Fix missing TOC loads on ppc64el

Bug: https://gitlab.com/freepascal.org/fpc/source/-/issues/39542
Origin: vendor, https://src.fedoraproject.org/rpms/fpc/c/01634e7f70132cc0bc1cab278edc8449e73ac1f9
Last-Update: 2022-02-07

A backport of upstream commits:
- https://gitlab.com/freepascal.org/fpc/source/-/commit/12f48c230bccd49f368be1e5a2855cb6c3a60c0f
- https://gitlab.com/freepascal.org/fpc/source/-/commit/9314bbbf080418827eef94a8bc392ce0497bf72b
- https://gitlab.com/freepascal.org/fpc/source/-/commit/2de72c854115908271912cd9b260a607c826eadb
- https://gitlab.com/freepascal.org/fpc/source/-/commit/83c18df69a79fe1035a0cf0bc0897c60d1af0293
- https://gitlab.com/freepascal.org/fpc/source/-/commit/68b5ca633ca71a83c29b78cd3669bf15477cd94f
Some modifications were made to make the changes apply cleanly to v3.2.2.

Gbp-Pq: Name ppc64el-toc-fixes.patch

diff --git a/fpcsrc/compiler/ncgvmt.pas b/fpcsrc/compiler/ncgvmt.pas
index ab93c7fe07a5c6a29dc947c2d12493d8389f20f5..bd7fd46a7eb1f4b5f6ea12e55dfab568aa2da4a2 100644 (file)
--- a/fpcsrc/compiler/ncgvmt.pas
+++ b/fpcsrc/compiler/ncgvmt.pas
@@ -1222,14 +1222,25 @@ implementation
         tmps : string;
         pd   : TProcdef;
         ImplIntf : TImplementedInterface;
-{$ifdef cpuhighleveltarget}
         wrapperpd: tprocdef;
         wrapperinfo: pskpara_interface_wrapper;
-{$else}
-       tmplist: tasmlist;
-       oldfileposinfo: tfileposinfo;
-{$endif cpuhighleveltarget}
+        tmplist: tasmlist;
+        oldfileposinfo: tfileposinfo;
+        usehighlevelwrapper: Boolean;
       begin
+{$if defined(cpuhighleveltarget)}
+        usehighlevelwrapper:=true;
+{$else defined(cpuhighleveltarget)}
+        { on PPC systems that use a TOC the linker needs to be able to insert
+          an instruction to restore the TOC register after every branch
+          between code fragments that use a different TOC (which has to be
+          executed when that "branch" returns). So we can't use tail call
+          branches to routines potentially using a different TOC there }
+        if target_info.system in systems_ppc_toc then
+          usehighlevelwrapper:=true
+        else
+          usehighlevelwrapper:=false;
+{$endif defined(cpuhighleveltarget)}
         for i:=0 to _class.ImplementedInterfaces.count-1 do
           begin
             ImplIntf:=TImplementedInterface(_class.ImplementedInterfaces[i]);
@@ -1246,43 +1257,47 @@ implementation
                        not is_objectpascal_helper(tprocdef(pd).struct) then
                       tobjectdef(tprocdef(pd).struct).register_vmt_call(tprocdef(pd).extnumber);
                     tmps:=CreateWrapperName(_Class,ImplIntf,j,pd);
-{$ifdef cpuhighleveltarget}
-                    new(wrapperinfo);
-                    wrapperinfo^.pd:=pd;
-                    wrapperinfo^.offset:=ImplIntf.ioffset;
-                    { insert the wrapper procdef in the current unit's local
-                      symbol table, but set the owning "struct" to the current
-                      class (so self will have the correct type) }
-                    wrapperpd:=create_procdef_alias(pd,tmps,tmps,
-                      current_module.localsymtable,_class,
-                      tsk_interface_wrapper,wrapperinfo);
-                    include(wrapperpd.procoptions,po_noreturn);
-{$else cpuhighleveltarget}
-                    oldfileposinfo:=current_filepos;
-                    if pd.owner.iscurrentunit then
-                      current_filepos:=pd.fileinfo
-                    else
-                      begin
-                        current_filepos.moduleindex:=current_module.unit_index;
-                        current_filepos.fileindex:=1;
-                        current_filepos.line:=1;
-                        current_filepos.column:=1;
-                      end;
-                    { create wrapper code }
-                    tmplist:=tasmlist.create;
-                    new_section(tmplist,sec_code,tmps,target_info.alignment.procalign);
-                    tmplist.Concat(tai_function_name.create(tmps));
-                    hlcg.init_register_allocators;
-                    hlcg.g_intf_wrapper(tmplist,pd,tmps,ImplIntf.ioffset);
-                    hlcg.done_register_allocators;
-                    if ((cs_debuginfo in current_settings.moduleswitches) or
-                       (cs_use_lineinfo in current_settings.globalswitches)) and
-                       (target_dbg.id<>dbg_stabx) then
-                         current_debuginfo.insertlineinfo(tmplist);
-                    list.concatlist(tmplist);
-                    tmplist.Free;
-                    current_filepos:=oldfileposinfo;
-{$endif cpuhighleveltarget}
+
+                  if usehighlevelwrapper then
+                    begin
+                      new(wrapperinfo);
+                      wrapperinfo^.pd:=pd;
+                      wrapperinfo^.offset:=ImplIntf.ioffset;
+                      { insert the wrapper procdef in the current unit's local
+                        symbol table, but set the owning "struct" to the current
+                        class (so self will have the correct type) }
+                      wrapperpd:=create_procdef_alias(pd,tmps,tmps,
+                        current_module.localsymtable,_class,
+                        tsk_interface_wrapper,wrapperinfo);
+                      include(wrapperpd.implprocoptions,pio_thunk);
+                    end
+                  else
+                    begin
+                      oldfileposinfo:=current_filepos;
+                      if pd.owner.iscurrentunit then
+                        current_filepos:=pd.fileinfo
+                      else
+                        begin
+                          current_filepos.moduleindex:=current_module.unit_index;
+                          current_filepos.fileindex:=1;
+                          current_filepos.line:=1;
+                          current_filepos.column:=1;
+                        end;
+                      { create wrapper code }
+                      tmplist:=tasmlist.create;
+                      new_section(tmplist,sec_code,tmps,target_info.alignment.procalign);
+                      tmplist.Concat(tai_function_name.create(tmps));
+                      hlcg.init_register_allocators;
+                      hlcg.g_intf_wrapper(tmplist,pd,tmps,ImplIntf.ioffset);
+                      hlcg.done_register_allocators;
+                      if ((cs_debuginfo in current_settings.moduleswitches) or
+                         (cs_use_lineinfo in current_settings.globalswitches)) and
+                         (target_dbg.id<>dbg_stabx) then
+                           current_debuginfo.insertlineinfo(tmplist);
+                      list.concatlist(tmplist);
+                      tmplist.Free;
+                      current_filepos:=oldfileposinfo;
+                    end;
                   end;
               end;
           end;

diff --git a/fpcsrc/compiler/powerpc64/cgcpu.pas b/fpcsrc/compiler/powerpc64/cgcpu.pas
index 9b0450614af8e8414267394205a5f7f3c2c5bc0c..0b7a58c1fcc3cb3e587cce1c302a50cce92e2df9 100644 (file)
--- a/fpcsrc/compiler/powerpc64/cgcpu.pas
+++ b/fpcsrc/compiler/powerpc64/cgcpu.pas
@@ -337,10 +337,6 @@ begin
     reference_reset_base(tmpref, reg, 0, ctempposinvalid, sizeof(pint), []);
     a_load_ref_reg(list, OS_ADDR, OS_ADDR, tmpref, tempreg);
 
-    { save TOC pointer in stackframe }
-    reference_reset_base(tmpref, NR_STACK_POINTER_REG, get_rtoc_offset, ctempposinvalid, 8, []);
-    a_load_reg_ref(list, OS_ADDR, OS_ADDR, NR_RTOC, tmpref);
-
     { move actual function pointer to CTR register }
     list.concat(taicpu.op_reg(A_MTCTR, tempreg));
 
@@ -1269,6 +1265,13 @@ begin
     end;
   end;
 
+  { save current RTOC for restoration after calls if necessary }
+  if pi_do_call in current_procinfo.flags then
+    begin
+      reference_reset_base(href,NR_STACK_POINTER_REG,get_rtoc_offset,ctempposinvalid,target_info.stackalign,[]);
+      a_load_reg_ref(list,OS_ADDR,OS_ADDR,NR_RTOC,href);
+    end;
+
   { CR register not used by FPC atm }
 
   { keep R1 allocated??? }

diff --git a/fpcsrc/compiler/ppcgen/cgppc.pas b/fpcsrc/compiler/ppcgen/cgppc.pas
index 35e0d7b1cfcf5bbcb3626558390ab15213e4dbdf..a71acf10f3a1995828a4b3ff18331012e5a414e6 100644 (file)
--- a/fpcsrc/compiler/ppcgen/cgppc.pas
+++ b/fpcsrc/compiler/ppcgen/cgppc.pas
@@ -1055,6 +1055,7 @@ unit cgppc;
            (assigned(ref.symbol) and
             not assigned(ref.relsymbol)) then
           begin
+            include(current_procinfo.flags,pi_needs_got);
             tmpreg := load_got_symbol(list, ref.symbol.name, asmsym2indsymflags(ref.symbol));
             if (ref.base = NR_NO) then
               ref.base := tmpreg

diff --git a/fpcsrc/compiler/ppcgen/hlcgppc.pas b/fpcsrc/compiler/ppcgen/hlcgppc.pas
index 4d747b3fcdd509a8e898b01d7cc60d302b6dafdc..2a991708d5008bfb43d3ba77193d348aa33f0f5b 100644 (file)
--- a/fpcsrc/compiler/ppcgen/hlcgppc.pas
+++ b/fpcsrc/compiler/ppcgen/hlcgppc.pas
@@ -183,11 +183,16 @@ implementation
           system_powerpc_darwin,
           system_powerpc64_darwin:
             list.concat(taicpu.op_sym(A_B,tcgppcgen(cg).get_darwin_call_stub(procdef.mangledname,false)));
-          else if use_dotted_functions then
-            {$note ts:todo add GOT change?? - think not needed :) }
-            list.concat(taicpu.op_sym(A_B,current_asmdata.RefAsmSymbol('.' + procdef.mangledname,AT_FUNCTION)))
           else
-            list.concat(taicpu.op_sym(A_B,current_asmdata.RefAsmSymbol(procdef.mangledname,AT_FUNCTION)))
+            begin
+              if use_dotted_functions then
+                {$note ts:todo add GOT change?? - think not needed :) }
+                list.concat(taicpu.op_sym(A_B,current_asmdata.RefAsmSymbol('.' + procdef.mangledname,AT_FUNCTION)))
+              else
+                list.concat(taicpu.op_sym(A_B,current_asmdata.RefAsmSymbol(procdef.mangledname,AT_FUNCTION)));
+              if (target_info.system in ([system_powerpc64_linux]+systems_aix)) then
+                list.concat(taicpu.op_none(A_NOP));
+            end;
         end;
       List.concat(Tai_symbol_end.Createname(labelname));
     end;

diff --git a/fpcsrc/compiler/psub.pas b/fpcsrc/compiler/psub.pas
index 6d04c694251c0e2d96c061acfc2951b2d44d0628..40d8419dbc8a911a75553777a3d4a4e2c4d13ec9 100644 (file)
--- a/fpcsrc/compiler/psub.pas
+++ b/fpcsrc/compiler/psub.pas
@@ -2306,8 +2306,10 @@ implementation
                  }
                  if (not pd.forwarddef) and
                     (pd.hasforward) and
-                    (proc_get_importname(pd)<>'') then
-                   call_through_new_name(pd,proc_get_importname(pd))
+                    (proc_get_importname(pd)<>'') then begin
+                     call_through_new_name(pd,proc_get_importname(pd))
+                     include(pd.implprocoptions,pio_thunk);
+                   end
                  else
 {$endif cpuhighleveltarget}
                    begin

diff --git a/fpcsrc/compiler/symconst.pas b/fpcsrc/compiler/symconst.pas
index fdc7a2007fed1782c0de2c5d129e98994864a3c0..865298730e65b8e40e7a8fdc89abbb6027ac725d 100644 (file)
--- a/fpcsrc/compiler/symconst.pas
+++ b/fpcsrc/compiler/symconst.pas
@@ -425,7 +425,17 @@ type
     { the routine contains no code }
     pio_empty,
     { the inline body of this routine is available }
-    pio_has_inlininginfo
+    pio_has_inlininginfo,
+    { inline is not possible (has assembler block, etc) }
+    pio_inline_not_possible,
+    { a nested routine accesses a local variable from this routine }
+    pio_nested_access,
+    { a stub/thunk }
+    pio_thunk,
+    { compiled with fastmath enabled } 
+    pio_fastmath, 
+    { inline is forbidden (calls get_frame) }
+    pio_inline_forbidden
   );
   timplprocoptions = set of timplprocoption;
 

diff --git a/fpcsrc/compiler/systems.pas b/fpcsrc/compiler/systems.pas
index e1cce1d5716fadf037b565898a780413843aae17..c322c60e65ad8e03f5d3feaa2660a80f498aaa55 100644 (file)
--- a/fpcsrc/compiler/systems.pas
+++ b/fpcsrc/compiler/systems.pas
@@ -393,6 +393,16 @@ interface
          on the caller side rather than on the callee side }
        systems_caller_copy_addr_value_para = [system_aarch64_ios,system_aarch64_darwin,system_aarch64_linux];
 
+       { all PPC systems that use a TOC register to address globals }
+       { TODO: not used by Darwin, but don't know about others (JM) }
+       systems_ppc_toc = [
+         system_powerpc_linux,
+         system_powerpc64_linux,
+         system_powerpc_aix,
+         system_powerpc64_aix,
+         system_powerpc_macosclassic
+       ];
+
        { pointer checking (requires special code in FPC_CHECKPOINTER,
          and can never work for libc-based targets or any other program
          linking to an external library)

diff --git a/fpcsrc/compiler/utils/ppuutils/ppudump.pp b/fpcsrc/compiler/utils/ppuutils/ppudump.pp
index 3418e628e0d99c8956118567d41bbe74a911ffa6..bb96d03914ef5306820119080bca0beda9031c68 100644 (file)
--- a/fpcsrc/compiler/utils/ppuutils/ppudump.pp
+++ b/fpcsrc/compiler/utils/ppuutils/ppudump.pp
@@ -3087,7 +3087,12 @@ type
 const
   piopt : array[low(timplprocoption)..high(timplprocoption)] of tpiopt=(
     (mask:pio_empty; str:'IsEmpty'),
-    (mask:pio_has_inlininginfo; str:'HasInliningInfo')
+    (mask:pio_has_inlininginfo; str:'HasInliningInfo'),
+    (mask:pio_inline_not_possible; str:'InlineNotPossible'),
+    (mask:pio_nested_access; str:'NestedAccess'),
+    (mask:pio_thunk; str:'Thunk'),
+    (mask:pio_fastmath; str:'FastMath'),
+    (mask:pio_inline_forbidden; str:'InlineForbidden')
   );
 var
   i: timplprocoption;