[AArch64] Optimise aarch64_add_offset for SVE VL constants

aarch64_add_offset contains code to decompose all SVE VL-based constants
into native operations.  The worst-case fallback is to load the number
of SVE elements into a register and use a general multiplication.
This patch improves that fallback by reusing expand_mult if
can_create_pseudo_p, rather than emitting a MULT pattern directly.

In order to increase the chances of being able to use a simple
add-and-shift, the patch also tries to compute VG * the lowest set
bit of the multiplier, rather than always using CNTD as the basis
for the multiplication path.

This is tested by the ACLE patches but is really an independent
improvement.

2019-08-15  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
	* config/aarch64/aarch64.c (aarch64_add_offset): In the fallback
	multiplication case, try to compute VG * (lowest set bit) directly
	rather than always basing the multiplication on VG.  Use
	expand_mult for the multiplication if we can.

gcc/testsuite/
	* gcc.target/aarch64/sve/loop_add_4.c: Expect 10 INCWs and
	INCDs rather than 8.

From-SVN: r274519

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index c3aa113..5fbf319 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,5 +1,12 @@
 2019-08-15  Richard Sandiford  <richard.sandiford@arm.com>
 
+       * config/aarch64/aarch64.c (aarch64_add_offset): In the fallback
+       multiplication case, try to compute VG * (lowest set bit) directly
+       rather than always basing the multiplication on VG.  Use
+       expand_mult for the multiplication if we can.
+
+2019-08-15  Richard Sandiford  <richard.sandiford@arm.com>
+
        * config/aarch64/aarch64-protos.h
        (aarch64_sve_scalar_inc_dec_immediate_p): Declare.
        (aarch64_sve_inc_dec_immediate_p): Rename to...

diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index b8c947e..3b0dcea 100644 (file)
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -73,6 +73,7 @@
 #include "selftest-rtl.h"
 #include "rtx-vector-builder.h"
 #include "intl.h"
+#include "expmed.h"
 
 /* This file should be included last.  */
 #include "target-def.h"
@@ -3465,20 +3466,36 @@ aarch64_add_offset (scalar_int_mode mode, rtx dest, rtx src,
        }
       else
        {
-         /* Use CNTD, then multiply it by FACTOR.  */
-         val = gen_int_mode (poly_int64 (2, 2), mode);
+         /* Base the factor on LOW_BIT if we can calculate LOW_BIT
+            directly, since that should increase the chances of being
+            able to use a shift and add sequence.  If LOW_BIT itself
+            is out of range, just use CNTD.  */
+         if (low_bit <= 16 * 8)
+           factor /= low_bit;
+         else
+           low_bit = 1;
+
+         val = gen_int_mode (poly_int64 (low_bit * 2, low_bit * 2), mode);
          val = aarch64_force_temporary (mode, temp1, val);
 
-         /* Go back to using a negative multiplication factor if we have
-            no register from which to subtract.  */
-         if (code == MINUS && src == const0_rtx)
+         if (can_create_pseudo_p ())
            {
-             factor = -factor;
-             code = PLUS;
+             rtx coeff1 = gen_int_mode (factor, mode);
+             val = expand_mult (mode, val, coeff1, NULL_RTX, false, true);
+           }
+         else
+           {
+             /* Go back to using a negative multiplication factor if we have
+                no register from which to subtract.  */
+             if (code == MINUS && src == const0_rtx)
+               {
+                 factor = -factor;
+                 code = PLUS;
+               }
+             rtx coeff1 = gen_int_mode (factor, mode);
+             coeff1 = aarch64_force_temporary (mode, temp2, coeff1);
+             val = gen_rtx_MULT (mode, val, coeff1);
            }
-         rtx coeff1 = gen_int_mode (factor, mode);
-         coeff1 = aarch64_force_temporary (mode, temp2, coeff1);
-         val = gen_rtx_MULT (mode, val, coeff1);
        }
 
       if (shift > 0)

diff --git a/gcc/testsuite/ChangeLog b/gcc/testsuite/ChangeLog
index 9f667f2..0008ff9 100644 (file)
--- a/gcc/testsuite/ChangeLog
+++ b/gcc/testsuite/ChangeLog
@@ -1,5 +1,10 @@
 2019-08-15  Richard Sandiford  <richard.sandiford@arm.com>
 
+       * gcc.target/aarch64/sve/loop_add_4.c: Expect 10 INCWs and
+       INCDs rather than 8.
+
+2019-08-15  Richard Sandiford  <richard.sandiford@arm.com>
+
        * gcc.target/aarch64/sve/revb_1.c: Restrict to little-endian targets.
        Avoid including stdint.h.
        * gcc.target/aarch64/sve/revh_1.c: Likewise.

diff --git a/gcc/testsuite/gcc.target/aarch64/sve/loop_add_4.c b/gcc/testsuite/gcc.target/aarch64/sve/loop_add_4.c
index 7f02497..9ead9c2 100644 (file)
--- a/gcc/testsuite/gcc.target/aarch64/sve/loop_add_4.c
+++ b/gcc/testsuite/gcc.target/aarch64/sve/loop_add_4.c
@@ -68,7 +68,8 @@ TEST_ALL (LOOP)
 /* { dg-final { scan-assembler-times {\tindex\tz[0-9]+\.s, w[0-9]+, w[0-9]+\n} 3 } } */
 /* { dg-final { scan-assembler-times {\tld1w\tz[0-9]+\.s, p[0-7]+/z, \[x[0-9]+, x[0-9]+, lsl 2\]} 8 } } */
 /* { dg-final { scan-assembler-times {\tst1w\tz[0-9]+\.s, p[0-7]+, \[x[0-9]+, x[0-9]+, lsl 2\]} 8 } } */
-/* { dg-final { scan-assembler-times {\tincw\tx[0-9]+\n} 8 } } */
+/* 2 for the calculations of -17 and 17.  */
+/* { dg-final { scan-assembler-times {\tincw\tx[0-9]+\n} 10 } } */
 
 /* { dg-final { scan-assembler-times {\tdecw\tz[0-9]+\.s, all, mul #16\n} 1 } } */
 /* { dg-final { scan-assembler-times {\tdecw\tz[0-9]+\.s, all, mul #15\n} 1 } } */
@@ -85,7 +86,8 @@ TEST_ALL (LOOP)
 /* { dg-final { scan-assembler-times {\tindex\tz[0-9]+\.d, x[0-9]+, x[0-9]+\n} 3 } } */
 /* { dg-final { scan-assembler-times {\tld1d\tz[0-9]+\.d, p[0-7]+/z, \[x[0-9]+, x[0-9]+, lsl 3\]} 8 } } */
 /* { dg-final { scan-assembler-times {\tst1d\tz[0-9]+\.d, p[0-7]+, \[x[0-9]+, x[0-9]+, lsl 3\]} 8 } } */
-/* { dg-final { scan-assembler-times {\tincd\tx[0-9]+\n} 8 } } */
+/* 2 for the calculations of -17 and 17.  */
+/* { dg-final { scan-assembler-times {\tincd\tx[0-9]+\n} 10 } } */
 
 /* { dg-final { scan-assembler-times {\tdecd\tz[0-9]+\.d, all, mul #16\n} 1 } } */
 /* { dg-final { scan-assembler-times {\tdecd\tz[0-9]+\.d, all, mul #15\n} 1 } } */