Hi!
The following patch fixes two issues in the new rpad pass.
One is that the insertion at the start of a basic block didn't work properly
if the basic block didn't contain any non-NOTE/non-DEBUG_INSN instructions.
next_nonnote_nondebug_insn hapilly turns through into another basic block
and the insertion can insert an instruction in between basic blocks or into
a different basic block from where we wanted to emit it.
I believe we want to emit it after CODE_LABEL, after NOTE_INSN_BASIC_BLOCK
and if possible, after debug insns in there, the patch emits it before the
first normal insn in the bb if any, or after the BB_END (thus extending
BB_END).
Another issue is that it is quite weird/dangerous to add the v4sf_const0
pseudo uses in lots of places in the IL, register those changes with df,
then do df_analyze with different flags and finally emit the setter.
I understand the goal was not to do df_analyze etc. in the usual case where
there are no instructions that need this treatment. This patch does the
df_analyze at the spot we find the first insn, but before we actually change
that instruction, so the changes are after the df_analyze.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2019-02-25 Jakub Jelinek <ja...@redhat.com>
PR target/89474
* config/i386/i386.c (remove_partial_avx_dependency): Call
df_analyze etc. before creation of the v4sf_const0 pseudo, rather than
after changing possibly many instructions to use that pseudo. Fix up
insertion of v4sf_const0 setter at the start of bb.
* gcc.target/i386/pr89474.c: New test.
--- gcc/config/i386/i386.c.jj 2019-02-22 23:02:47.805117610 +0100
+++ gcc/config/i386/i386.c 2019-02-25 14:20:05.793608879 +0100
@@ -2835,7 +2835,14 @@ remove_partial_avx_dependency (void)
continue;
if (!v4sf_const0)
- v4sf_const0 = gen_reg_rtx (V4SFmode);
+ {
+ calculate_dominance_info (CDI_DOMINATORS);
+ df_set_flags (DF_DEFER_INSN_RESCAN);
+ df_chain_add_problem (DF_DU_CHAIN | DF_UD_CHAIN);
+ df_md_add_problem ();
+ df_analyze ();
+ v4sf_const0 = gen_reg_rtx (V4SFmode);
+ }
/* Convert PARTIAL_XMM_UPDATE_TRUE insns, DF -> SF, SF -> DF,
SI -> SF, SI -> DF, DI -> SF, DI -> DF, to vec_dup and
@@ -2883,12 +2890,6 @@ remove_partial_avx_dependency (void)
if (v4sf_const0)
{
- calculate_dominance_info (CDI_DOMINATORS);
- df_set_flags (DF_DEFER_INSN_RESCAN);
- df_chain_add_problem (DF_DU_CHAIN | DF_UD_CHAIN);
- df_md_add_problem ();
- df_analyze ();
-
/* (Re-)discover loops so that bb->loop_father can be used in the
analysis below. */
loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
@@ -2904,11 +2905,23 @@ remove_partial_avx_dependency (void)
bb = get_immediate_dominator (CDI_DOMINATORS,
bb->loop_father->header);
- insn = BB_HEAD (bb);
- if (!NONDEBUG_INSN_P (insn))
- insn = next_nonnote_nondebug_insn (insn);
set = gen_rtx_SET (v4sf_const0, CONST0_RTX (V4SFmode));
- set_insn = emit_insn_before (set, insn);
+
+ insn = BB_HEAD (bb);
+ while (insn && !NONDEBUG_INSN_P (insn))
+ {
+ if (insn == BB_END (bb))
+ {
+ insn = NULL;
+ break;
+ }
+ insn = NEXT_INSN (insn);
+ }
+ if (insn == BB_HEAD (bb))
+ set_insn = emit_insn_before (set, insn);
+ else
+ set_insn = emit_insn_after (set,
+ insn ? PREV_INSN (insn) : BB_END (bb));
df_insn_rescan (set_insn);
df_process_deferred_rescans ();
loop_optimizer_finalize ();
--- gcc/testsuite/gcc.target/i386/pr89474.c.jj 2019-02-25 14:21:51.651867104
+0100
+++ gcc/testsuite/gcc.target/i386/pr89474.c 2019-02-25 14:21:34.373151405
+0100
@@ -0,0 +1,14 @@
+/* PR target/89474 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx" } */
+
+int a;
+void foo (double);
+int baz (void);
+
+void
+bar (void)
+{
+ while (baz ())
+ foo (a);
+}
Jakub
