diff mbox

[ARM] Use UBFX for some and-immediate operations

Message ID 501AA9D8.3080506@arm.com
State New
Headers show

Commit Message

Richard Earnshaw Aug. 2, 2012, 4:24 p.m. UTC 
This patch adds the ability to spot and exploit the UBFX instruction
(bit-field extract) to perform AND operations with an immediate which is
2^N - 1.  On some benchmarks this can lead to a significant win over the
existing behaviour on Cortex-A15.

Tested on arm-eabi with some additional testing on arm-linux-gnueabi
platforms.

R.

2012-08-02  Richard Earnshaw  <rearnsha@arm.com>

	* arm.c (arm_gen_constant): Use UBFX for some AND operations when
	available.

Comments

Richard Henderson Aug. 3, 2012, 6:10 p.m. UTC | #1 
On 2012-08-02 09:24, Richard Earnshaw wrote:
> +	    /* Extz only supports SImode, but we can coerce the operands
> +	       into that mode.  */
> +	    emit_constant_insn (cond,
> +				gen_extzv_t2 (gen_lowpart (mode, target),
> +					      gen_lowpart (mode, source),
> +					      GEN_INT (i), const0_rtx));

Didn't you mean gen_lowpart (SImode, ...) ?


r~
diff mbox

Patch

--- arm.c	(revision 190101)
+++ arm.c	(local)
@@ -2982,6 +2982,31 @@  arm_gen_constant (enum rtx_code code, en
       return 1;
     }
 
+  /* On targets with UXTH/UBFX, we can deal with AND (2^N)-1 in a single
+     insn.  */
+  if (code == AND && (i = exact_log2 (remainder + 1)) > 0
+      && (arm_arch_thumb2 || (i == 16 && arm_arch6 && mode == SImode)))
+    {
+      if (generate)
+	{
+	  if (mode == SImode && i == 16)
+	    /* Use UXTH in preference to UBFX, since on Thumb2 it's a 
+	       smaller insn.  */
+	    emit_constant_insn (cond,
+				gen_zero_extendhisi2
+				(target, gen_lowpart (HImode, source)));
+	  else
+	    /* Extz only supports SImode, but we can coerce the operands
+	       into that mode.  */
+	    emit_constant_insn (cond,
+				gen_extzv_t2 (gen_lowpart (mode, target),
+					      gen_lowpart (mode, source),
+					      GEN_INT (i), const0_rtx));
+	}
+
+      return 1;
+    }
+
   /* Calculate a few attributes that may be useful for specific
      optimizations.  */
   /* Count number of leading zeros.  */