From: Yann Collet <cyan@fb.com>
Date: Fri, 3 Mar 2017 01:09:21 +0000 (-0800)
Subject: disable prefetch-decode for 32-bits target
X-Git-Tag: v1.1.4~1^2~22
X-Git-Url: http://git.ipfire.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=fe5d27062e83f6f55fdee3c80f31f51fbf10f7bb;p=thirdparty%2Fzstd.git

disable prefetch-decode for 32-bits target

This decoder variant is detrimental to x86 architecture
likely due to register pressure.

Note that the variant is disabled for all 32-bits targets.
It's unclear if it would help for different architectures,
such as ARM, MIPS or PowerPC.
---

diff --git a/lib/decompress/zstd_decompress.c b/lib/decompress/zstd_decompress.c
index 4dfdf20fb..482c334ff 100644
--- a/lib/decompress/zstd_decompress.c
+++ b/lib/decompress/zstd_decompress.c
@@ -1410,13 +1410,18 @@ static size_t ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
 
     if (srcSize >= ZSTD_BLOCKSIZE_ABSOLUTEMAX) return ERROR(srcSize_wrong);
 
-    /* Decode literals sub-block */
+    /* Decode literals section */
     {   size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize);
         if (ZSTD_isError(litCSize)) return litCSize;
         ip += litCSize;
         srcSize -= litCSize;
     }
-    if (dctx->fParams.windowSize > (1<<23)) return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize);
+    if (sizeof(size_t) > 4)  /* do not enable prefetching on 32-bits x86, as it's performance detrimental */
+                             /* likely because of register pressure */
+                             /* if that's the correct cause, then 32-bits ARM should be affected differently */
+                             /* it would be good to test this on ARM real hardware, to see if prefetch version improves speed */
+        if (dctx->fParams.windowSize > (1<<23))
+            return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize);
     return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize);
 }