[PATCH] rs6000: Fix for PR63187

[PATCH] rs6000: Fix for PR63187

[Segher Boessenkool]

I made a thinko in r214080.  That patch makes the "andi." pattern not match
for insns that can match "rlwinm", and also makes "rlwinm." not match insns
that can match "andi." (and similar for "andis." and "rldicl" etc.)  This
ensures we always choose the cheapest instruction.

However, the splitter for "andi." does not take this into account, and
generates a clobber always, leading to unrecognisable instructions.  This
patch fixes it.

Okay if it survives bootstrap+regcheck?


Segher


2014-09-05  Segher Boessenkool  <segher@kernel.crashing.org>

	PR target/63187
	* config/rs6000/rs6000.md (*and<mode>3_imm_dot, *and<mode>3_imm_dot2):
	Do not allow any_mask_operand for operands[2].
	(*and<mode>3_imm_mask_dot, *and<mode>3_imm_mask_dot2): New.


---
 gcc/config/rs6000/rs6000.md | 56 +++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 54 insertions(+), 2 deletions(-)

diff --git a/gcc/config/rs6000/rs6000.md b/gcc/config/rs6000/rs6000.md
index 8f3549e..2df8e41 100644
--- a/gcc/config/rs6000/rs6000.md
+++ b/gcc/config/rs6000/rs6000.md
@@ -2736,7 +2736,8 @@ (define_insn_and_split "*and<mode>3_imm_dot"
    (clobber (match_scratch:GPR 0 "=r,r"))
    (clobber (match_scratch:CC 4 "=X,x"))]
   "(<MODE>mode == Pmode || UINTVAL (operands[2]) <= 0x7fffffff)
-   && rs6000_gen_cell_microcode"
+   && rs6000_gen_cell_microcode
+   && !any_mask_operand (operands[2], <MODE>mode)"
   "@
    andi%e2. %0,%1,%u2
    #"
@@ -2763,7 +2764,8 @@ (define_insn_and_split "*and<mode>3_imm_dot2"
 		 (match_dup 2)))
    (clobber (match_scratch:CC 4 "=X,x"))]
   "(<MODE>mode == Pmode || UINTVAL (operands[2]) <= 0x7fffffff)
-   && rs6000_gen_cell_microcode"
+   && rs6000_gen_cell_microcode
+   && !any_mask_operand (operands[2], <MODE>mode)"
   "@
    andi%e2. %0,%1,%u2
    #"
@@ -2780,6 +2782,56 @@ (define_insn_and_split "*and<mode>3_imm_dot2"
    (set_attr "dot" "yes")
    (set_attr "length" "4,8")])
 
+(define_insn_and_split "*and<mode>3_imm_mask_dot"
+  [(set (match_operand:CC 3 "cc_reg_operand" "=x,??y")
+	(compare:CC (and:GPR (match_operand:GPR 1 "gpc_reg_operand" "%r,r")
+			     (match_operand:GPR 2 "logical_const_operand" "n,n"))
+		    (const_int 0)))
+   (clobber (match_scratch:GPR 0 "=r,r"))]
+  "(<MODE>mode == Pmode || UINTVAL (operands[2]) <= 0x7fffffff)
+   && rs6000_gen_cell_microcode
+   && any_mask_operand (operands[2], <MODE>mode)"
+  "@
+   andi%e2. %0,%1,%u2
+   #"
+  "&& reload_completed && cc_reg_not_cr0_operand (operands[3], CCmode)"
+  [(set (match_dup 0)
+	(and:GPR (match_dup 1)
+		 (match_dup 2)))
+   (set (match_dup 3)
+	(compare:CC (match_dup 0)
+		    (const_int 0)))]
+  ""
+  [(set_attr "type" "logical")
+   (set_attr "dot" "yes")
+   (set_attr "length" "4,8")])
+
+(define_insn_and_split "*and<mode>3_imm_mask_dot2"
+  [(set (match_operand:CC 3 "cc_reg_operand" "=x,??y")
+	(compare:CC (and:GPR (match_operand:GPR 1 "gpc_reg_operand" "%r,r")
+			     (match_operand:GPR 2 "logical_const_operand" "n,n"))
+		    (const_int 0)))
+   (set (match_operand:GPR 0 "gpc_reg_operand" "=r,r")
+	(and:GPR (match_dup 1)
+		 (match_dup 2)))]
+  "(<MODE>mode == Pmode || UINTVAL (operands[2]) <= 0x7fffffff)
+   && rs6000_gen_cell_microcode
+   && any_mask_operand (operands[2], <MODE>mode)"
+  "@
+   andi%e2. %0,%1,%u2
+   #"
+  "&& reload_completed && cc_reg_not_cr0_operand (operands[3], CCmode)"
+  [(set (match_dup 0)
+	(and:GPR (match_dup 1)
+		 (match_dup 2)))
+   (set (match_dup 3)
+	(compare:CC (match_dup 0)
+		    (const_int 0)))]
+  ""
+  [(set_attr "type" "logical")
+   (set_attr "dot" "yes")
+   (set_attr "length" "4,8")])
+
 
 (define_insn "*and<mode>3_mask"
   [(set (match_operand:GPR 0 "gpc_reg_operand" "=r,r")
-- 
1.8.1.4

Re: [PATCH] rs6000: Fix for PR63187

[David Edelsohn]

On Fri, Sep 5, 2014 at 1:48 PM, Segher Boessenkool
<segher@kernel.crashing.org> wrote:
> I made a thinko in r214080.  That patch makes the "andi." pattern not match
> for insns that can match "rlwinm", and also makes "rlwinm." not match insns
> that can match "andi." (and similar for "andis." and "rldicl" etc.)  This
> ensures we always choose the cheapest instruction.
>
> However, the splitter for "andi." does not take this into account, and
> generates a clobber always, leading to unrecognisable instructions.  This
> patch fixes it.
>
> Okay if it survives bootstrap+regcheck?
>
>
> Segher
>
>
> 2014-09-05  Segher Boessenkool  <segher@kernel.crashing.org>
>
>         PR target/63187
>         * config/rs6000/rs6000.md (*and<mode>3_imm_dot, *and<mode>3_imm_dot2):
>         Do not allow any_mask_operand for operands[2].
>         (*and<mode>3_imm_mask_dot, *and<mode>3_imm_mask_dot2): New.

Okay.

Thanks, David