From 0c4ef36360627686a0f3b56d64409ffb8bfbcb8c Mon Sep 17 00:00:00 2001
From: =?utf8?q?Kristian=20H=C3=B8gsberg=20Kristensen?=
 <kristian.h.kristensen@intel.com>
Date: Fri, 29 Jan 2016 12:10:12 -0800
Subject: [PATCH] anv: clflush is only orderered against mfence

We can't use the more fine-grained load and store fence commands (lfence
and mfence), since clflush is only guaranteed to be ordered with respect
to mfence.
---
 src/vulkan/anv_batch_chain.c |  4 ++--
 src/vulkan/anv_device.c      | 17 +++++++++--------
 src/vulkan/anv_private.h     |  2 +-
 3 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/src/vulkan/anv_batch_chain.c b/src/vulkan/anv_batch_chain.c
index e9bd67c9442..d74c5995168 100644
--- a/src/vulkan/anv_batch_chain.c
+++ b/src/vulkan/anv_batch_chain.c
@@ -755,7 +755,7 @@ anv_cmd_buffer_add_secondary(struct anv_cmd_buffer *primary,
       if (!primary->device->info.has_llc) {
          void *inst = secondary->batch.next - inst_size;
          void *p = (void *) (((uintptr_t) inst) & ~CACHELINE_MASK);
-         __builtin_ia32_sfence();
+         __builtin_ia32_mfence();
          while (p < secondary->batch.next) {
             __builtin_ia32_clflush(p);
             p += CACHELINE_SIZE;
@@ -1047,7 +1047,7 @@ anv_cmd_buffer_prepare_execbuf(struct anv_cmd_buffer *cmd_buffer)
    anv_cmd_buffer_process_relocs(cmd_buffer, &cmd_buffer->surface_relocs);
 
    if (!cmd_buffer->device->info.has_llc) {
-      __builtin_ia32_sfence();
+      __builtin_ia32_mfence();
       anv_vector_foreach(bbo, &cmd_buffer->seen_bbos) {
          for (uint32_t i = 0; i < (*bbo)->length; i += CACHELINE_SIZE)
             __builtin_ia32_clflush((*bbo)->bo.map + i);
diff --git a/src/vulkan/anv_device.c b/src/vulkan/anv_device.c
index c7a9fd15c1d..5bb9fec0085 100644
--- a/src/vulkan/anv_device.c
+++ b/src/vulkan/anv_device.c
@@ -1173,7 +1173,7 @@ VkResult anv_FlushMappedMemoryRanges(
       return VK_SUCCESS;
 
    /* Make sure the writes we're flushing have landed. */
-   __builtin_ia32_sfence();
+   __builtin_ia32_mfence();
 
    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
 
@@ -1193,7 +1193,7 @@ VkResult anv_InvalidateMappedMemoryRanges(
    clflush_mapped_ranges(device, memoryRangeCount, pMemoryRanges);
 
    /* Make sure no reads get moved up above the invalidate. */
-   __builtin_ia32_lfence();
+   __builtin_ia32_mfence();
 
    return VK_SUCCESS;
 }
@@ -1342,7 +1342,7 @@ VkResult anv_CreateFence(
    if (!device->info.has_llc) {
       assert(((uintptr_t) fence->bo.map & CACHELINE_MASK) == 0);
       assert(batch.next - fence->bo.map <= CACHELINE_SIZE);
-      __builtin_ia32_sfence();
+      __builtin_ia32_mfence();
       __builtin_ia32_clflush(fence->bo.map);
    }
 
@@ -1510,7 +1510,7 @@ VkResult anv_CreateEvent(
 
    if (!device->info.has_llc) {
       /* Make sure the writes we're flushing have landed. */
-      __builtin_ia32_sfence();
+      __builtin_ia32_mfence();
       __builtin_ia32_clflush(event);
    }
 
@@ -1538,9 +1538,10 @@ VkResult anv_GetEventStatus(
    ANV_FROM_HANDLE(anv_event, event, _event);
 
    if (!device->info.has_llc) {
-      /* Make sure the writes we're flushing have landed. */
+      /* Invalidate read cache before reading event written by GPU. */
       __builtin_ia32_clflush(event);
-      __builtin_ia32_lfence();
+      __builtin_ia32_mfence();
+
    }
 
    return event->semaphore;
@@ -1557,7 +1558,7 @@ VkResult anv_SetEvent(
 
    if (!device->info.has_llc) {
       /* Make sure the writes we're flushing have landed. */
-      __builtin_ia32_sfence();
+      __builtin_ia32_mfence();
       __builtin_ia32_clflush(event);
    }
 
@@ -1575,7 +1576,7 @@ VkResult anv_ResetEvent(
 
    if (!device->info.has_llc) {
       /* Make sure the writes we're flushing have landed. */
-      __builtin_ia32_sfence();
+      __builtin_ia32_mfence();
       __builtin_ia32_clflush(event);
    }
 
diff --git a/src/vulkan/anv_private.h b/src/vulkan/anv_private.h
index b1d4577f93e..c5ce1484bc2 100644
--- a/src/vulkan/anv_private.h
+++ b/src/vulkan/anv_private.h
@@ -433,7 +433,7 @@ anv_state_clflush(struct anv_state state)
    void *end = state.map + state.alloc_size;
    void *p = (void *) (((uintptr_t) state.map) & ~CACHELINE_MASK);
 
-   __builtin_ia32_sfence();
+   __builtin_ia32_mfence();
    while (p < end) {
       __builtin_ia32_clflush(p);
       p += CACHELINE_SIZE;
-- 
2.30.2