diff mbox

rs6000: Make rs6000_split_logical handle inverted 2nd operand (PR64358)

Message ID 026b6a6be3c17513aedc8272f4fd6942e6db0b92.1419087240.git.segher@kernel.crashing.org
State New
Headers show

Commit Message

Segher Boessenkool Dec. 20, 2014, 3:18 p.m. UTC 
If rs6000_split_logical is asked to invert the second operand (but not
the first), it emits RTL that is just that; but canonical RTL has the
first arm inverted if only one, not the second.  This patch fixes that.

With that in place, the bug in PR target/64358 is easily fixed.  Do
that.  Also change the other 128-bit boolc splitter to have the same
operand order as everything else.  And use BOOL_REGS_OP1 and _OP2 for
operands 1 resp. 2 (not the other way around); this doesn't actually
matter, but it was a bit confusing.

This fixes builtin-arith-overflow-{10,11}.c as well.

Bootstrapped (no fortran this time, it doesn't build right now) and
tested on powerpc64-linux, -m32,-m32/-mpowerpc64,-m64,-m64/-mlra; no
regressions (and the mentioned cases fixed).

Okay for mainline?  Does this need backporting as well?


Segher


2014-12-20  Segher Boessenkool  <segher@kernel.crashing.org>

gcc/
	PR target/64358
	* config/rs6000/rs6000.c (rs6000_split_logical_inner): Swap the
	input operands if only the second is inverted.
	* config/rs6000/rs6000.md (*boolc<mode>3_internal1 for BOOL_128):
	Swap BOOL_REGS_OP1 and BOOL_REGS_OP2.  Correct arguments to
	rs6000_split_logical.
	(*boolc<mode>3_internal2 for TI2): Swap operands[1] and operands[2].

---
 gcc/config/rs6000/rs6000.c  |  4 ++++
 gcc/config/rs6000/rs6000.md | 12 ++++++------
 2 files changed, 10 insertions(+), 6 deletions(-)

Comments

David Edelsohn Dec. 20, 2014, 3:29 p.m. UTC | #1 
On Sat, Dec 20, 2014 at 10:18 AM, Segher Boessenkool
<segher@kernel.crashing.org> wrote:
> If rs6000_split_logical is asked to invert the second operand (but not
> the first), it emits RTL that is just that; but canonical RTL has the
> first arm inverted if only one, not the second.  This patch fixes that.
>
> With that in place, the bug in PR target/64358 is easily fixed.  Do
> that.  Also change the other 128-bit boolc splitter to have the same
> operand order as everything else.  And use BOOL_REGS_OP1 and _OP2 for
> operands 1 resp. 2 (not the other way around); this doesn't actually
> matter, but it was a bit confusing.
>
> This fixes builtin-arith-overflow-{10,11}.c as well.
>
> Bootstrapped (no fortran this time, it doesn't build right now) and
> tested on powerpc64-linux, -m32,-m32/-mpowerpc64,-m64,-m64/-mlra; no
> regressions (and the mentioned cases fixed).
>
> Okay for mainline?  Does this need backporting as well?
>
>
> Segher
>
>
> 2014-12-20  Segher Boessenkool  <segher@kernel.crashing.org>
>
> gcc/
>         PR target/64358
>         * config/rs6000/rs6000.c (rs6000_split_logical_inner): Swap the
>         input operands if only the second is inverted.
>         * config/rs6000/rs6000.md (*boolc<mode>3_internal1 for BOOL_128):
>         Swap BOOL_REGS_OP1 and BOOL_REGS_OP2.  Correct arguments to
>         rs6000_split_logical.
>         (*boolc<mode>3_internal2 for TI2): Swap operands[1] and operands[2].

Okay.

Thanks, David
diff mbox

Patch

diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index 73152ce..bbed888 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -33167,6 +33167,10 @@  rs6000_split_logical_inner (rtx dest,
   if (complement_op2_p)
     op2 = gen_rtx_NOT (mode, op2);
 
+  /* For canonical RTL, if only one arm is inverted it is the first.  */
+  if (!complement_op1_p && complement_op2_p)
+    std::swap (op1, op2);
+
   bool_rtx = ((code == NOT)
 	      ? gen_rtx_NOT (mode, op1)
 	      : gen_rtx_fmt_ee (code, mode, op1, op2));
diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index 36e6182..2d3a5d3 100644
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -7570,8 +7570,8 @@  (define_insn_and_split "*boolc<mode>3_internal1"
   [(set (match_operand:BOOL_128 0 "vlogical_operand" "=<BOOL_REGS_OUTPUT>")
 	(match_operator:BOOL_128 3 "boolean_operator"
 	 [(not:BOOL_128
-	   (match_operand:BOOL_128 2 "vlogical_operand" "<BOOL_REGS_OP1>"))
-	  (match_operand:BOOL_128 1 "vlogical_operand" "<BOOL_REGS_OP2>")]))]
+	   (match_operand:BOOL_128 2 "vlogical_operand" "<BOOL_REGS_OP2>"))
+	  (match_operand:BOOL_128 1 "vlogical_operand" "<BOOL_REGS_OP1>")]))]
   "TARGET_P8_VECTOR || (GET_CODE (operands[3]) == AND)"
 {
   if (TARGET_VSX && vsx_register_operand (operands[0], <MODE>mode))
@@ -7586,7 +7586,7 @@  (define_insn_and_split "*boolc<mode>3_internal1"
    && reload_completed && int_reg_operand (operands[0], <MODE>mode)"
   [(const_int 0)]
 {
-  rs6000_split_logical (operands, GET_CODE (operands[3]), false, true, false);
+  rs6000_split_logical (operands, GET_CODE (operands[3]), false, false, true);
   DONE;
 }
   [(set (attr "type")
@@ -7607,14 +7607,14 @@  (define_insn_and_split "*boolc<mode>3_internal2"
   [(set (match_operand:TI2 0 "int_reg_operand" "=&r,r,r")
 	(match_operator:TI2 3 "boolean_operator"
 	 [(not:TI2
-	   (match_operand:TI2 1 "int_reg_operand" "r,0,r"))
-	  (match_operand:TI2 2 "int_reg_operand" "r,r,0")]))]
+	   (match_operand:TI2 2 "int_reg_operand" "r,0,r"))
+	  (match_operand:TI2 1 "int_reg_operand" "r,r,0")]))]
   "!TARGET_P8_VECTOR && (GET_CODE (operands[3]) != AND)"
   "#"
   "reload_completed && !TARGET_P8_VECTOR && (GET_CODE (operands[3]) != AND)"
   [(const_int 0)]
 {
-  rs6000_split_logical (operands, GET_CODE (operands[3]), false, true, false);
+  rs6000_split_logical (operands, GET_CODE (operands[3]), false, false, true);
   DONE;
 }
   [(set_attr "type" "integer")