From 108d19e6b57d598e9addf233870a43711d59e035 Mon Sep 17 00:00:00 2001 From: Chandler Carruth Date: Fri, 14 Dec 2012 13:37:17 +0000 Subject: Add a rough draft of some content about the new SROA. I'll try to proof read this and clean it up tomorrow, but hopefully it's a good placeholder. llvm-svn: 170196 --- llvm/docs/ReleaseNotes.rst | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-) (limited to 'llvm/docs/ReleaseNotes.rst') diff --git a/llvm/docs/ReleaseNotes.rst b/llvm/docs/ReleaseNotes.rst index 486074914b8..31c567b33ac 100644 --- a/llvm/docs/ReleaseNotes.rst +++ b/llvm/docs/ReleaseNotes.rst @@ -351,7 +351,21 @@ We vectorize under the following loops: '``noalias``' and are checked at runtime. #. ... -SROA - We've re-written SROA to be significantly more powerful. +SROA - We've re-written SROA to be significantly more powerful and generate +code which is much more friendly to the rest of the optimization pipeline. +Previously this pass had scaling problems that required it to only operate on +relatively small aggregates, and at times it would mistakenly replace a large +aggregate with a single very large integer in order to make it a scalar SSA +value. The result was a large number of i1024 and i2048 values representing any +small stack buffer. These in turn slowed down many subsequent optimization +paths. + +The new SROA pass uses a different algorithm that allows it to only promote to +scalars the pieces of the aggregate actively in use. Because of this it doesn't +require any thresholds. It also always deduces the scalar values from the uses +of the aggregate rather than the specific LLVM type of the aggregate. These +features combine to both optimize more code with the pass but to improve the +compile time of many functions dramatically. #. Branch weight metadata is preseved through more of the optimizer. #. ... -- cgit v1.2.3