Merge lp:~michaelh1/gcc-linaro/dont-peel-neon into lp:gcc-linaro/4.6

[Michael Hope]

Disable peeling for NEON as NEON can do unaligned loads and the setup cost is too high for small loops.

This request is for testing and benchmarking. If the benchmarks are good then it can be merged.

[Linaro Toolchain Builder (cbuild)]

cbuild has taken a snapshot of this branch at r106769 and queued it for build.

The snapshot is available at:
 http://ex.seabright.co.nz/snapshots/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon.tar.xdelta3.xz

and will be built on the following builders:
 a9-builder armv5-builder i686 x86_64

You can track the build queue at:
 http://ex.seabright.co.nz/helpers/scheduler

cbuild-snapshot: gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon
cbuild-ancestor: lp:gcc-linaro/4.6+bzr106768
cbuild-state: check

[Linaro Toolchain Builder (cbuild)]

cbuild successfully built this on x86_64-natty-cbuild152-oort3-x86_64r1.

The build results are available at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/x86_64-natty-cbuild152-oort3-x86_64r1

The test suite results were unchanged compared to the branch point lp:gcc-linaro/4.6+bzr106768.

The full testsuite results are at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/x86_64-natty-cbuild152-oort3-x86_64r1/gcc-testsuite.txt

cbuild-checked: x86_64-natty-cbuild152-oort3-x86_64r1

[Linaro Toolchain Builder (cbuild)]

cbuild successfully built this on armv7l-natty-cbuild152-ursa4-cortexa9r1.

The build results are available at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/armv7l-natty-cbuild152-ursa4-cortexa9r1

The test suite results changed compared to the branch point lp:gcc-linaro/4.6+bzr106768:
 -PASS: gcc.dg/vect/no-section-anchors-vect-69.c scan-tree-dump-times vect "Alignment of access forced using peeling" 2
 +PASS: gcc.dg/vect/no-section-anchors-vect-69.c scan-tree-dump-times vect "Alignment of access forced using peeling" 1
 +XFAIL: gcc.dg/vect/no-section-anchors-vect-69.c scan-tree-dump-times vect "Alignment of access forced using peeling" 2
 +FAIL: gcc.dg/vect/no-section-anchors-vect-69.c scan-tree-dump-times vect "Alignment of access forced using versioning" 1
 -PASS: gcc.dg/vect/no-section-anchors-vect-69.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 +FAIL: gcc.dg/vect/no-section-anchors-vect-69.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 -PASS: gcc.dg/vect/section-anchors-vect-69.c scan-tree-dump-times vect "Alignment of access forced using peeling" 4
 +FAIL: gcc.dg/vect/section-anchors-vect-69.c scan-tree-dump-times vect "Alignment of access forced using versioning" 4
 -PASS: gcc.dg/vect/section-anchors-vect-69.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 +FAIL: gcc.dg/vect/section-anchors-vect-69.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 -PASS: gcc.dg/vect/vect-28.c -flto scan-tree-dump-times vect "Alignment of access forced using peeling" 1
 +FAIL: gcc.dg/vect/vect-28.c -flto scan-tree-dump-times vect "Alignment of access forced using versioning" 1
 -PASS: gcc.dg/vect/vect-28.c -flto scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 +FAIL: gcc.dg/vect/vect-28.c -flto scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 -PASS: gcc.dg/vect/vect-28.c scan-tree-dump-times vect "Alignment of access forced using peeling" 1
 +FAIL: gcc.dg/vect/vect-28.c scan-tree-dump-times vect "Alignment of access forced using versioning" 1
 -PASS: gcc.dg/vect/vect-28.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 +FAIL: gcc.dg/vect/vect-28.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 -PASS: gcc.dg/vect/vect-33.c -flto scan-tree-dump-times vect "Alignment of access forced using peeling" 1
 +FAIL: gcc.dg/vect/vect-33.c -flto scan-tree-dump-times vect "Alignment of access forced using versioning" 1
 -PASS: gcc.dg/vect/vect-33.c -flto scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 +FAIL: gcc.dg/vect/vect-33.c -flto scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 -PASS: gcc.dg/vect/vect-33.c scan-tree-dump-times vect "Alignment of access forced using peeling" 1
 +FAIL: gcc.dg/vect/vect-33.c scan-tree-dump-times vect "Alignment of access forced using versioning" 1
 -PASS: gcc.dg/vect/vect-33.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 +FAIL: gcc.dg/vect/vect-33.c scan-tree-dump-times vect "Vectorizing an unaligned access" 0
 -PASS: gcc.dg/vect/vect-44.c -flto scan-tree-dump-times vect "Alignment of access forced using peeling" 1...

[Linaro Toolchain Builder (cbuild)]

cbuild successfully built this on i686-natty-cbuild152-oort4-i686r1.

The build results are available at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/i686-natty-cbuild152-oort4-i686r1

The test suite results were unchanged compared to the branch point lp:gcc-linaro/4.6+bzr106768.

The full testsuite results are at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/i686-natty-cbuild152-oort4-i686r1/gcc-testsuite.txt

cbuild-checked: i686-natty-cbuild152-oort4-i686r1

[Michael Hope (michaelh1)]

Self-rejecting. I have a build that we can benchmark with now.

review: Disapprove

[Michael Hope (michaelh1)]

cbuild successfully built this on i686-natty-cbuild152-oort4-i686r1.

The build results are available at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/i686-natty-cbuild152-oort4-i686r1

The testsuite results are the same as the branch point lp:gcc-linaro/4.6+bzr106768

The full testsuite results are at:
 http://ex.seabright.co.nz/build/gcc-linaro-4.6+bzr106769~michaelh1~dont-peel-neon/logs/i686-natty-cbuild152-oort4-i686r1/gcc-testsuite.txt

[Loïc Minier (lool)]

Rejecting per your earlier self-reject

Unmerged revisions

106769. By Michael Hope on 2011-07-11

Disable peeling for NEON as NEON can do unaligned loads and the
setup cost is too high for small loops.

This request is for testing and benchmarking. If the benchmarks are
good then it can be merged.

=== modified file 'ChangeLog.linaro'
--- ChangeLog.linaro	2011-07-07 13:15:23 +0000
+++ ChangeLog.linaro	2011-07-11 05:35:32 +0000
@@ -1,3 +1,20 @@
[phone]  Michael Hope  <[e-mail]>
+
+	Merge from Sourcery G++ 4.5:
+
+	2011-04-01  Julian Brown  <[e-mail]>
+
+	Issue #6223
+
+	gcc/
+	* config/arm/arm.c (arm_vector_alignment_reachable): Don't use
+	peeling for alignment for NEON.
+
+	gcc/testsuite/
+	* lib/target-supports.exp
+	(check_effective_target_vector_alignment_reachable): Not true for
+	ARM NEON.
+
 2011-07-07  Richard Sandiford  <[e-mail]>
 
 	gcc/

=== modified file 'gcc/config/arm/arm.c'
--- gcc/config/arm/arm.c	2011-06-29 09:13:17 +0000
+++ gcc/config/arm/arm.c	2011-07-11 05:35:32 +0000
@@ -23874,10 +23874,13 @@
 static bool
 arm_vector_alignment_reachable (const_tree type, bool is_packed)
 {
-  /* Vectors which aren't in packed structures will not be less aligned than
-     the natural alignment of their element type, so this is safe.  */
+  /* NOTE: returning true here will unconditionally peel loop iterations so
+     that aligned accesses can be used.  This is undesirable when misaligned
+     accesses are available, particularly for small loop iteration counts,
+     since the overhead for dispatching to multiple versions of the loop is
+     quite high.  */
   if (TARGET_NEON && !BYTES_BIG_ENDIAN)
-    return !is_packed;
+    return false;
 
   return default_builtin_vector_alignment_reachable (type, is_packed);
 }

=== modified file 'gcc/testsuite/lib/target-supports.exp'
--- gcc/testsuite/lib/target-supports.exp	2011-07-04 11:13:51 +0000
+++ gcc/testsuite/lib/target-supports.exp	2011-07-11 05:35:32 +0000
@@ -3027,8 +3027,9 @@
     if [info exists et_vector_alignment_reachable_saved] {
         verbose "check_effective_target_vector_alignment_reachable: using cached result" 2
     } else {
-        if { [check_effective_target_vect_aligned_arrays]
-             || [check_effective_target_natural_alignment_32] } {
+        if { ([check_effective_target_vect_aligned_arrays]
+              || [check_effective_target_natural_alignment_32])
+	     && !([istarget arm*-*-*] && [check_effective_target_arm_neon]) } {
             set et_vector_alignment_reachable_saved 1
         } else {
             set et_vector_alignment_reachable_saved 0