Message ID | f0e99301-5101-1d90-de69-d9cb8eff9e2d@redhat.com |
---|---|
State | New |
Headers | show |
Series | [COMMITTED] patch for PR95464 | expand |
On Thu, Jun 04, 2020 at 12:13:51PM -0400, Vladimir Makarov via Gcc-patches wrote: > The following patch fixes > >           https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95464 > > The patch was successfully bootstrapped and tested on x86-64. I did not > test it on other major targets as they do not deal with STRICT_LOW_PART. Thanks. > There is also no test case for this PR as it is hard to check the right code > generation. The testcase from the PR was a dg-do run testcase that FAILed without your patch, can't we just use that testcase (in gcc.target/i386/, restricted to lp64 and possibly linux only to make sure frame pointer can be safely omitted)? Jakub
On 2020-06-04 12:17 p.m., Jakub Jelinek wrote: > On Thu, Jun 04, 2020 at 12:13:51PM -0400, Vladimir Makarov via Gcc-patches wrote: >> The following patch fixes >> >> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95464 >> >> The patch was successfully bootstrapped and tested on x86-64. I did not >> test it on other major targets as they do not deal with STRICT_LOW_PART. > Thanks. > >> There is also no test case for this PR as it is hard to check the right code >> generation. > The testcase from the PR was a dg-do run testcase that FAILed without your > patch, can't we just use that testcase (in gcc.target/i386/, restricted to > lp64 and possibly linux only to make sure frame pointer can be safely > omitted)? > > OK, Jakub. I'll try to create a test for the PR.
On Thu, Jun 4, 2020 at 9:17 AM Jakub Jelinek via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > On Thu, Jun 04, 2020 at 12:13:51PM -0400, Vladimir Makarov via Gcc-patches wrote: > > The following patch fixes > > > > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95464 > > > > The patch was successfully bootstrapped and tested on x86-64. I did not > > test it on other major targets as they do not deal with STRICT_LOW_PART. > > Thanks. > > > There is also no test case for this PR as it is hard to check the right code > > generation. > > The testcase from the PR was a dg-do run testcase that FAILed without your > patch, can't we just use that testcase (in gcc.target/i386/, restricted to > lp64 and possibly linux only to make sure frame pointer can be safely > omitted)? Why lp64? Shouldn't it be !ia32?
On 2020-06-04 12:55 p.m., H.J. Lu wrote: > On Thu, Jun 4, 2020 at 9:17 AM Jakub Jelinek via Gcc-patches > <gcc-patches@gcc.gnu.org> wrote: >> >> The testcase from the PR was a dg-do run testcase that FAILed without your >> patch, can't we just use that testcase (in gcc.target/i386/, restricted to >> lp64 and possibly linux only to make sure frame pointer can be safely >> omitted)? > Why lp64? Shouldn't it be !ia32? > Thank you, H.J. I'll use ! ia32.
Add processing STRICT_LOW_PART for matched reloads. 2020-06-04 Vladimir Makarov <vmakarov@redhat.com> PR middle-end/95464 * lra.c (lra_emit_move): Add processing STRICT_LOW_PART. * lra-constraints.c (match_reload): Use STRICT_LOW_PART in output reload if the original insn has it too. diff --git a/gcc/lra-constraints.c b/gcc/lra-constraints.c index bf6d4a2fd4b..421c453997b 100644 --- a/gcc/lra-constraints.c +++ b/gcc/lra-constraints.c @@ -1071,6 +1071,8 @@ match_reload (signed char out, signed char *ins, signed char *outs, if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX) { start_sequence (); + if (out >= 0 && curr_static_id->operand[out].strict_low) + out_rtx = gen_rtx_STRICT_LOW_PART (VOIDmode, out_rtx); lra_emit_move (out_rtx, copy_rtx (new_out_reg)); emit_insn (*after); *after = get_insns (); diff --git a/gcc/lra.c b/gcc/lra.c index 3435cff6a1d..caa09d86ca6 100644 --- a/gcc/lra.c +++ b/gcc/lra.c @@ -490,13 +490,16 @@ void lra_emit_move (rtx x, rtx y) { int old; - + rtx_insn *insn; + if (GET_CODE (y) != PLUS) { if (rtx_equal_p (x, y)) return; old = max_reg_num (); - rtx_insn *insn = emit_move_insn (x, y); + + insn = (GET_CODE (x) != STRICT_LOW_PART + ? emit_move_insn (x, y) : emit_insn (gen_rtx_SET (x, y))); /* The move pattern may require scratch registers, so convert them into real registers now. */ if (insn != NULL_RTX)