[Libreoffice-commits] core.git: i18npool/source

[Noel Grandin (via logerrit)]

 i18npool/source/transliteration/transliteration_body.cxx |    7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

New commits:
commit 16d645e5b8f11b4ddb49a2b58bde388b28960abc
Author:     Noel Grandin <noel.grandin at collabora.co.uk>
AuthorDate: Thu Sep 16 10:38:04 2021 +0200
Commit:     Noel Grandin <noel.grandin at collabora.co.uk>
CommitDate: Thu Sep 16 12:09:23 2021 +0200

    speedup Transliteration_body::transliterateImpl
    
    use alloca in Transliteration_body::transliterateImpl to avoid
    over-allocating stack-space (which tends to unnecessarily flush some
    cache)
    
    Change-Id: I1843fdcb830a3e948a8bbd0a9c7eb143b21a804c
    Reviewed-on: https://gerrit.libreoffice.org/c/core/+/122184
    Tested-by: Jenkins
    Reviewed-by: Noel Grandin <noel.grandin at collabora.co.uk>

diff --git a/i18npool/source/transliteration/transliteration_body.cxx b/i18npool/source/transliteration/transliteration_body.cxx
index 9fd89df75e42..3581212af8b3 100644
--- a/i18npool/source/transliteration/transliteration_body.cxx
+++ b/i18npool/source/transliteration/transliteration_body.cxx
@@ -102,9 +102,12 @@ Transliteration_body::transliterateImpl(
     // Allocate the max possible buffer. Try to use stack instead of heap,
     // which would have to be reallocated most times anyways.
     constexpr sal_Int32 nLocalBuf = 2048;
-    sal_Unicode aLocalBuf[ nLocalBuf * NMAPPINGMAX ], *out = aLocalBuf;
+    sal_Unicode* out;
     std::unique_ptr<sal_Unicode[]> pHeapBuf;
-    if (nCount > nLocalBuf)
+    size_t nBytes = (nCount + 1) * sizeof(sal_Unicode);
+    if (nBytes <= nLocalBuf * NMAPPINGMAX)
+        out = static_cast<sal_Unicode*>(alloca(nBytes));
+    else
     {
         pHeapBuf.reset(new sal_Unicode[ nCount * NMAPPINGMAX ]);
         out = pHeapBuf.get();