Revert of Switch uses of SkChecksum::Compute to Murmur3. (patchset #2 id:20001 of https://codereview.chromium.org/1436973003/ )

Reason for revert:
I would not have expected this to change any images.

Looks like it's changed both MSAA and non-MSAA GMs:
1d024a3c909ae5cefa5e8b339e2b52dc73ee85ac&unt=true&head=true&query=source_type%3Dgm">https://gold.skia.org/search2?blame=1d024a3c909ae5cefa5e8b339e2b52dc73ee85ac&unt=true&head=true&query=source_type%3Dgm

Original issue's description:
> Switch uses of SkChecksum::Compute to Murmur3.
>
> SkChecksum::Compute is a very, very poorly distributed hash function.
> This replaces all remaining uses with Murmur3.
>
> The only interesting stuff is in src/gpu.
>
> BUG=skia:
>
> Committed: https://skia.googlesource.com/skia/+/1d024a3c909ae5cefa5e8b339e2b52dc73ee85ac

TBR=bsalomon@google.com,mtklein@chromium.org
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=skia:

Review URL: https://codereview.chromium.org/1445523003

1 files changed, 89 insertions, 0 deletions

diff --git a/include/private/SkChecksum.h b/include/private/SkChecksum.h
index 6289a444ae..4526416fc1 100644
--- a/include/private/SkChecksum.h
+++ b/include/private/SkChecksum.h
@@ -12,7 +12,31 @@
 #include "SkTLogic.h"
 #include "SkTypes.h"
 
+/**
+ *  Computes a 32bit checksum from a blob of 32bit aligned data. This is meant
+ *  to be very very fast, as it is used internally by the font cache, in
+ *  conjuction with the entire raw key. This algorithm does not generate
+ *  unique values as well as others (e.g. MD5) but it performs much faster.
+ *  Skia's use cases can survive non-unique values (since the entire key is
+ *  always available). Clients should only be used in circumstances where speed
+ *  over uniqueness is at a premium.
+ */
 class SkChecksum : SkNoncopyable {
+private:
+    /*
+     *  Our Rotate and Mash helpers are meant to automatically do the right
+     *  thing depending if sizeof(uintptr_t) is 4 or 8.
+     */
+    enum {
+        ROTR = 17,
+        ROTL = sizeof(uintptr_t) * 8 - ROTR,
+        HALFBITS = sizeof(uintptr_t) * 4
+    };
+
+    static inline uintptr_t Mash(uintptr_t total, uintptr_t value) {
+        return ((total >> ROTR) | (total << ROTL)) ^ value;
+    }
+
 public:
     /**
      * uint32_t -> uint32_t hash, useful for when you're about to trucate this hash but you
@@ -44,6 +68,7 @@ public:
 
     /**
      * Calculate 32-bit Murmur hash (murmur3).
+     * This should take 2-3x longer than SkChecksum::Compute, but is a considerably better hash.
      * See en.wikipedia.org/wiki/MurmurHash.
      *
      *  @param data Memory address of the data block to be processed.
@@ -52,6 +77,70 @@ public:
      *  @return hash result
      */
     static uint32_t Murmur3(const void* data, size_t bytes, uint32_t seed=0);
+
+    /**
+     *  Compute a 32-bit checksum for a given data block
+     *
+     *  WARNING: this algorithm is tuned for efficiency, not backward/forward
+     *  compatibility.  It may change at any time, so a checksum generated with
+     *  one version of the Skia code may not match a checksum generated with
+     *  a different version of the Skia code.
+     *
+     *  @param data Memory address of the data block to be processed. Must be
+     *              32-bit aligned.
+     *  @param size Size of the data block in bytes. Must be a multiple of 4.
+     *  @return checksum result
+     */
+    static uint32_t Compute(const uint32_t* data, size_t size) {
+        // Use may_alias to remind the compiler we're intentionally violating strict aliasing,
+        // and so not to apply strict-aliasing-based optimizations.
+        typedef uint32_t SK_ATTRIBUTE(may_alias) aliased_uint32_t;
+        const aliased_uint32_t* safe_data = (const aliased_uint32_t*)data;
+
+        SkASSERT(SkIsAlign4(size));
+
+        /*
+         *  We want to let the compiler use 32bit or 64bit addressing and math
+         *  so we use uintptr_t as our magic type. This makes the code a little
+         *  more obscure (we can't hard-code 32 or 64 anywhere, but have to use
+         *  sizeof()).
+         */
+        uintptr_t result = 0;
+        const uintptr_t* ptr = reinterpret_cast<const uintptr_t*>(safe_data);
+
+        /*
+         *  count the number of quad element chunks. This takes into account
+         *  if we're on a 32bit or 64bit arch, since we use sizeof(uintptr_t)
+         *  to compute how much to shift-down the size.
+         */
+        size_t n4 = size / (sizeof(uintptr_t) << 2);
+        for (size_t i = 0; i < n4; ++i) {
+            result = Mash(result, *ptr++);
+            result = Mash(result, *ptr++);
+            result = Mash(result, *ptr++);
+            result = Mash(result, *ptr++);
+        }
+        size &= ((sizeof(uintptr_t) << 2) - 1);
+
+        safe_data = reinterpret_cast<const aliased_uint32_t*>(ptr);
+        const aliased_uint32_t* stop = safe_data + (size >> 2);
+        while (safe_data < stop) {
+            result = Mash(result, *safe_data++);
+        }
+
+        /*
+         *  smash us down to 32bits if we were 64. Note that when uintptr_t is
+         *  32bits, this code-path should go away, but I still got a warning
+         *  when I wrote
+         *      result ^= result >> 32;
+         *  since >>32 is undefined for 32bit ints, hence the wacky HALFBITS
+         *  define.
+         */
+        if (8 == sizeof(result)) {
+            result ^= result >> HALFBITS;
+        }
+        return static_cast<uint32_t>(result);
+    }
 };
 
 // SkGoodHash should usually be your first choice in hashing data.