After

commit cbc56384029c9224280b0a1018fb9502797f243d
Author: H.J. Lu <[email protected]>
Date:   Fri Jun 26 08:43:20 2026 +0800

    determine_block_size: Set len_rtx to min size if min size == max size

ix86_expand_set_or_cpymem may use vector loop on a block of memory whose
size < epilogue size.  Don't use misaligned prologue if count <= epilogue
size.

gcc/

PR target/126275
* config/i386/i386-expand.cc (ix86_expand_set_or_cpymem): Don't
use misaligned prologue if count <= epilogue size.

gcc/testsuite/

PR target/126275
* gcc.target/i386/pr126275.c: New.


-- 
H.J.
From e638d794f466922cbbcb1dd15ca68324da756207 Mon Sep 17 00:00:00 2001
From: "H.J. Lu" <[email protected]>
Date: Thu, 16 Jul 2026 05:39:33 +0800
Subject: [PATCH] x86: Avoid misaligned prologue if count <= epilogue size

After

commit cbc56384029c9224280b0a1018fb9502797f243d
Author: H.J. Lu <[email protected]>
Date:   Fri Jun 26 08:43:20 2026 +0800

    determine_block_size: Set len_rtx to min size if min size == max size

ix86_expand_set_or_cpymem may use vector loop on a block of memory whose
size < epilogue size.  Don't use misaligned prologue if count <= epilogue
size.

gcc/

	PR target/126275
	* config/i386/i386-expand.cc (ix86_expand_set_or_cpymem): Don't
	use misaligned prologue if count <= epilogue size.

gcc/testsuite/

	PR target/126275
	* gcc.target/i386/pr126275.c: New.

Signed-off-by: H.J. Lu <[email protected]>
---
 gcc/config/i386/i386-expand.cc           | 13 +++++----
 gcc/testsuite/gcc.target/i386/pr126275.c | 36 ++++++++++++++++++++++++
 2 files changed, 44 insertions(+), 5 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/i386/pr126275.c

diff --git a/gcc/config/i386/i386-expand.cc b/gcc/config/i386/i386-expand.cc
index 2f6b401e3ab..1841428b3dc 100644
--- a/gcc/config/i386/i386-expand.cc
+++ b/gcc/config/i386/i386-expand.cc
@@ -9765,11 +9765,14 @@ ix86_expand_set_or_cpymem (rtx dst, rtx src, rtx count_exp, rtx val_exp,
       /* Destination is aligned after the misaligned prologue.  */
       aligned_dstmem = misaligned_prologue_used;
 
-      if (noalign && !misaligned_prologue_used)
-	{
-	  /* Also use misaligned prologue if alignment isn't needed and
-	     destination isn't aligned.   Since alignment isn't needed,
-	     the destination after prologue won't be aligned.  */
+      if (noalign
+	  && !misaligned_prologue_used
+	  && (!count || count > (unsigned int) epilogue_size_needed))
+	{
+	  /* Also use misaligned prologue if count > epilogue size,
+	     alignment isn't needed and destination isn't aligned.
+	     Since alignment isn't needed, the destination after
+	     prologue won't be aligned.  */
 	  aligned_dstmem = (GET_MODE_ALIGNMENT (move_mode)
 			    <= MEM_ALIGN (dst));
 	  if (!aligned_dstmem)
diff --git a/gcc/testsuite/gcc.target/i386/pr126275.c b/gcc/testsuite/gcc.target/i386/pr126275.c
new file mode 100644
index 00000000000..c57cc0ebd0a
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr126275.c
@@ -0,0 +1,36 @@
+/* { dg-do compile { target fpic } } */
+/* { dg-options "-O1 -fno-pic -ftree-loop-distribute-patterns -fno-tree-dominator-opts -fno-tree-scev-cprop -march=x86-64" } */
+
+/* Keep labels and directives ('.cfi_startproc', '.cfi_endproc').  */
+/* { dg-final { check-function-bodies "**" "" "" { target { *-*-linux* && lp64 } } {^\t?\.} } } */
+
+/*
+**func:
+**.LFB0:
+**	.cfi_startproc
+**	movl	\$0, %eax
+**	.p2align 3
+**.L2:
+**	addl	\$1, %eax
+**	cmpl	\$18, %eax
+**	jne	.L2
+**	cltq
+**	movabsq	\$506381209866536711, %rdx
+**	movq	%rdx, %xmm0
+**	punpcklqdq	%xmm0, %xmm0
+**	movups	%xmm0, buf\(%rax\)
+**	movb	\$7, buf\+16\(%rax\)
+**	ret
+**	.cfi_endproc
+**...
+*/
+
+extern char buf[512];
+
+void
+func (void)
+{
+  int i = 0;
+  for (; i < 18; i++);
+  for (; i < 35; i++) buf[i] = 7;
+}
-- 
2.55.0

Reply via email to