thank you
>>>>>>>>> very much.
>>>>>>>>>
>>>>>>>>> 1. test case
>>>>>>>>> #969 and #970 can be simplified to the following case:
>>>>>>>>> struct I3 {
>>>>>>>>> int a, b, c;
>>>>>>>>> };
>>>>>>>>> extern struct I3 foo();
>>>>>>>>> void bar() {
>>>>>>>>> struct I3 a;
>>>>>>>>> a = foo();
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> The struct I1 can also be replaced by:
>>>>>>>>> struct F3 {
>>>>>>>>> float a, b, c;
>>>>>>>>> };
>>>>>>>>>
>>>>>>>>> 2. IR dump
>>>>>>>>> Open64 will allocate a temporary on stack for the return value. The IR
>>>>>>>>> after MSTID lowering looks like:
>>>>>>>>>
>>>>>>>>> MCALL 126 <1,51,foo> # flags 0x7e {line: 1/7}
>>>>>>>>> I8I8LDID 1 <1,5,.preg_I8> T<5,.predef_I8,8> # $r1
>>>>>>>>> I8STID 0 <2,2,_temp_.Mreturn.foo0> T<5,.predef_I8,8> {line: 0/0}
>>>>>>>>> I8I8LDID 7 <1,5,.preg_I8> T<5,.predef_I8,8> # $r7
>>>>>>>>> I8STID 8 <2,2,_temp_.Mreturn.foo0> T<5,.predef_I8,8> {line: 0/0}
>>>>>>>>>
>>>>>>>>> [2]: _temp_.Mreturn.foo0 <2,2> Variable of type I1 (#53,
>>>>>>>>> KIND_STRUCT)
>>>>>>>>> Address: 0(_temp_.Mreturn.foo0<2,2>) Alignment: 8
>>>>>>>>> bytes
>>>>>>>>> location: file (null), line 0
>>>>>>>>> Flags: 0x00400000 temp, XLOCAL
>>>>>>>>> Sclass: AUTO
>>>>>>>>>
>>>>>>>>> [53]: I3 : (f: 0x1000 content_seen) size 12 M: STRUCT
>>>>>>>>> 0 a .predef_I4 (#4) align 4
>>>>>>>>> fl:0x0000
>>>>>>>>> 4 b .predef_I4 (#4) align 4
>>>>>>>>> fl:0x0000
>>>>>>>>> 8 c .predef_I4 (#4) align 4
>>>>>>>>> fl:0x0001 last_field
>>>>>>>>>
>>>>>>>>> The size of _temp_.Mreturn.foo0 is only 12 byte. The I8STID will
>>>>>>>>> overwrite some other variables on stack. For float struct F3, two
>>>>>>>>> F8STID will be generated.
>>>>>>>>>
>>>>>>>>> 3. Patch
>>>>>>>>> Index: osprey/be/com/wn_lower.cxx
>>>>>>>>> ===================================================================
>>>>>>>>> --- osprey/be/com/wn_lower.cxx (revision 3998)
>>>>>>>>> +++ osprey/be/com/wn_lower.cxx (working copy)
>>>>>>>>> @@ -7299,6 +7299,7 @@
>>>>>>>>> }
>>>>>>>>> else { // return via 1 or more return registers
>>>>>>>>> TYPE_ID mtype, desc;
>>>>>>>>> + WN_OFFSET st_ofst = WN_store_offset(tree);
>>>>>>>>> for (INT32 i = 0; i < RETURN_INFO_count(return_info); i++) {
>>>>>>>>> if (i != 0)
>>>>>>>>> WN_INSERT_BlockLast (block, wn); // insert the last STID
>>>>>>>>> created
>>>>>>>>> @@ -7334,8 +7335,9 @@
>>>>>>>>> }
>>>>>>>>> #endif
>>>>>>>>> wn = WN_CreateStid(OPR_STID, MTYPE_V, desc,
>>>>>>>>> -
>>>>>>>>> WN_store_offset(tree)+i*MTYPE_byte_size(mtype),
>>>>>>>>> + st_ofst,
>>>>>>>>> WN_st_idx(tree), Be_Type_Tbl(desc), n_rhs);
>>>>>>>>> + st_ofst += MTYPE_byte_size(mtype);
>>>>>>>>> wn = lower_store (block, wn, actions);
>>>>>>>>> WN_Set_Linenum(wn, current_srcpos);
>>>>>>>>> }
>>>>>>>>> @@ -7394,6 +7396,7 @@
>>>>>>>>> }
>>>>>>>>> else { // return via 1 or more return registers
>>>>>>>>> WN *base_expr;
>>>>>>>>> + WN_OFFSET st_ofst = WN_store_offset(tree);
>>>>>>>>> for (INT32 i = 0; i < RETURN_INFO_count(return_info); i++) {
>>>>>>>>> if (i != 0)
>>>>>>>>> WN_INSERT_BlockLast (block, wn); // insert the last STID
>>>>>>>>> created
>>>>>>>>> @@ -7406,8 +7409,9 @@
>>>>>>>>> // OSP-438: The type of Istore node should be a pointer to mtype.
>>>>>>>>> // (Previously, we just use Be_Type_Tbl(mtype) as the type of Istore).
>>>>>>>>> wn = WN_CreateIstore(OPR_ISTORE, MTYPE_V, mtype,
>>>>>>>>> -
>>>>>>>>> WN_store_offset(tree)+i*MTYPE_byte_size(mtype),
>>>>>>>>> + st_ofst,
>>>>>>>>> Make_Pointer_Type(Be_Type_Tbl(mtype),
>>>>>>>>> FALSE), n_rhs, base_expr);
>>>>>>>>> + st_ofst += MTYPE_byte_size(mtype);
>>>>>>>>> WN_Set_Linenum(wn, WN_Get_Linenum(tree));
>>>>>>>>> wn = lower_store (block, wn, actions);
>>>>>>>>> WN_Set_Linenum (wn, WN_Get_Linenum(tree));
>>>>>>>>> @@ -12285,6 +12289,7 @@
>>>>>>>>> ("expected RETURN_VAL kid not type M"));
>>>>>>>>>
>>>>>>>>> algn = TY_align(ty_idx);
>>>>>>>>> + WN_OFFSET ld_ofst = WN_load_offset(o_rhs);
>>>>>>>>> for (i = 0; i < RETURN_INFO_count(return_info); i++) {
>>>>>>>>> mtype = RETURN_INFO_mtype(return_info, i);
>>>>>>>>> ty_idx_used = Be_Type_Tbl(mtype);
>>>>>>>>> @@ -12312,15 +12317,14 @@
>>>>>>>>> #endif
>>>>>>>>> #if defined(TARG_PPC32)
>>>>>>>>> n_rhs = WN_CreateLdid (OPR_LDID, mtype, desc,
>>>>>>>>> - WN_load_offset(o_rhs)
>>>>>>>>> - + i * MTYPE_byte_size(mtype),
>>>>>>>>> + ld_ofst,
>>>>>>>>> WN_st_idx(o_rhs), ty_idx_used);
>>>>>>>>> #else
>>>>>>>>> n_rhs = WN_CreateLdid (OPR_LDID, mtype, mtype,
>>>>>>>>> - WN_load_offset(o_rhs)
>>>>>>>>> - + i * MTYPE_byte_size(mtype),
>>>>>>>>> + ld_ofst,
>>>>>>>>> WN_st_idx(o_rhs), ty_idx_used);
>>>>>>>>> #endif
>>>>>>>>> + ld_ofst += MTYPE_byte_size(mtype);
>>>>>>>>> wn = WN_CreateStid(OPR_STID, MTYPE_V, mtype, preg, preg_st,
>>>>>>>>> Be_Type_Tbl(mtype), n_rhs);
>>>>>>>>> WN_Set_Linenum(wn, current_srcpos); // Bug 1268
>>>>>>>>> @@ -12332,6 +12336,7 @@
>>>>>>>>> ("expected RETURN_VAL kid not type M"));
>>>>>>>>>
>>>>>>>>> algn = TY_align(ty_idx);
>>>>>>>>> + WN_OFFSET ld_ofst = WN_load_offset(o_rhs);
>>>>>>>>> for (i = 0; i < RETURN_INFO_count(return_info); i++) {
>>>>>>>>> mtype = RETURN_INFO_mtype(return_info, i);
>>>>>>>>> ty_idx_used = Be_Type_Tbl(mtype);
>>>>>>>>> @@ -12340,10 +12345,10 @@
>>>>>>>>> n_rhs = WN_kid0(o_rhs);
>>>>>>>>> else n_rhs = WN_COPY_Tree(WN_kid0(o_rhs));
>>>>>>>>> n_rhs = WN_CreateIload(OPR_ILOAD, mtype, mtype,
>>>>>>>>> - WN_load_offset(o_rhs)
>>>>>>>>> - + i * MTYPE_byte_size(mtype),
>>>>>>>>> + ld_ofst,
>>>>>>>>> ty_idx_used,
>>>>>>>>> Make_Pointer_Type(ty_idx_used),
>>>>>>>>> n_rhs);
>>>>>>>>> + ld_ofst += MTYPE_byte_size(mtype);
>>>>>>>>> n_rhs = lower_expr(block, n_rhs, actions);
>>>>>>>>> preg = RETURN_INFO_preg (return_info, i);
>>>>>>>>> preg_st = Standard_Preg_For_Mtype(mtype);
>>>>>>>>> @@ -12357,6 +12362,7 @@
>>>>>>>>> Is_True(WN_operator(WN_kid1(o_rhs)) == OPR_INTCONST,
>>>>>>>>> ("expected RETURN_VAL's MLOAD kid to be of constant
>>>>>>>>> size"));
>>>>>>>>> algn = TY_align(TY_pointed(WN_load_addr_ty(o_rhs)));
>>>>>>>>> + WN_OFFSET ld_ofst = WN_load_offset(o_rhs);
>>>>>>>>> for (i = 0; i < RETURN_INFO_count(return_info); i++) {
>>>>>>>>> mtype = RETURN_INFO_mtype(return_info, i);
>>>>>>>>> ty_idx_used = Be_Type_Tbl(mtype);
>>>>>>>>> @@ -12365,10 +12371,10 @@
>>>>>>>>> n_rhs = WN_kid0(o_rhs);
>>>>>>>>> else n_rhs = WN_COPY_Tree(WN_kid0(o_rhs));
>>>>>>>>> n_rhs = WN_CreateIload(OPR_ILOAD, mtype, mtype,
>>>>>>>>> - WN_load_offset(o_rhs)
>>>>>>>>> - + i * MTYPE_byte_size(mtype),
>>>>>>>>> + ld_ofst,
>>>>>>>>> ty_idx_used,
>>>>>>>>> Make_Pointer_Type(ty_idx_used),
>>>>>>>>> n_rhs);
>>>>>>>>> + ld_ofst += MTYPE_byte_size(mtype);
>>>>>>>>> n_rhs = lower_expr(block, n_rhs, actions);
>>>>>>>>> preg = RETURN_INFO_preg (return_info, i);
>>>>>>>>> preg_st = Standard_Preg_For_Mtype(mtype);
>>>>>>>>> Index: osprey/common/com/x8664/targ_sim.cxx
>>>>>>>>> ===================================================================
>>>>>>>>> --- osprey/common/com/x8664/targ_sim.cxx (revision 3998)
>>>>>>>>> +++ osprey/common/com/x8664/targ_sim.cxx (working copy)
>>>>>>>>> @@ -673,13 +673,13 @@
>>>>>>>>> }
>>>>>>>>> else {
>>>>>>>>> if (classes[0] == X86_64_SSE_CLASS) {
>>>>>>>>> - info.mtype[0] = SIM_INFO.dbl_type;
>>>>>>>>> + info.mtype[0] = (size > 4) ? MTYPE_F8 : MTYPE_F4;
>>>>>>>>> info.preg[0] = PR_first_reg(SIM_INFO.dbl_results);
>>>>>>>>> next_float_return_num =
>>>>>>>>> PR_last_reg(SIM_INFO.dbl_results);
>>>>>>>>> next_int_return_num =
>>>>>>>>> PR_first_reg(SIM_INFO.int_results);
>>>>>>>>> }
>>>>>>>>> else {
>>>>>>>>> - info.mtype[0] = SIM_INFO.int_type;
>>>>>>>>> + info.mtype[0] = (size > 4) ? MTYPE_I8 : MTYPE_I4;
>>>>>>>>> info.preg[0] = PR_first_reg(SIM_INFO.int_results);
>>>>>>>>> next_float_return_num =
>>>>>>>>> PR_first_reg(SIM_INFO.dbl_results);
>>>>>>>>> next_int_return_num =
>>>>>>>>> PR_last_reg(SIM_INFO.int_results);
>>>>>>>>> @@ -687,11 +687,11 @@
>>>>>>>>>
>>>>>>>>> if (n > 1) {
>>>>>>>>> if (classes[1] == X86_64_SSE_CLASS) {
>>>>>>>>> - info.mtype[1] = SIM_INFO.dbl_type;
>>>>>>>>> + info.mtype[1] = (size > 12) ? MTYPE_F8 : MTYPE_F4;
>>>>>>>>> info.preg[1] = next_float_return_num;
>>>>>>>>> }
>>>>>>>>> else {
>>>>>>>>> - info.mtype[1] = SIM_INFO.int_type;
>>>>>>>>> + info.mtype[1] = (size > 12) ? MTYPE_I8 : MTYPE_I4;
>>>>>>>>> info.preg[1] = next_int_return_num;
>>>>>>>>> }
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> --
>>>>>>>>> Regards,
>>>>>>>>> Lai Jian-Xin
>>>>>>>>>
>>>>>>>>> ------------------------------------------------------------------------------
>>>>>>>>> Live Security Virtual Conference
>>>>>>>>> Exclusive live event will cover all the ways today's security and
>>>>>>>>> threat landscape has changed and how IT managers can respond.
>>>>>>>>> Discussions
>>>>>>>>> will include endpoint security, mobile security and the latest in
>>>>>>>>> malware
>>>>>>>>> threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/
>>>>>>>>> _______________________________________________
>>>>>>>>> Open64-devel mailing list
>>>>>>>>> Open64-devel@lists.sourceforge.net
>>>>>>>>> https://lists.sourceforge.net/lists/listinfo/open64-devel
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> --
>>>>>>> Regards,
>>>>>>> Lai Jian-Xin
>>>>>
>>>>>
>>>>>
>>>>> --
>>>>> Regards,
>>>>> Lai Jian-Xin
>>>
>>>
>>>
>>> --
>>> Regards,
>>> Lai Jian-Xin
>>
>>
>>
>> --
>> Regards,
>> Lai Jian-Xin
--
Regards,
Lai Jian-Xin
------------------------------------------------------------------------------
Live Security Virtual Conference
Exclusive live event will cover all the ways today's security and
threat landscape has changed and how IT managers can respond. Discussions
will include endpoint security, mobile security and the latest in malware
threats. http://www.accelacomm.com/jaw/sfrnl04242012/114/50122263/
_______________________________________________
Open64-devel mailing list
Open64-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/open64-devel
