[PATCH] Fix out-of-ssa with unsupported vector types (PR rtl-optimization/90756)

[PATCH] Fix out-of-ssa with unsupported vector types (PR rtl-optimization/90756)

[Jakub Jelinek]

Hi!

This weird hink has been added by Alex in r228175, it isn't clear why
nor how it ever can be correct.  While say for DECL_MODE we have the problem
that for global vars when switching between functions with different ISA
selections the mode might not be ok, TYPE_MODE is stored as a raw vector
mode that a function overrides to BLKmode if that particular vector mode is
not supported.  This hunk breaks that assumption and leaks unsupported
vector modes in the IL of the functions which then have no way to handle
that, but even before that happens usually it breaks because we try to
convert_mode between BLKmode and the unsupported vector mode or vice versa
on PHI nodes.

Alex, do you remember why this has been done?

Patch has been bootstrapped/regtested on x86_64-linux and i686-linux (the
latter didn't have SSE enabled by default), Jeff said he'll test it on many
crosses.  Ok for trunk if that testing succeeds?

2019-07-03  Jakub Jelinek  <jakub@redhat.com>

	PR rtl-optimization/90756
	* explow.c (promote_ssa_mode): Always use TYPE_MODE, don't bypass it
	for VECTOR_TYPE_P.

	* gcc.dg/pr90756.c: New test.

--- gcc/explow.c.jj	2019-06-26 13:51:54.289358743 +0200
+++ gcc/explow.c	2019-07-02 20:32:11.318360013 +0200
@@ -892,16 +892,7 @@ promote_ssa_mode (const_tree name, int *
 
   tree type = TREE_TYPE (name);
   int unsignedp = TYPE_UNSIGNED (type);
-  machine_mode mode = TYPE_MODE (type);
-
-  /* Bypass TYPE_MODE when it maps vector modes to BLKmode.  */
-  if (mode == BLKmode)
-    {
-      gcc_assert (VECTOR_TYPE_P (type));
-      mode = type->type_common.mode;
-    }
-
-  machine_mode pmode = promote_mode (type, mode, &unsignedp);
+  machine_mode pmode = promote_mode (type, TYPE_MODE (type), &unsignedp);
   if (punsignedp)
     *punsignedp = unsignedp;
 
--- gcc/testsuite/gcc.dg/pr90756.c.jj	2019-07-02 20:30:17.651131684 +0200
+++ gcc/testsuite/gcc.dg/pr90756.c	2019-07-02 20:29:49.038574696 +0200
@@ -0,0 +1,26 @@
+/* PR rtl-optimization/90756 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -Wno-psabi" } */
+/* { dg-additional-options "-mno-sse" { target ia32 } } */
+
+typedef float B __attribute__((vector_size(4 * sizeof (float))));
+typedef unsigned long long C __attribute__((vector_size(4 * sizeof (long long))));
+typedef short D __attribute__((vector_size(4 * sizeof (short))));
+B z;
+void foo (C);
+C bar (D);
+B baz ();
+D qux (B);
+
+void
+quux (int x)
+{
+  B n = z, b = z;
+  while (1)
+    switch (x)
+      {
+      case 0: n = baz (); /* FALLTHRU */
+      case 1: { B o = n; n = b; b = o; } /* FALLTHRU */
+      case 2: { D u = qux (b); C v = bar (u); foo (v); }
+      }
+}

	Jakub

Re: [PATCH] Fix out-of-ssa with unsupported vector types (PR rtl-optimization/90756)

[Richard Biener]

[-- Attachment #1: Type: text/plain, Size: 3038 bytes --]

On Wed, 3 Jul 2019, Jakub Jelinek wrote:

> Hi!
> 
> This weird hink has been added by Alex in r228175, it isn't clear why
> nor how it ever can be correct.  While say for DECL_MODE we have the problem
> that for global vars when switching between functions with different ISA
> selections the mode might not be ok, TYPE_MODE is stored as a raw vector
> mode that a function overrides to BLKmode if that particular vector mode is
> not supported.  This hunk breaks that assumption and leaks unsupported
> vector modes in the IL of the functions which then have no way to handle
> that, but even before that happens usually it breaks because we try to
> convert_mode between BLKmode and the unsupported vector mode or vice versa
> on PHI nodes.
> 
> Alex, do you remember why this has been done?
> 
> Patch has been bootstrapped/regtested on x86_64-linux and i686-linux (the
> latter didn't have SSE enabled by default), Jeff said he'll test it on many
> crosses.  Ok for trunk if that testing succeeds?

OK.

Thanks,
Richard.

> 2019-07-03  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR rtl-optimization/90756
> 	* explow.c (promote_ssa_mode): Always use TYPE_MODE, don't bypass it
> 	for VECTOR_TYPE_P.
> 
> 	* gcc.dg/pr90756.c: New test.
> 
> --- gcc/explow.c.jj	2019-06-26 13:51:54.289358743 +0200
> +++ gcc/explow.c	2019-07-02 20:32:11.318360013 +0200
> @@ -892,16 +892,7 @@ promote_ssa_mode (const_tree name, int *
>  
>    tree type = TREE_TYPE (name);
>    int unsignedp = TYPE_UNSIGNED (type);
> -  machine_mode mode = TYPE_MODE (type);
> -
> -  /* Bypass TYPE_MODE when it maps vector modes to BLKmode.  */
> -  if (mode == BLKmode)
> -    {
> -      gcc_assert (VECTOR_TYPE_P (type));
> -      mode = type->type_common.mode;
> -    }
> -
> -  machine_mode pmode = promote_mode (type, mode, &unsignedp);
> +  machine_mode pmode = promote_mode (type, TYPE_MODE (type), &unsignedp);
>    if (punsignedp)
>      *punsignedp = unsignedp;
>  
> --- gcc/testsuite/gcc.dg/pr90756.c.jj	2019-07-02 20:30:17.651131684 +0200
> +++ gcc/testsuite/gcc.dg/pr90756.c	2019-07-02 20:29:49.038574696 +0200
> @@ -0,0 +1,26 @@
> +/* PR rtl-optimization/90756 */
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -Wno-psabi" } */
> +/* { dg-additional-options "-mno-sse" { target ia32 } } */
> +
> +typedef float B __attribute__((vector_size(4 * sizeof (float))));
> +typedef unsigned long long C __attribute__((vector_size(4 * sizeof (long long))));
> +typedef short D __attribute__((vector_size(4 * sizeof (short))));
> +B z;
> +void foo (C);
> +C bar (D);
> +B baz ();
> +D qux (B);
> +
> +void
> +quux (int x)
> +{
> +  B n = z, b = z;
> +  while (1)
> +    switch (x)
> +      {
> +      case 0: n = baz (); /* FALLTHRU */
> +      case 1: { B o = n; n = b; b = o; } /* FALLTHRU */
> +      case 2: { D u = qux (b); C v = bar (u); foo (v); }
> +      }
> +}
> 
> 	Jakub
> 

-- 
Richard Biener <rguenther@suse.de>
SUSE Linux GmbH, Maxfeldstrasse 5, 90409 Nuernberg, Germany;
GF: Felix Imendörffer, Mary Higgins, Sri Rasiah; HRB 21284 (AG NÌrnberg)

Re: [PATCH] Fix out-of-ssa with unsupported vector types (PR rtl-optimization/90756)

[Jeff Law]

On 7/3/19 1:59 AM, Jakub Jelinek wrote:
> Hi!
> 
> This weird hink has been added by Alex in r228175, it isn't clear why
> nor how it ever can be correct.  While say for DECL_MODE we have the problem
> that for global vars when switching between functions with different ISA
> selections the mode might not be ok, TYPE_MODE is stored as a raw vector
> mode that a function overrides to BLKmode if that particular vector mode is
> not supported.  This hunk breaks that assumption and leaks unsupported
> vector modes in the IL of the functions which then have no way to handle
> that, but even before that happens usually it breaks because we try to
> convert_mode between BLKmode and the unsupported vector mode or vice versa
> on PHI nodes.
> 
> Alex, do you remember why this has been done?
> 
> Patch has been bootstrapped/regtested on x86_64-linux and i686-linux (the
> latter didn't have SSE enabled by default), Jeff said he'll test it on many
> crosses.  Ok for trunk if that testing succeeds?
> 
> 2019-07-03  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR rtl-optimization/90756
> 	* explow.c (promote_ssa_mode): Always use TYPE_MODE, don't bypass it
> 	for VECTOR_TYPE_P.
> 
> 	* gcc.dg/pr90756.c: New test.
Nothing tripped related to this patch in the various targets in my tester.

jeff