Using ranger from cfgexpand

Using ranger from cfgexpand

[Andrew Pinski]

[-- Attachment #1: Type: text/plain, Size: 2306 bytes --]

Hi,
  I was trying to use the ranger from inside of cfgexpand but since at
this point we have a mix of RTL and gimple basic blocks, things fall
over very fast.

First does it make sense to use the ranger from expand or should we
change the gimple IR in isel instead if we have a better way of doing
an expansion.

In this case we have MIN_EXPR<a, 1> where a has a range of [0,INF] and
I want to expand it as (a != 0) to get better code.

Attached is the current patch which I did and a testcase which ICEs at -O2.

The backtrace of the ICE is:
0x88c885 gsi_last_bb(basic_block_def*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-iterator.h:165
0x88c885 gsi_last_nondebug_bb(basic_block_def*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-iterator.h:325
0x88c885 gimple_outgoing_range_stmt_p(basic_block_def*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-edge.cc:42
0xe05739 gimple_outgoing_range::edge_range_p(irange&, edge_def*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-edge.cc:194
0xe12afa gori_compute::outgoing_edge_range_p(vrange&, edge_def*,
tree_node*, range_query&)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-gori.cc:1404
0xe00dde ranger_cache::edge_range(vrange&, edge_def*, tree_node*,
ranger_cache::rfd_mode)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-cache.cc:1185
0xe01e48 ranger_cache::propagate_cache(tree_node*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-cache.cc:1305
0xe0306f ranger_cache::fill_block_cache(tree_node*, basic_block_def*,
basic_block_def*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-cache.cc:1592
0xe039b1 ranger_cache::block_range(vrange&, basic_block_def*, tree_node*, bool)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range-cache.cc:1260
0xdf9776 gimple_ranger::range_on_entry(vrange&, basic_block_def*, tree_node*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range.cc:164
0xdfcc29 gimple_ranger::range_of_expr(vrange&, tree_node*, gimple*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/gimple-range.cc:138
0xd1a334 expand_expr_real_2(separate_ops*, rtx_def*, machine_mode,
expand_modifier, gimple*)
        /home/apinski/src/upstream-gcc-git/gcc/gcc/expr.cc:9969

Thanks,
Andrew Pinski

[-- Attachment #2: expand1.diff.txt --]
[-- Type: text/plain, Size: 3315 bytes --]

diff --git a/gcc/cfgexpand.cc b/gcc/cfgexpand.cc
index edf292cfbe9..b908d329f1e 100644
--- a/gcc/cfgexpand.cc
+++ b/gcc/cfgexpand.cc
@@ -74,6 +74,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "output.h"
 #include "builtins.h"
 #include "opts.h"
+#include "gimple-range.h"
 
 /* Some systems use __main in a way incompatible with its use in gcc, in these
    cases use the macros NAME__MAIN to give a quoted symbol and SYMBOL__MAIN to
@@ -3981,7 +3982,7 @@ expand_gimple_stmt_1 (gimple *stmt)
 	       don't directly expand to target.  */
 	    temp = nontemporal || promoted ? NULL_RTX : target;
 	    temp = expand_expr_real_2 (&ops, temp, GET_MODE (target),
-				       EXPAND_NORMAL);
+				       EXPAND_NORMAL, stmt);
 
 	    if (temp == target)
 	      ;
@@ -6606,6 +6607,9 @@ pass_expand::execute (function *fun)
   rtx_insn *var_seq, *var_ret_seq;
   unsigned i;
 
+  if (optimize >= 2)
+    enable_ranger (cfun, false);
+
   timevar_push (TV_OUT_OF_SSA);
   rewrite_out_of_ssa (&SA);
   timevar_pop (TV_OUT_OF_SSA);
@@ -7054,6 +7058,9 @@ pass_expand::execute (function *fun)
 
   timevar_pop (TV_POST_EXPAND);
 
+  if (optimize >= 2)
+    disable_ranger (cfun);
+
   return 0;
 }
 
diff --git a/gcc/expr.cc b/gcc/expr.cc
index 9a37bff1fdd..934fcbb9f63 100644
--- a/gcc/expr.cc
+++ b/gcc/expr.cc
@@ -65,6 +65,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "rtx-vector-builder.h"
 #include "tree-pretty-print.h"
 #include "flags.h"
+#include "value-query.h"
 
 
 /* If this is nonzero, we do not bother generating VOLATILE
@@ -9207,7 +9208,7 @@ expand_expr_divmod (tree_code code, machine_mode mode, tree treeop0,
 
 rtx
 expand_expr_real_2 (sepops ops, rtx target, machine_mode tmode,
-		    enum expand_modifier modifier)
+		    enum expand_modifier modifier, gimple *stmt)
 {
   rtx op0, op1, op2, temp;
   rtx_code_label *lab;
@@ -9958,6 +9959,27 @@ expand_expr_real_2 (sepops ops, rtx target, machine_mode tmode,
       expand_operands (treeop0, treeop1,
 		       target, &op0, &op1, EXPAND_NORMAL);
 
+      /* For MIN<a, 1> if a has a range with a lower bound of 0,
+         just expand as a != 0 if we can. */
+      if (INTEGRAL_TYPE_P (type)
+	  && op1 == const1_rtx
+	  && code == MIN_EXPR)
+	{
+	  value_range vr;
+	  get_range_query (cfun)->range_of_expr (vr, treeop0, stmt);
+	  if (vr.undefined_p ())
+	    vr.set_varying (type);
+	  if (vr.lower_bound() == 0)
+	    {
+	      rtx tmp = emit_store_flag_force (target, NE,
+					       op0, const0_rtx,
+					       mode, unsignedp,
+					       1);
+	      if (tmp)
+		return tmp;
+	    }
+	}
+
       /* First try to do it with a special MIN or MAX instruction.
 	 If that does not win, use a conditional jump to select the proper
 	 value.  */
diff --git a/gcc/expr.h b/gcc/expr.h
index 11bff531862..8aec76e4f31 100644
--- a/gcc/expr.h
+++ b/gcc/expr.h
@@ -298,7 +298,7 @@ extern rtx expand_expr_real (tree, rtx, machine_mode,
 extern rtx expand_expr_real_1 (tree, rtx, machine_mode,
 			       enum expand_modifier, rtx *, bool);
 extern rtx expand_expr_real_2 (sepops, rtx, machine_mode,
-			       enum expand_modifier);
+			       enum expand_modifier, gimple * = nullptr);
 
 /* Generate code for computing expression EXP.
    An rtx for the computed value is returned.  The value is never null.

[-- Attachment #3: t3.c --]
[-- Type: application/octet-stream, Size: 116 bytes --]

int g(int min_need_stall)
{
  return  min_need_stall < 0 ? 1 : ((min_need_stall) < (1) ? (min_need_stall) : (1));
}

Re: Using ranger from cfgexpand

[Jeff Law]

On 9/2/23 18:40, Andrew Pinski via Gcc wrote:
> Hi,
>    I was trying to use the ranger from inside of cfgexpand but since at
> this point we have a mix of RTL and gimple basic blocks, things fall
> over very fast.
> 
> First does it make sense to use the ranger from expand or should we
> change the gimple IR in isel instead if we have a better way of doing
> an expansion.
> 
> In this case we have MIN_EXPR<a, 1> where a has a range of [0,INF] and
> I want to expand it as (a != 0) to get better code.
I very much want to get to a point where we can use ranges during 
expansion, but we're probably going to be limited to global ranges since 
as you note, we're in a mixed IL state and ranger is probably going to 
be confused as hell.


Jeff

Re: Using ranger from cfgexpand

[Richard Biener]

> Am 03.09.2023 um 03:23 schrieb Jeff Law via Gcc <gcc@gcc.gnu.org>:
> 
> 
> 
>> On 9/2/23 18:40, Andrew Pinski via Gcc wrote:
>> Hi,
>>   I was trying to use the ranger from inside of cfgexpand but since at
>> this point we have a mix of RTL and gimple basic blocks, things fall
>> over very fast.
>> First does it make sense to use the ranger from expand or should we
>> change the gimple IR in isel instead if we have a better way of doing
>> an expansion.
>> In this case we have MIN_EXPR<a, 1> where a has a range of [0,INF] and
>> I want to expand it as (a != 0) to get better code.
> I very much want to get to a point where we can use ranges during expansion, but we're probably going to be limited to global ranges since as you note, we're in a mixed IL state and ranger is probably going to be confused as hell.

I think we want to move towards doing all the clever bits of RTL expansion as a preparation on GIMPLE which is what we have ISEL for.

Richard 

> 
> Jeff