Diff for /wikisrc/users/maya/vax-gcc.mdwn between versions 1.1 and 1.12

version 1.1, 2019/08/29 12:43:34 version 1.12, 2019/09/15 12:56:38
Line 1 Line 1
 # Various GCC/vax issues and invesigation #  # Various GCC/vax issues and invesigation #
   
 # libgomp crash with -O2 #  [WIP tree](gcc-mirror/)
   
   ## libgomp crash with -O2 [#58901](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58901) ##
 Repro:  Repro:
 ````  
 #!/bin/sh  
   
 cat << EOF > oacc.i      #!/bin/sh
 int a, b;      
 int e() {      cat << EOF > oacc.i
   short *c = 0;      int a, b;
   char d = c[a + 1];      int e() {
   b = d;        short *c = 0;
   a += 2;        char d = c[a + 1];
 }        b = d;
 EOF        a += 2;
 vax--netbsdelf-gcc -O2 -c oacc.i      }
 ````      EOF
       vax--netbsdelf-gcc -O2 -c oacc.i
   
 Asserts at emit-rtl.c:2310 gcc_assert (memory_address_addr_space_p (mode, addr, as));  Asserts at emit-rtl.c:2310 gcc_assert (memory_address_addr_space_p (mode, addr, as));
   
 NetBSD avoids it with a local diff to reload.c/dse.c.  NetBSD avoids it with a [local diff](https://github.com/NetBSD/src/commit/e437e96750193b86d0464965661f616e011056fa) to reload.c/dse.c.  
   It looks like a real bug in vax_legitimate_address_p not handling some addressing modes.  
   The code to handle offset(reg)[index] might be off: it doesn't like if neither xfoo0, xfoo1 are constant.
   
   
   ## change_address_1, at emit-rtl.c:2286 [#85152](https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85152) ##
   
   Repro:
   
       #!/bin/sh
   
       cat << EOF > oacc.i
       void fn2();
       unsigned a;
       int fn1() {
         int b = a >> 16;
         if (b)
           fn2();
       }
       EOF
       vax--netbsdelf-gcc -O2 -c oacc.i
   
   ## unrecognizable insn with any optimization ##
   
   Repro:
   
       #!/bin/sh
   
       cat << EOF > decl.ii
       struct a {
         long long *b() const;
       };
       template <typename> class d : a {
       public:
         template <typename e> d(e);
         long long f() const;
       };
       template <typename c> long long d<c>::f() const {
         int g;
         return b()[g];
       }
       int h;
       bool i() {
         d<int> j(h);
         return j.f();
       }
       EOF
       vax--netbsdelf-g++ -O2 -c decl.ii
   
   This errors out with:
   
       $ env PATH=$PWD:$PATH ./xgcc -fno-use-linker-plugin -c -O2 ~/oacc/decl.ii -fno-tree-vectorize
       /home/fly/oacc/decl.ii: In function 'bool i()':
       /home/fly/oacc/decl.ii:17:1: error: unrecognizable insn:
          17 | }
             | ^
       (insn 13 12 14 2 (set (reg:SI 33)
               (subreg:SI (mem:DI (plus:SI (mult:SI (reg/v:SI 24 [ g ])
                               (const_int 8 [0x8]))
                           (reg/f:SI 23 [ _6 ])) [1 *_10+0 S8 A32]) 4)) "/home/fly/oacc/decl.ii":16:14 -1
            (nil))
       during RTL pass: vregs
       /home/fly/oacc/decl.ii:17:1: internal compiler error: in extract_insn, at recog.c:2310
       
   
   This is in VAX-specific code. We need to figure out what instruction pattern this is, that we're missing.
   
   <!--Providing a trivial scheduling will avoid subregs of mem.  
   However, this doesn't work for us. Our crash and mem subregs happen at a too early pass, most likely.-->
   
   ## DWARF ##
   
       flag_dwarf2_cfi_asm = 0;
   
   is kind of a big deal. We should probably get rid of it, but [binutils hates that.](https://github.com/NetBSD/src/commit/1df57dce3744cf4743358b098728a0658b80b8fb)
   
       ld: warning: dynamic relocation to `?' in readonly section `.eh_frame'
       ld: warning: dynamic relocation to `?' in readonly section `.eh_frame'
       ld: warning: creating a DT_TEXTREL in a shared object
       ld: .eh_frame_hdr refers to overlapping FDEs
       ld: final link failed: bad value
   
   [riscv seemed to share the issue](https://github.com/riscv/riscv-binutils-gdb/issues/76) and had fixed it.
   
   ## review comments ##
   
 It looks like a real bug in vax_legitimate_address_p not handling some addressing modes  Jeff Law [provided some feedback](https://gcc.gnu.org/ml/gcc/2019-04/msg00281.html). More of it should be addressed.

Removed from v.1.1  
changed lines
  Added in v.1.12


CVSweb for NetBSD wikisrc <wikimaster@NetBSD.org> software: FreeBSD-CVSweb