[gcc r12-2385] x86: Don't issue vzeroupper if callee returns AVX register

[gcc r12-2385] x86: Don't issue vzeroupper if callee returns AVX register

[H.J. Lu]

https://gcc.gnu.org/g:5586e7e85de381f3df843091494889f8ad8e3e1f

commit r12-2385-g5586e7e85de381f3df843091494889f8ad8e3e1f
Author: H.J. Lu <hjl.tools@gmail.com>
Date:   Sun Jul 18 06:12:29 2021 -0700

    x86: Don't issue vzeroupper if callee returns AVX register
    
    Don't issue vzeroupper before function call if callee returns AVX
    register since callee must be compiled with AVX.
    
    gcc/
    
            PR target/101495
            * config/i386/i386.c (ix86_check_avx_upper_stores): Moved before
            ix86_avx_u128_mode_needed.
            (ix86_avx_u128_mode_needed): Return AVX_U128_DIRTY if callee
            returns AVX register.
    
    gcc/testsuite/
    
            PR target/101495
            * gcc.target/i386/avx-vzeroupper-28.c: New test.

Diff:
---
 gcc/config/i386/i386.c                            | 32 ++++++++++++++---------
 gcc/testsuite/gcc.target/i386/avx-vzeroupper-28.c | 17 ++++++++++++
 2 files changed, 37 insertions(+), 12 deletions(-)

diff --git a/gcc/config/i386/i386.c b/gcc/config/i386/i386.c
index 9d74b7a191b..8481693ab7f 100644
--- a/gcc/config/i386/i386.c
+++ b/gcc/config/i386/i386.c
@@ -14093,6 +14093,18 @@ ix86_check_avx_upper_register (const_rtx exp)
 	  && GET_MODE_BITSIZE (GET_MODE (exp)) > 128);
 }
 
+/* Check if a 256bit or 512bit AVX register is referenced in stores.   */
+
+static void
+ix86_check_avx_upper_stores (rtx dest, const_rtx, void *data)
+ {
+   if (ix86_check_avx_upper_register (dest))
+    {
+      bool *used = (bool *) data;
+      *used = true;
+    }
+ }
+
 /* Return needed mode for entity in optimize_mode_switching pass.  */
 
 static int
@@ -14117,6 +14129,14 @@ ix86_avx_u128_mode_needed (rtx_insn *insn)
 	    }
 	}
 
+      /* Needed mode is set to AVX_U128_CLEAN if there are no 256bit
+	 nor 512bit registers used in the function return register.  */
+      bool avx_upper_reg_found = false;
+      note_stores (insn, ix86_check_avx_upper_stores,
+		   &avx_upper_reg_found);
+      if (avx_upper_reg_found)
+	return AVX_U128_DIRTY;
+
       /* If the function is known to preserve some SSE registers,
 	 RA and previous passes can legitimately rely on that for
 	 modes wider than 256 bits.  It's only safe to issue a
@@ -14217,18 +14237,6 @@ ix86_mode_needed (int entity, rtx_insn *insn)
   return 0;
 }
 
-/* Check if a 256bit or 512bit AVX register is referenced in stores.   */
- 
-static void
-ix86_check_avx_upper_stores (rtx dest, const_rtx, void *data)
- {
-   if (ix86_check_avx_upper_register (dest))
-    {
-      bool *used = (bool *) data;
-      *used = true;
-    }
- } 
-
 /* Calculate mode of upper 128bit AVX registers after the insn.  */
 
 static int
diff --git a/gcc/testsuite/gcc.target/i386/avx-vzeroupper-28.c b/gcc/testsuite/gcc.target/i386/avx-vzeroupper-28.c
new file mode 100644
index 00000000000..381ee9a7f96
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/avx-vzeroupper-28.c
@@ -0,0 +1,17 @@
+/* PR target/101495  */
+/* { dg-do compile } */
+/* { dg-options "-O2 -mavx2 -mtune=generic -dp" } */
+
+#include <immintrin.h>
+
+extern __m256 x, y;
+extern __m256 bar (void);
+
+__m256
+foo ()
+{
+  x = y;
+  return bar ();
+}
+
+/* { dg-final { scan-assembler-not "avx_vzeroupper" } } */