diff mbox

PATCH: Split AVX 32byte unalignd load/store

Message ID AANLkTi=2d9W+u86CLmAHzEh91=AMaZrzPpAzLiZ_6-=V@mail.gmail.com
State New
Headers show

Commit Message

H.J. Lu March 27, 2011, 6:57 p.m. UTC
On Sun, Mar 27, 2011 at 10:53 AM, Uros Bizjak <ubizjak@gmail.com> wrote:
> On Sun, Mar 27, 2011 at 3:44 PM, H.J. Lu <hjl.tools@gmail.com> wrote:
>
>> Here is a patch to split AVX 32byte unalignd load/store:
>>
>> http://gcc.gnu.org/ml/gcc-patches/2011-02/msg00743.html
>>
>> It speeds up some SPEC CPU 2006 benchmarks by up to 6%.
>> OK for trunk?
>
>> 2011-02-11  H.J. Lu  <hongjiu.lu@intel.com>
>>
>>       * config/i386/i386.c (flag_opts): Add -mavx256-split-unaligned-load
>>       and -mavx256-split-unaligned-store.
>>       (ix86_option_override_internal): Split 32-byte AVX unaligned
>>       load/store by default.
>>       (ix86_avx256_split_vector_move_misalign): New.
>>       (ix86_expand_vector_move_misalign): Use it.
>>
>>       * config/i386/i386.opt: Add -mavx256-split-unaligned-load and
>>       -mavx256-split-unaligned-store.
>>
>>       * config/i386/sse.md (*avx_mov<mode>_internal): Verify unaligned
>>       256bit load/store.  Generate unaligned store on misaligned memory
>>       operand.
>>       (*avx_movu<ssemodesuffix><avxmodesuffix>): Verify unaligned
>>       256bit load/store.
>>       (*avx_movdqu<avxmodesuffix>): Likewise.
>>
>>       * doc/invoke.texi: Document -mavx256-split-unaligned-load and
>>       -mavx256-split-unaligned-store.
>>
>> gcc/testsuite/
>>
>> 2011-02-11  H.J. Lu  <hongjiu.lu@intel.com>
>>
>>       * gcc.target/i386/avx256-unaligned-load-1.c: New.
>>       * gcc.target/i386/avx256-unaligned-load-2.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-load-3.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-load-4.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-load-5.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-load-6.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-load-7.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-1.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-2.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-3.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-4.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-5.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-6.c: Likewise.
>>       * gcc.target/i386/avx256-unaligned-store-7.c: Likewise.
>>
>
>
>
>> @@ -203,19 +203,37 @@
>>        return standard_sse_constant_opcode (insn, operands[1]);
>>      case 1:
>>      case 2:
>> +      if (GET_MODE_ALIGNMENT (<MODE>mode) == 256
>> +       && ((TARGET_AVX256_SPLIT_UNALIGNED_STORE
>> +            && MEM_P (operands[0])
>> +            && MEM_ALIGN (operands[0]) < 256)
>> +           || (TARGET_AVX256_SPLIT_UNALIGNED_LOAD
>> +               && MEM_P (operands[1])
>> +               && MEM_ALIGN (operands[1]) < 256)))
>> +     gcc_unreachable ();
>
> Please use "misaligned_operand (operands[...], <MODE>mode)" instead of
> MEM_P && MEM_ALIGN combo in a couple of places.
>
> OK with that change.
>

This is the patch I checked in.

Thanks.

Comments

Richard Henderson March 28, 2011, 8:17 p.m. UTC | #1
On 03/27/2011 11:57 AM, H.J. Lu wrote:
> +    { "-mavx256-split-unaligned-stroe",	MASK_AVX256_SPLIT_UNALIGNED_STORE},

Typo.


r~
diff mbox

Patch

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 4e8ca69..a4ca762 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -3130,6 +3130,8 @@  ix86_target_string (int isa, int flags, const char *arch, const char *tune,
     { "-mvect8-ret-in-mem",		MASK_VECT8_RETURNS },
     { "-m8bit-idiv",			MASK_USE_8BIT_IDIV },
     { "-mvzeroupper",			MASK_VZEROUPPER },
+    { "-mavx256-split-unaligned-load",	MASK_AVX256_SPLIT_UNALIGNED_LOAD},
+    { "-mavx256-split-unaligned-stroe",	MASK_AVX256_SPLIT_UNALIGNED_STORE},
   };
 
   const char *opts[ARRAY_SIZE (isa_opts) + ARRAY_SIZE (flag_opts) + 6][2];
@@ -4274,11 +4276,18 @@  ix86_option_override_internal (bool main_args_p)
   if (TARGET_AVX)
     {
       /* When not optimize for size, enable vzeroupper optimization for
-	 TARGET_AVX with -fexpensive-optimizations.  */
-      if (!optimize_size
-	  && flag_expensive_optimizations
-	  && !(target_flags_explicit & MASK_VZEROUPPER))
-	target_flags |= MASK_VZEROUPPER;
+	 TARGET_AVX with -fexpensive-optimizations and split 32-byte
+	 AVX unaligned load/store.  */
+      if (!optimize_size)
+	{
+	  if (flag_expensive_optimizations
+	      && !(target_flags_explicit & MASK_VZEROUPPER))
+	    target_flags |= MASK_VZEROUPPER;
+	  if (!(target_flags_explicit & MASK_AVX256_SPLIT_UNALIGNED_LOAD))
+	    target_flags |= MASK_AVX256_SPLIT_UNALIGNED_LOAD;
+	  if (!(target_flags_explicit & MASK_AVX256_SPLIT_UNALIGNED_STORE))
+	    target_flags |= MASK_AVX256_SPLIT_UNALIGNED_STORE;
+	}
     }
   else 
     {
@@ -15588,6 +15597,57 @@  ix86_expand_vector_move (enum machine_mode mode, rtx operands[])
   emit_insn (gen_rtx_SET (VOIDmode, op0, op1));
 }
 
+/* Split 32-byte AVX unaligned load and store if needed.  */
+
+static void
+ix86_avx256_split_vector_move_misalign (rtx op0, rtx op1)
+{
+  rtx m;
+  rtx (*extract) (rtx, rtx, rtx);
+  rtx (*move_unaligned) (rtx, rtx);
+  enum machine_mode mode;
+
+  switch (GET_MODE (op0))
+    {
+    default:
+      gcc_unreachable ();
+    case V32QImode:
+      extract = gen_avx_vextractf128v32qi;
+      move_unaligned = gen_avx_movdqu256;
+      mode = V16QImode;
+      break;
+    case V8SFmode:
+      extract = gen_avx_vextractf128v8sf;
+      move_unaligned = gen_avx_movups256;
+      mode = V4SFmode;
+      break;
+    case V4DFmode:
+      extract = gen_avx_vextractf128v4df;
+      move_unaligned = gen_avx_movupd256;
+      mode = V2DFmode;
+      break;
+    }
+
+  if (MEM_P (op1) && TARGET_AVX256_SPLIT_UNALIGNED_LOAD)
+    {
+      rtx r = gen_reg_rtx (mode);
+      m = adjust_address (op1, mode, 0);
+      emit_move_insn (r, m);
+      m = adjust_address (op1, mode, 16);
+      r = gen_rtx_VEC_CONCAT (GET_MODE (op0), r, m);
+      emit_move_insn (op0, r);
+    }
+  else if (MEM_P (op0) && TARGET_AVX256_SPLIT_UNALIGNED_STORE)
+    {
+      m = adjust_address (op0, mode, 0);
+      emit_insn (extract (m, op1, const0_rtx));
+      m = adjust_address (op0, mode, 16);
+      emit_insn (extract (m, op1, const1_rtx));
+    }
+  else
+    emit_insn (move_unaligned (op0, op1));
+}
+
 /* Implement the movmisalign patterns for SSE.  Non-SSE modes go
    straight to ix86_expand_vector_move.  */
 /* Code generation for scalar reg-reg moves of single and double precision data:
@@ -15672,7 +15732,7 @@  ix86_expand_vector_move_misalign (enum machine_mode mode, rtx operands[])
 	    case 32:
 	      op0 = gen_lowpart (V32QImode, op0);
 	      op1 = gen_lowpart (V32QImode, op1);
-	      emit_insn (gen_avx_movdqu256 (op0, op1));
+	      ix86_avx256_split_vector_move_misalign (op0, op1);
 	      break;
 	    default:
 	      gcc_unreachable ();
@@ -15688,7 +15748,7 @@  ix86_expand_vector_move_misalign (enum machine_mode mode, rtx operands[])
 	      emit_insn (gen_avx_movups (op0, op1));
 	      break;
 	    case V8SFmode:
-	      emit_insn (gen_avx_movups256 (op0, op1));
+	      ix86_avx256_split_vector_move_misalign (op0, op1);
 	      break;
 	    case V2DFmode:
 	      if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
@@ -15701,7 +15761,7 @@  ix86_expand_vector_move_misalign (enum machine_mode mode, rtx operands[])
 	      emit_insn (gen_avx_movupd (op0, op1));
 	      break;
 	    case V4DFmode:
-	      emit_insn (gen_avx_movupd256 (op0, op1));
+	      ix86_avx256_split_vector_move_misalign (op0, op1);
 	      break;
 	    default:
 	      gcc_unreachable ();
diff --git a/gcc/config/i386/i386.opt b/gcc/config/i386/i386.opt
index e02d098..f63a406 100644
--- a/gcc/config/i386/i386.opt
+++ b/gcc/config/i386/i386.opt
@@ -420,3 +420,11 @@  Emit profiling counter call at function entry before prologue.
 m8bit-idiv
 Target Report Mask(USE_8BIT_IDIV) Save
 Expand 32bit/64bit integer divide into 8bit unsigned integer divide with run-time check
+
+mavx256-split-unaligned-load
+Target Report Mask(AVX256_SPLIT_UNALIGNED_LOAD) Save
+Split 32-byte AVX unaligned load
+
+mavx256-split-unaligned-store
+Target Report Mask(AVX256_SPLIT_UNALIGNED_STORE) Save
+Split 32-byte AVX unaligned store
diff --git a/gcc/config/i386/sse.md b/gcc/config/i386/sse.md
index 70a0b34..de11f73 100644
--- a/gcc/config/i386/sse.md
+++ b/gcc/config/i386/sse.md
@@ -203,19 +203,35 @@ 
       return standard_sse_constant_opcode (insn, operands[1]);
     case 1:
     case 2:
+      if (GET_MODE_ALIGNMENT (<MODE>mode) == 256
+	  && ((TARGET_AVX256_SPLIT_UNALIGNED_STORE
+	       && misaligned_operand (operands[0], <MODE>mode))
+	      || (TARGET_AVX256_SPLIT_UNALIGNED_LOAD
+		  && misaligned_operand (operands[1], <MODE>mode))))
+	gcc_unreachable ();
       switch (get_attr_mode (insn))
         {
 	case MODE_V8SF:
 	case MODE_V4SF:
-	  return "vmovaps\t{%1, %0|%0, %1}";
+	  if (misaligned_operand (operands[0], <MODE>mode)
+	      || misaligned_operand (operands[1], <MODE>mode))
+	    return "vmovups\t{%1, %0|%0, %1}";
+	  else
+	    return "vmovaps\t{%1, %0|%0, %1}";
 	case MODE_V4DF:
 	case MODE_V2DF:
-	  if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
+	  if (misaligned_operand (operands[0], <MODE>mode)
+	      || misaligned_operand (operands[1], <MODE>mode))
+	    return "vmovupd\t{%1, %0|%0, %1}";
+	  else if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
 	    return "vmovaps\t{%1, %0|%0, %1}";
 	  else
 	    return "vmovapd\t{%1, %0|%0, %1}";
 	default:
-	  if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
+	  if (misaligned_operand (operands[0], <MODE>mode)
+	      || misaligned_operand (operands[1], <MODE>mode))
+	    return "vmovdqu\t{%1, %0|%0, %1}";
+	  else if (TARGET_SSE_PACKED_SINGLE_INSN_OPTIMAL)
 	    return "vmovaps\t{%1, %0|%0, %1}";
 	  else
 	    return "vmovdqa\t{%1, %0|%0, %1}";
@@ -400,7 +416,15 @@ 
 	  UNSPEC_MOVU))]
   "AVX_VEC_FLOAT_MODE_P (<MODE>mode)
    && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "vmovu<ssemodesuffix>\t{%1, %0|%0, %1}"
+{
+  if (GET_MODE_ALIGNMENT (<MODE>mode) == 256
+      && ((TARGET_AVX256_SPLIT_UNALIGNED_STORE
+	   && misaligned_operand (operands[0], <MODE>mode))
+	  || (TARGET_AVX256_SPLIT_UNALIGNED_LOAD
+	      && misaligned_operand (operands[1], <MODE>mode))))
+    gcc_unreachable ();
+  return "vmovu<ssemodesuffix>\t{%1, %0|%0, %1}";
+}
   [(set_attr "type" "ssemov")
    (set_attr "movu" "1")
    (set_attr "prefix" "vex")
@@ -459,7 +483,15 @@ 
 	  [(match_operand:AVXMODEQI 1 "nonimmediate_operand" "xm,x")]
 	  UNSPEC_MOVU))]
   "TARGET_AVX && !(MEM_P (operands[0]) && MEM_P (operands[1]))"
-  "vmovdqu\t{%1, %0|%0, %1}"
+{
+  if (GET_MODE_ALIGNMENT (<MODE>mode) == 256
+      && ((TARGET_AVX256_SPLIT_UNALIGNED_STORE
+	   && misaligned_operand (operands[0], <MODE>mode))
+	  || (TARGET_AVX256_SPLIT_UNALIGNED_LOAD
+	      && misaligned_operand (operands[1], <MODE>mode))))
+    gcc_unreachable ();
+  return "vmovdqu\t{%1, %0|%0, %1}";
+}
   [(set_attr "type" "ssemov")
    (set_attr "movu" "1")
    (set_attr "prefix" "vex")
diff --git a/gcc/doc/invoke.texi b/gcc/doc/invoke.texi
index 925455d..85bf2b4 100644
--- a/gcc/doc/invoke.texi
+++ b/gcc/doc/invoke.texi
@@ -602,7 +602,8 @@  Objective-C and Objective-C++ Dialects}.
 -momit-leaf-frame-pointer  -mno-red-zone -mno-tls-direct-seg-refs @gol
 -mcmodel=@var{code-model} -mabi=@var{name} @gol
 -m32  -m64 -mlarge-data-threshold=@var{num} @gol
--msse2avx -mfentry -m8bit-idiv}
+-msse2avx -mfentry -m8bit-idiv @gol
+-mavx256-split-unaligned-load -mavx256-split-unaligned-store}
 
 @emph{i386 and x86-64 Windows Options}
 @gccoptlist{-mconsole -mcygwin -mno-cygwin -mdll @gol
@@ -12669,6 +12670,12 @@  runt-time check.  If both dividend and divisor are within range of 0
 to 255, 8bit unsigned integer divide will be used instead of
 32bit/64bit integer divide.
 
+@item -mavx256-split-unaligned-load
+@item -mavx256-split-unaligned-store
+@opindex avx256-split-unaligned-load
+@opindex avx256-split-unaligned-store
+Split 32-byte AVX unaligned load and store.
+
 @end table
 
 These @samp{-m} switches are supported in addition to the above
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c
new file mode 100644
index 0000000..023e859
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-1.c
@@ -0,0 +1,19 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-load" } */
+
+#define N 1024
+
+float a[N], b[N+3], c[N];
+
+void
+avx_test (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    c[i] = a[i] * b[i+3];
+}
+
+/* { dg-final { scan-assembler-not "\\*avx_movups256/1" } } */
+/* { dg-final { scan-assembler "\\*avx_movups/1" } } */
+/* { dg-final { scan-assembler "vinsertf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-2.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-2.c
new file mode 100644
index 0000000..8394e27
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-2.c
@@ -0,0 +1,29 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target lp64 } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-load" } */
+
+#define N 1024
+
+char **ep;
+char **fp;
+
+void
+avx_test (void)
+{
+  int i;
+  char **ap;
+  char **bp;
+  char **cp;
+
+  ap = ep;
+  bp = fp;
+  for (i = 128; i >= 0; i--)
+    {
+      *ap++ = *cp++;
+      *bp++ = 0;
+    }
+}
+
+/* { dg-final { scan-assembler-not "\\*avx_movdqu256/1" } } */
+/* { dg-final { scan-assembler "\\*avx_movdqu/1" } } */
+/* { dg-final { scan-assembler "vinsertf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-3.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-3.c
new file mode 100644
index 0000000..ec7d59d
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-3.c
@@ -0,0 +1,19 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-load" } */
+
+#define N 1024
+
+double a[N], b[N+3], c[N];
+
+void
+avx_test (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    c[i] = a[i] * b[i+3];
+}
+
+/* { dg-final { scan-assembler-not "\\*avx_movupd256/1" } } */
+/* { dg-final { scan-assembler "\\*avx_movupd/1" } } */
+/* { dg-final { scan-assembler "vinsertf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-4.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-4.c
new file mode 100644
index 0000000..0d3ef33
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-4.c
@@ -0,0 +1,19 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O3 -dp -mavx -mno-avx256-split-unaligned-load -mno-avx256-split-unaligned-store" } */
+
+#define N 1024
+
+float a[N], b[N+3];
+
+void
+avx_test (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i] = a[i+3] * 2;
+}
+
+/* { dg-final { scan-assembler "\\*avx_movups256/1" } } */
+/* { dg-final { scan-assembler-not "\\*avx_movups/1" } } */
+/* { dg-final { scan-assembler-not "vinsertf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-5.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-5.c
new file mode 100644
index 0000000..153b66f
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-5.c
@@ -0,0 +1,43 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-load" } */
+
+#include "avx-check.h"
+
+#define N 8
+
+float a[N+3] = { -1, -1, -1, 24.43, 68.346, 43.35,
+		 546.46, 46.79, 82.78, 82.7, 9.4 };
+float b[N];
+float c[N];
+
+void
+foo (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i] = a[i+3] * 2;
+}
+
+__attribute__ ((noinline))
+float
+bar (float x)
+{
+  return x * 2;
+}
+
+void
+avx_test (void)
+{
+  int i;
+
+  foo ();
+
+  for (i = 0; i < N; i++)
+    c[i] = bar (a[i+3]);
+
+  for (i = 0; i < N; i++)
+    if (b[i] != c[i])
+      abort ();
+}
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-6.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-6.c
new file mode 100644
index 0000000..2fa984c
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-6.c
@@ -0,0 +1,42 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-load" } */
+
+#include "avx-check.h"
+
+#define N 4
+
+double a[N+3] = { -1, -1, -1, 24.43, 68.346, 43.35, 546.46 };
+double b[N];
+double c[N];
+
+void
+foo (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i] = a[i+3] * 2;
+}
+
+__attribute__ ((noinline))
+double
+bar (double x)
+{
+  return x * 2;
+}
+
+void
+avx_test (void)
+{
+  int i;
+
+  foo ();
+
+  for (i = 0; i < N; i++)
+    c[i] = bar (a[i+3]);
+
+  for (i = 0; i < N; i++)
+    if (b[i] != c[i])
+      abort ();
+}
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-7.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-7.c
new file mode 100644
index 0000000..ad16a53
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-load-7.c
@@ -0,0 +1,60 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-load" } */
+
+#include "avx-check.h"
+
+#define N 128
+
+char **ep;
+char **fp;
+char **mp;
+char **lp;
+
+__attribute__ ((noinline))
+void
+foo (void)
+{
+  mp = (char **) malloc (N);
+  lp = (char **) malloc (N);
+  ep = (char **) malloc (N);
+  fp = (char **) malloc (N);
+}
+
+void
+avx_test (void)
+{
+  int i;
+  char **ap, **bp, **cp, **dp;
+  char *str = "STR";
+
+  foo ();
+
+  cp = mp;
+  dp = lp;
+
+  for (i = N; i >= 0; i--)
+    {
+      *cp++ = str;
+      *dp++ = str;
+    }
+
+  ap = ep;
+  bp = fp;
+  cp = mp;
+  dp = lp;
+
+  for (i = N; i >= 0; i--)
+    {
+      *ap++ = *cp++;
+      *bp++ = *dp++;
+    }
+
+  for (i = N; i >= 0; i--)
+    {
+      if (strcmp (*--ap, "STR") != 0)
+	abort ();
+      if (strcmp (*--bp, "STR") != 0)
+	abort ();
+    }
+}
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c
new file mode 100644
index 0000000..99db55c
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-1.c
@@ -0,0 +1,22 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-store" } */
+
+#define N 1024
+
+float a[N], b[N+3], c[N], d[N];
+
+void
+avx_test (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i+3] = a[i] * 10.0;
+
+  for (i = 0; i < N; i++)
+    d[i] = c[i] * 20.0;
+}
+
+/* { dg-final { scan-assembler-not "\\*avx_movups256/2" } } */
+/* { dg-final { scan-assembler "movups.*\\*avx_movv4sf_internal/3" } } */
+/* { dg-final { scan-assembler "vextractf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c
new file mode 100644
index 0000000..38ee9e2
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-2.c
@@ -0,0 +1,29 @@ 
+/* { dg-do compile } */
+/* { dg-require-effective-target lp64 } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-store" } */
+
+#define N 1024
+
+char **ep;
+char **fp;
+
+void
+avx_test (void)
+{
+  int i;
+  char **ap;
+  char **bp;
+  char **cp;
+
+  ap = ep;
+  bp = fp;
+  for (i = 128; i >= 0; i--)
+    {
+      *ap++ = *cp++;
+      *bp++ = 0;
+    }
+}
+
+/* { dg-final { scan-assembler-not "\\*avx_movdqu256/2" } } */
+/* { dg-final { scan-assembler "movdqu.*\\*avx_movv16qi_internal/3" } } */
+/* { dg-final { scan-assembler "vextractf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c
new file mode 100644
index 0000000..eaab6fd
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-3.c
@@ -0,0 +1,22 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-store" } */
+
+#define N 1024
+
+double a[N], b[N+3], c[N], d[N];
+
+void
+avx_test (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i+3] = a[i] * 10.0;
+
+  for (i = 0; i < N; i++)
+    d[i] = c[i] * 20.0;
+}
+
+/* { dg-final { scan-assembler-not "\\*avx_movupd256/2" } } */
+/* { dg-final { scan-assembler "movupd.*\\*avx_movv2df_internal/3" } } */
+/* { dg-final { scan-assembler "vextractf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-4.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-4.c
new file mode 100644
index 0000000..96cca66
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-4.c
@@ -0,0 +1,20 @@ 
+/* { dg-do compile } */
+/* { dg-options "-O3 -dp -mavx -mno-avx256-split-unaligned-load -mno-avx256-split-unaligned-store" } */
+
+#define N 1024
+
+float a[N], b[N+3], c[N];
+
+void
+avx_test (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i+3] = a[i] * c[i];
+}
+
+/* { dg-final { scan-assembler "\\*avx_movups256/2" } } */
+/* { dg-final { scan-assembler-not "\\*avx_movups/2" } } */
+/* { dg-final { scan-assembler-not "\\*avx_movv4sf_internal/3" } } */
+/* { dg-final { scan-assembler-not "vextractf128" } } */
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-5.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-5.c
new file mode 100644
index 0000000..642da3c
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-5.c
@@ -0,0 +1,42 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-store" } */
+
+#include "avx-check.h"
+
+#define N 8
+
+float a[N] = { 24.43, 68.346, 43.35, 546.46, 46.79, 82.78, 82.7, 9.4 };
+float b[N+3];
+float c[N+3];
+
+void
+foo (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i+3] = a[i] * 2;
+}
+
+__attribute__ ((noinline))
+float
+bar (float x)
+{
+  return x * 2;
+}
+
+void
+avx_test (void)
+{
+  int i;
+
+  foo ();
+
+  for (i = 0; i < N; i++)
+    c[i+3] = bar (a[i]);
+
+  for (i = 0; i < N; i++)
+    if (b[i+3] != c[i+3])
+      abort ();
+}
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-6.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-6.c
new file mode 100644
index 0000000..a0de7a5
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-6.c
@@ -0,0 +1,42 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-store" } */
+
+#include "avx-check.h"
+
+#define N 4
+
+double a[N] = { 24.43, 68.346, 43.35, 546.46 };
+double b[N+3];
+double c[N+3];
+
+void
+foo (void)
+{
+  int i;
+
+  for (i = 0; i < N; i++)
+    b[i+3] = a[i] * 2;
+}
+
+__attribute__ ((noinline))
+double
+bar (double x)
+{
+  return x * 2;
+}
+
+void
+avx_test (void)
+{
+  int i;
+
+  foo ();
+
+  for (i = 0; i < N; i++)
+    c[i+3] = bar (a[i]);
+
+  for (i = 0; i < N; i++)
+    if (b[i+3] != c[i+3])
+      abort ();
+}
diff --git a/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-7.c b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-7.c
new file mode 100644
index 0000000..4272dc3
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx256-unaligned-store-7.c
@@ -0,0 +1,45 @@ 
+/* { dg-do run } */
+/* { dg-require-effective-target avx } */
+/* { dg-options "-O3 -dp -mavx -mavx256-split-unaligned-store" } */
+
+#include "avx-check.h"
+
+#define N 128
+
+char **ep;
+char **fp;
+
+__attribute__ ((noinline))
+void
+foo (void)
+{
+  ep = (char **) malloc (N);
+  fp = (char **) malloc (N);
+}
+
+void
+avx_test (void)
+{
+  int i;
+  char **ap, **bp;
+  char *str = "STR";
+
+  foo ();
+
+  ap = ep;
+  bp = fp;
+
+  for (i = N; i >= 0; i--)
+    {
+      *ap++ = str;
+      *bp++ = str;
+    }
+
+  for (i = N; i >= 0; i--)
+    {
+      if (strcmp (*--ap, "STR") != 0)
+	abort ();
+      if (strcmp (*--bp, "STR") != 0)
+	abort ();
+    }
+}