| Message ID | 201103232154.28597.ebotcazou@adacore.com |
|---|---|
| State | New |
| Headers | show |
On Wed, Mar 23, 2011 at 9:54 PM, Eric Botcazou <ebotcazou@adacore.com> wrote: > Hi, > > this is a fallout of one of my gimplifier changes. For the attached testcase, > the compiler generates invalid GIMPLE at -O1 or above: > > p.adb: In function 'P.F': > p.adb:24:1: error: invalid operand in unary operation > D.2361_107 = (integer[1:1] *) &*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4}; > > if (!is_gimple_val (rhs1)) > { > error ("invalid operand in unary operation"); > return true; > } > > (gdb) p rhs_code > $6 = NOP_EXPR > (gdb) p debug_generic_expr(rhs1) > &*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4} > > i.e ADDR_EXPR of ARRAY_RANGE_REF with non-constant operand #1. > > The invalid statement is generated by the early SRA pass: > > integer a1[1:1]; > > VIEW_CONVERT_EXPR<integer[1:1]>(*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4}) = > a1; > > is rewritten into: > > Created a replacement for a1 offset: 0, size: 32: a1$1 > > D.2361_107 = (integer[1:1] *) &*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4}; > MEM[(integer[1:D.2312] *)D.2361_107] = a1$1_92; > > > It turns out that the problematic NOP_EXPR is a useless type conversion emitted > by build_fold_addr_expr: if you add a call to force_gimple_operand at the hot > spot, it ends up removing the NOP_EXPR and nothing else. So the proposed fix > is just to add a call to STRIP_USELESS_TYPE_CONVERSION there. > > Tested on i586-suse-linux, OK for the mainline? Ok. At some point I hope we get gimple variants for most of the tree building helpers (that build_fold_addr_expr, and also fold_convert). Thanks, Richard. > > 2011-03-23 Eric Botcazou <ebotcazou@adacore.com> > > * tree-sra.c (build_ref_for_offset): Call STRIP_USELESS_TYPE_CONVERSION > on the address built for a reference with variable offset. > > > 2011-03-23 Eric Botcazou <ebotcazou@adacore.com> > > * gnat.dg/array15.ad[sb]: New test. > * gnat.dg/array15_pkg.ads: New helper. > > > -- > Eric Botcazou >
Index: tree-sra.c =================================================================== --- tree-sra.c (revision 171345) +++ tree-sra.c (working copy) @@ -1376,6 +1376,7 @@ build_ref_for_offset (location_t loc, tr add_referenced_var (tmp); tmp = make_ssa_name (tmp, NULL); addr = build_fold_addr_expr (unshare_expr (prev_base)); + STRIP_USELESS_TYPE_CONVERSION (addr); stmt = gimple_build_assign (tmp, addr); gimple_set_location (stmt, loc); SSA_NAME_DEF_STMT (tmp) = stmt;
Hi, this is a fallout of one of my gimplifier changes. For the attached testcase, the compiler generates invalid GIMPLE at -O1 or above: p.adb: In function 'P.F': p.adb:24:1: error: invalid operand in unary operation D.2361_107 = (integer[1:1] *) &*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4}; if (!is_gimple_val (rhs1)) { error ("invalid operand in unary operation"); return true; } (gdb) p rhs_code $6 = NOP_EXPR (gdb) p debug_generic_expr(rhs1) &*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4} i.e ADDR_EXPR of ARRAY_RANGE_REF with non-constant operand #1. The invalid statement is generated by the early SRA pass: integer a1[1:1]; VIEW_CONVERT_EXPR<integer[1:1]>(*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4}) = a1; is rewritten into: Created a replacement for a1 offset: 0, size: 32: a1$1 D.2361_107 = (integer[1:1] *) &*S15b.16_56[D.2327_91 ...]{lb: 1 sz: 4}; MEM[(integer[1:D.2312] *)D.2361_107] = a1$1_92; It turns out that the problematic NOP_EXPR is a useless type conversion emitted by build_fold_addr_expr: if you add a call to force_gimple_operand at the hot spot, it ends up removing the NOP_EXPR and nothing else. So the proposed fix is just to add a call to STRIP_USELESS_TYPE_CONVERSION there. Tested on i586-suse-linux, OK for the mainline? 2011-03-23 Eric Botcazou <ebotcazou@adacore.com> * tree-sra.c (build_ref_for_offset): Call STRIP_USELESS_TYPE_CONVERSION on the address built for a reference with variable offset. 2011-03-23 Eric Botcazou <ebotcazou@adacore.com> * gnat.dg/array15.ad[sb]: New test. * gnat.dg/array15_pkg.ads: New helper.