Attachment 43408 Details for Bug 33069 – 0001-Workaround-for-_mm_empty-related-miscompilation.patch

[patch] 0001-Workaround-for-_mm_empty-related-miscompilation.patch

0001-Workaround-for-_mm_empty-related-miscompilation-prob.patch (text/plain), 5.42 KB, created by Siarhei Siamashka on 2011-02-15 15:40:51 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Siarhei Siamashka

Created: 2011-02-15 15:40:51 UTC

Size: 5.42 KB

patch

obsolete

>From 7657d503b0b1ec8b235910f58602ed631e3fa18f Mon Sep 17 00:00:00 2001
>From: Siarhei Siamashka <siarhei.siamashka@nokia.com>
>Date: Wed, 16 Feb 2011 01:21:49 +0200
>Subject: [PATCH] Workaround for _mm_empty() related miscompilation problem
>
>In some cases floating point registers may get corrupted and as a result
>pixman tests fail. The most likely cause is some kind of miscompilation
>of inlined function containing MMX/SSE2 code:
>    https://bugs.freedesktop.org/show_bug.cgi?id=33069
>
>This patch just puts MMX/SSE2 code into a normal function which now
>gets called from an additional small inline wrapper function. The
>size of compiled code also becomes smaller.
>---
> pixman/pixman-sse2.c |   60 +++++++++++++++++++++++++++++++++++---------------
> 1 files changed, 42 insertions(+), 18 deletions(-)
>
>diff --git a/pixman/pixman-sse2.c b/pixman/pixman-sse2.c
>index 2e135e2..c323d1f 100644
>--- a/pixman/pixman-sse2.c
>+++ b/pixman/pixman-sse2.c
>@@ -5794,14 +5794,12 @@ sse2_composite_over_8888_8888_8888 (pixman_implementation_t *imp,
> }
> 
> /* A variant of 'core_combine_over_u_sse2' with minor tweaks */
>-static force_inline void
>+static void
> scaled_nearest_scanline_sse2_8888_8888_OVER (uint32_t*       pd,
>                                              const uint32_t* ps,
>                                              int32_t         w,
>                                              pixman_fixed_t  vx,
>-                                             pixman_fixed_t  unit_x,
>-                                             pixman_fixed_t  max_vx,
>-                                             pixman_bool_t   fully_transparent_src)
>+                                             pixman_fixed_t  unit_x)
> {
>     uint32_t s, d;
>     const uint32_t* pm = NULL;
>@@ -5810,9 +5808,6 @@ scaled_nearest_scanline_sse2_8888_8888_OVER (uint32_t*       pd,
>     __m128i xmm_src_lo, xmm_src_hi;
>     __m128i xmm_alpha_lo, xmm_alpha_hi;
> 
>-    if (fully_transparent_src)
>-	return;
>-
>     /* Align dst on a 16-byte boundary */
>     while (w && ((unsigned long)pd & 15))
>     {
>@@ -5888,32 +5883,45 @@ scaled_nearest_scanline_sse2_8888_8888_OVER (uint32_t*       pd,
>     _mm_empty ();
> }
> 
>+static force_inline void
>+scaled_nearest_scanline_sse2_8888_8888_OVER_wrapper (uint32_t*       pd,
>+                                                     const uint32_t* ps,
>+                                                     int32_t         w,
>+                                                     pixman_fixed_t  vx,
>+                                                     pixman_fixed_t  unit_x,
>+                                                     pixman_fixed_t  max_vx,
>+                                                     pixman_bool_t   zero_src)
>+{
>+    if (zero_src)
>+	return;
>+    scaled_nearest_scanline_sse2_8888_8888_OVER (pd, ps, w, vx, unit_x);
>+}
>+
>+
> FAST_NEAREST_MAINLOOP (sse2_8888_8888_cover_OVER,
>-		       scaled_nearest_scanline_sse2_8888_8888_OVER,
>+		       scaled_nearest_scanline_sse2_8888_8888_OVER_wrapper,
> 		       uint32_t, uint32_t, COVER)
> FAST_NEAREST_MAINLOOP (sse2_8888_8888_none_OVER,
>-		       scaled_nearest_scanline_sse2_8888_8888_OVER,
>+		       scaled_nearest_scanline_sse2_8888_8888_OVER_wrapper,
> 		       uint32_t, uint32_t, NONE)
> FAST_NEAREST_MAINLOOP (sse2_8888_8888_pad_OVER,
>-		       scaled_nearest_scanline_sse2_8888_8888_OVER,
>+		       scaled_nearest_scanline_sse2_8888_8888_OVER_wrapper,
> 		       uint32_t, uint32_t, PAD)
> 
>-static force_inline void
>+static void
> scaled_nearest_scanline_sse2_8888_n_8888_OVER (const uint32_t * mask,
> 					       uint32_t *       dst,
> 					       const uint32_t * src,
> 					       int32_t          w,
> 					       pixman_fixed_t   vx,
>-					       pixman_fixed_t   unit_x,
>-					       pixman_fixed_t   max_vx,
>-					       pixman_bool_t    zero_src)
>+					       pixman_fixed_t   unit_x)
> {
>     __m128i xmm_mask;
>     __m128i xmm_src, xmm_src_lo, xmm_src_hi;
>     __m128i xmm_dst, xmm_dst_lo, xmm_dst_hi;
>     __m128i xmm_alpha_lo, xmm_alpha_hi;
> 
>-    if (zero_src || (*mask >> 24) == 0)
>+    if ((*mask >> 24) == 0)
> 	return;
> 
>     xmm_mask = create_mask_16_128 (*mask >> 24);
>@@ -6001,14 +6009,30 @@ scaled_nearest_scanline_sse2_8888_n_8888_OVER (const uint32_t * mask,
>     _mm_empty ();
> }
> 
>+static force_inline void
>+scaled_nearest_scanline_sse2_8888_n_8888_OVER_wrapper (const uint32_t * mask,
>+						       uint32_t *       dst,
>+						       const uint32_t * src,
>+						       int32_t          w,
>+						       pixman_fixed_t   vx,
>+						       pixman_fixed_t   unit_x,
>+						       pixman_fixed_t   max_vx,
>+						       pixman_bool_t    zero_src)
>+{
>+    if (zero_src)
>+	return;
>+    scaled_nearest_scanline_sse2_8888_n_8888_OVER (mask, dst, src, w,
>+						   vx, unit_x);
>+}
>+
> FAST_NEAREST_MAINLOOP_COMMON (sse2_8888_n_8888_cover_OVER,
>-			      scaled_nearest_scanline_sse2_8888_n_8888_OVER,
>+			      scaled_nearest_scanline_sse2_8888_n_8888_OVER_wrapper,
> 			      uint32_t, uint32_t, uint32_t, COVER, TRUE, TRUE)
> FAST_NEAREST_MAINLOOP_COMMON (sse2_8888_n_8888_pad_OVER,
>-			      scaled_nearest_scanline_sse2_8888_n_8888_OVER,
>+			      scaled_nearest_scanline_sse2_8888_n_8888_OVER_wrapper,
> 			      uint32_t, uint32_t, uint32_t, PAD, TRUE, TRUE)
> FAST_NEAREST_MAINLOOP_COMMON (sse2_8888_n_8888_none_OVER,
>-			      scaled_nearest_scanline_sse2_8888_n_8888_OVER,
>+			      scaled_nearest_scanline_sse2_8888_n_8888_OVER_wrapper,
> 			      uint32_t, uint32_t, uint32_t, NONE, TRUE, TRUE)
> 
> static const pixman_fast_path_t sse2_fast_paths[] =
>-- 
>1.7.3.4
>

Actions: View | Splinter Review

Attachments on bug 33069: 41978 | 41979 | 42143 | 42144 | 43408

Use of freedesktop.org services, including Bugzilla, is subject to our Code of Conduct. How we collect and use information is described in our Privacy Policy.