[gcc r14-8319] PR rtl-optimization/111267: Improved forward propagation.

[gcc r14-8319] PR rtl-optimization/111267: Improved forward propagation.

[Roger Sayle]

https://gcc.gnu.org/g:86de9b66480b710202a2898cf513db105d8c432f

commit r14-8319-g86de9b66480b710202a2898cf513db105d8c432f
Author: Roger Sayle <roger@nextmovesoftware.com>
Date:   Sun Jan 21 21:22:28 2024 +0000

    PR rtl-optimization/111267: Improved forward propagation.
    
    This patch resolves PR rtl-optimization/111267 by improving RTL-level
    forward propagation.  This x86_64 code quality regression was caused
    (exposed) by my changes to improve how x86's (TImode) argument passing
    is represented at the RTL-level (reducing the use of SUBREGs to catch
    more optimization opportunities in combine).  The pitfall is that the
    more complex RTL representations expose a limitation in RTL's fwprop
    pass.
    
    At the heart of fwprop, in try_fwprop_subst_pattern, the logic can
    be summarized as three steps.  Step 1 is a heuristic that rejects the
    propagation attempt if the expression is too complex, step 2 calls
    the backend's recog to see if the propagated/simplified instruction
    is recognizable/valid, and step 3 then calls set_src_cost to compare
    the rtx costs of the replacement vs. the original, and accepts the
    transformation if the final cost is the same of better.
    
    The logic error (or missed optimization opportunity) is that the
    step 1 heuristic that attempts to predict (second guess) the
    process is flawed.  Ultimately the decision on whether to fwprop
    or not should depend solely on actual improvement, as measured
    by RTX costs.  Hence the prototype fix in the bugzilla PR removes
    the heuristic of calling prop.profitable_p entirely, relying
    entirely on the cost comparison in step 3.
    
    Unfortunately, things are a tiny bit more complicated.  The cost
    comparison in fwprop uses the older set_src_cost API and not the
    newer (preffered) insn_cost API as currently used in combine.
    This means that the cost improvement comparisons are only done
    for single_set instructions (more complex PARALLELs etc. aren't
    supported).  Hence we can only rely on skipping step 1 for that
    subset of instructions actually evaluated by step 3.
    
    The other subtlety is that to avoid potential infinite loops
    in fwprop we should only reply purely on rtx costs when the
    transformation is obviously an improvement.  If the replacement
    has the same cost as the original, we can use the prop.profitable_p
    test to preserve the current behavior.
    
    Finally, to answer Richard Biener's remaining question about this
    approach: yes, there is an asymmetry between how patterns are
    handled and how REG_EQUAL notes are handled.  For example, at
    the moment propagation into notes doesn't use rtx costs at all,
    and ultimately when fwprop is updated to use insn_cost, this
    (and recog) obviously isn't applicable to notes.  There's no reason
    the logic need be identical between patterns and notes, and during
    stage4 we only need update propagation into patterns to fix this
    P1 regression (notes and use of cost_insn can be done for GCC 15).
    
    For Jakub's reduced testcase:
    
    struct S { float a, b, c, d; };
    int bar (struct S x, struct S y) {
      return x.b <= y.d && x.c >= y.a;
    }
    
    On x86_64-pc-linux-gnu with -O2 gcc currently generates:
    
    bar:    movq    %xmm2, %rdx
            movq    %xmm3, %rax
            movq    %xmm0, %rsi
            xchgq   %rdx, %rax
            movq    %rsi, %rcx
            movq    %rax, %rsi
            movq    %rdx, %rax
            shrq    $32, %rcx
            shrq    $32, %rax
            movd    %ecx, %xmm4
            movd    %eax, %xmm0
            comiss  %xmm4, %xmm0
            jb      .L6
            movd    %esi, %xmm0
            xorl    %eax, %eax
            comiss  %xmm0, %xmm1
            setnb   %al
            ret
    .L6:    xorl    %eax, %eax
            ret
    
    with this simple patch to fwprop, we now generate:
    
    bar:    shufps  $85, %xmm0, %xmm0
            shufps  $85, %xmm3, %xmm3
            comiss  %xmm0, %xmm3
            jb      .L6
            xorl    %eax, %eax
            comiss  %xmm2, %xmm1
            setnb   %al
            ret
    .L6:    xorl    %eax, %eax
            ret
    
    2024-01-21  Roger Sayle  <roger@nextmovesoftware.com>
                Richard Biener  <rguenther@suse.de>
    
    gcc/ChangeLog
            PR rtl-optimization/111267
            * fwprop.cc (fwprop_propagation::profitabe_p): Rename
            profitable_p method to likely_profitable_p.
            (try_fwprop_subst_node): Update call to likely_profitable_p.
            Only bail-out early when !prop.likely_profitable_p for instructions
            that are not single sets.  When comparing costs, bail-out if the
            cost is unchanged and !prop.likely_profitable_p.
    
    gcc/testsuite/ChangeLog
            PR rtl-optimization/111267
            * gcc.target/i386/pr111267.c: New test case.

Diff:
---
 gcc/fwprop.cc                            | 14 +++++++++-----
 gcc/testsuite/gcc.target/i386/pr111267.c | 14 ++++++++++++++
 2 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/gcc/fwprop.cc b/gcc/fwprop.cc
index 0c588f83325..7872609b336 100644
--- a/gcc/fwprop.cc
+++ b/gcc/fwprop.cc
@@ -180,7 +180,7 @@ namespace
 
     bool changed_mem_p () const { return result_flags & CHANGED_MEM; }
     bool folded_to_constants_p () const;
-    bool profitable_p () const;
+    bool likely_profitable_p () const;
 
     bool check_mem (int, rtx) final override;
     void note_simplification (int, uint16_t, rtx, rtx) final override;
@@ -323,7 +323,7 @@ fwprop_propagation::folded_to_constants_p () const
    false if it would increase the complexity of the pattern too much.  */
 
 bool
-fwprop_propagation::profitable_p () const
+fwprop_propagation::likely_profitable_p () const
 {
   if (changed_mem_p ())
     return true;
@@ -398,7 +398,7 @@ try_fwprop_subst_note (insn_info *use_insn, set_info *def,
     }
   else
     {
-      if (!prop.folded_to_constants_p () && !prop.profitable_p ())
+      if (!prop.folded_to_constants_p () && !prop.likely_profitable_p ())
 	{
 	  if (dump_file && (dump_flags & TDF_DETAILS))
 	    fprintf (dump_file, "cannot propagate from insn %d into"
@@ -449,7 +449,10 @@ try_fwprop_subst_pattern (obstack_watermark &attempt, insn_change &use_change,
   if (prop.num_replacements == 0)
     return false;
 
-  if (!prop.profitable_p ())
+  if (!prop.likely_profitable_p ()
+      && (prop.changed_mem_p ()
+	  || use_insn->is_asm ()
+	  || !single_set (use_rtl)))
     {
       if (dump_file && (dump_flags & TDF_DETAILS))
 	fprintf (dump_file, "cannot propagate from insn %d into"
@@ -481,7 +484,8 @@ try_fwprop_subst_pattern (obstack_watermark &attempt, insn_change &use_change,
 	redo_changes (0);
 	auto new_cost = set_src_cost (SET_SRC (use_set),
 				      GET_MODE (SET_DEST (use_set)), speed);
-	if (new_cost > old_cost)
+	if (new_cost > old_cost
+	    || (new_cost == old_cost && !prop.likely_profitable_p ()))
 	  {
 	    if (dump_file)
 	      fprintf (dump_file, "change not profitable"
diff --git a/gcc/testsuite/gcc.target/i386/pr111267.c b/gcc/testsuite/gcc.target/i386/pr111267.c
new file mode 100644
index 00000000000..e3d549d8cba
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr111267.c
@@ -0,0 +1,14 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+struct S { float a, b, c, d; };
+
+int
+bar (struct S x, struct S y)
+{
+  return x.b <= y.d && x.c >= y.a;
+}
+
+/* { dg-final { scan-assembler-not "movq" } } */
+/* { dg-final { scan-assembler-not "xchgq" } } */
+/* { dg-final { scan-assembler-not "shrq" } } */
+/* { dg-final { scan-assembler-not "movd" } } */