Remove alignment constraint on SkMemoryCtxPatch
The compiler was generating some code in SkRasterPipeline::run
like:
vmovaps %ymm0,0xc0(%rcx,%rdi,1)
which assumes memory that is aligned to 32 bytes (256 bits) but
this was not the case because AutoSTMalloc allocates an internal
buffer that is only 4 byte aligned (it's a buffer of uint32_t).
By removing this, the compiler has to generate the unaligned version
vmovups %ymm0,0xe0(%rcx,%rdi,1)
which may be slower, but doesn't crash.
Change-Id: Ie61ffe7d2b14b0319cb2e15693243a4db10f30f0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/1030296
Reviewed-by: Ben Wagner <bungeman@google.com>
Commit-Queue: Kaylee Lubick <kjlubick@google.com>
diff --git a/src/core/SkRasterPipelineOpContexts.h b/src/core/SkRasterPipelineOpContexts.h
index 783e81a..fd1bb8c 100644
--- a/src/core/SkRasterPipelineOpContexts.h
+++ b/src/core/SkRasterPipelineOpContexts.h
@@ -60,10 +60,7 @@
bool store;
};
-// Some SIMD instructions operate faster if we read from aligned memory. 64 bytes (512 bits) is
-// the widest we have (AVX-512), so if we have the scratch field be first and the whole struct
-// aligned that way, the memory for our tail pixels should also be aligned to 64 bytes.
-struct alignas(64) MemoryCtxPatch {
+struct MemoryCtxPatch {
std::byte scratch[kMaxScratchPerPatch];
MemoryCtxInfo info;
