this change saves a bunch of space but then more breaks in the wrong place:

[samth@huor:~/.../papers/rajan-papers/fast-dsl (master) plt] git diff diff --git a/fast-dsl/fastdsl.tex b/fast-dsl/fastdsl.tex index b61c3dd..f8e7cdd 100644 --- a/fast-dsl/fastdsl.tex +++ b/fast-dsl/fastdsl.tex @@ -256,7 +256,7 @@ difficult to express in a safe, high-level language. Faced with these limitations, DSL authors have several unpalatable choices. Either they limit themselves to the performance available in the host general-purpose language, they abandon smooth -interoperability and simply generate code in a low-level language like +interoperability and generate code in a low-level language like C, or they take on the substantial extra workload of designing a custom integration for their DSL. Each of these choices is sadly well-represented in modern DSL design---the first is too numerous to @@ -295,7 +295,7 @@ programs which generate LLVM IR. % However, the surrounding levels remain the same---code is still generated with -Racket macros, and the entire system still relies on the Racket +Racket macros, and the entire system relies on the Racket runtime. \begin{figure} @@ -413,7 +413,7 @@ Last in \cref{sec:hakaru}, our most substantial case study is a new backend for the Hakaru probabilistic programming system. Using Sham, we replaced an existing Haskell-generating backend and implemented new optimizations with substantial performance benefits. The new system is -up to $20\times$ faster than before. +up to $20\times$ faster. %% Acknowledgments

It doesn’t look like any space is needed? It already fits, no?