| Message ID | 7c9ac3d29a29c4b1bc78294afb955f30fd36ad77.1455137361.git.segher@kernel.crashing.org |
|---|---|
| State | New |
| Headers | show |
On Thu, Feb 11, 2016 at 9:04 AM, Segher Boessenkool <segher@kernel.crashing.org> wrote: > PR64682 is a problem in distribute_notes, where it has trouble putting > a REG_DEAD note for a reg that is set twice in the right spot. My fix > for that did the wrong thing for PR69567. And then my attempted fix > for that one made PR64682 fail again. > > Instead, let's just lose the note in such complicated cases, like we > already do in certain similar cases. > > Tested on powerpc64-linux and x86_64-linux. Also built Linux kernels > for some 30 supported targets; no difference in generated code was > observed. > > Committing to trunk. > > HJ, I tested this on GCC 5 for x86_64-linux, the failure is gone; > could you test it on your setup before I apply it there though? > > > Segher > > > 2016-02-11 Segher Boessenkool <segher@kernel.crashing.org> > > PR rtl-optimization/64682 > PR rtl-optimization/69567 > PR rtl-optimization/69737 > * combine.c (distribute_notes) <REG_DEAD>: If the register is set > in I2 as well, just lose it. > > Yes, it fixed the regressions on ia32 and x96-64: New passes: FAIL: gcc.c-torture/execute/pr64682.c -O2 execution test FAIL: gcc.c-torture/execute/pr64682.c -O2 execution test FAIL: gcc.c-torture/execute/pr64682.c -O2 -flto -fno-use-linker-plugin -flto-partition=none execution test FAIL: gcc.c-torture/execute/pr64682.c -O2 -flto -fno-use-linker-plugin -flto-partition=none execution test FAIL: gcc.c-torture/execute/pr64682.c -O2 -flto -fuse-linker-plugin -fno-fat-lto-objects execution test FAIL: gcc.c-torture/execute/pr64682.c -O2 -flto -fuse-linker-plugin -fno-fat-lto-objects execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -fomit-frame-pointer execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -fomit-frame-pointer execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -fomit-frame-pointer -funroll-all-loops -finline-functions execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -fomit-frame-pointer -funroll-all-loops -finline-functions execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -fomit-frame-pointer -funroll-loops execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -fomit-frame-pointer -funroll-loops execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -g execution test FAIL: gcc.c-torture/execute/pr64682.c -O3 -g execution test Thanks.
diff --git a/gcc/combine.c b/gcc/combine.c index 3609b94..24dcefa 100644 --- a/gcc/combine.c +++ b/gcc/combine.c @@ -13901,7 +13901,6 @@ distribute_notes (rtx notes, rtx_insn *from_insn, rtx_insn *i3, rtx_insn *i2, tem_insn = from_insn; else { - tem_insn = i3; if (from_insn && CALL_P (from_insn) && find_reg_fusage (from_insn, USE, XEXP (note, 0))) @@ -13910,14 +13909,7 @@ distribute_notes (rtx notes, rtx_insn *from_insn, rtx_insn *i3, rtx_insn *i2, place = i3; else if (i2 != 0 && next_nonnote_nondebug_insn (i2) == i3 && reg_referenced_p (XEXP (note, 0), PATTERN (i2))) - { - place = i2; - /* If the new I2 sets the same register that is marked dead - in the note, the note now should not be put on I2, as the - note refers to a previous incarnation of the reg. */ - if (reg_set_p (XEXP (note, 0), PATTERN (i2))) - tem_insn = i2; - } + place = i2; else if ((rtx_equal_p (XEXP (note, 0), elim_i2) && !(i2mod && reg_overlap_mentioned_p (XEXP (note, 0), @@ -13925,6 +13917,12 @@ distribute_notes (rtx notes, rtx_insn *from_insn, rtx_insn *i3, rtx_insn *i2, || rtx_equal_p (XEXP (note, 0), elim_i1) || rtx_equal_p (XEXP (note, 0), elim_i0)) break; + tem_insn = i3; + /* If the new I2 sets the same register that is marked dead + in the note, we do not know where to put the note. + Give up. */ + if (i2 != 0 && reg_set_p (XEXP (note, 0), PATTERN (i2))) + break; } if (place == 0)