Interprocedural MemorySSA Optimizations

lib/Optimization/IPO/ contains the interprocedural ShadowMem and MemorySSA optimizations used by the optimization tooling.

Location: lib/Optimization/IPO/

Library target: CanaryOptimizationIPO

Prerequisites

These passes expect IR that has already been prepared with SeaDsa-based ShadowMem instrumentation. The lotus-opt-ipo frontend inserts the required setup passes before running any IPO optimization:

RemovePtrToInt
AllocWrapInfo
DsaLibFuncInfo
AllocSiteInfo
DsaAnalysis
ShadowMem

Without that instrumentation, the IPO passes have no shadow.mem.* def-use chains to analyze.

Main passes

IPDeadStoreElimination (pass name ipdse) removes stores and some global initializers whose ShadowMem def-use chains never reach a shadow.mem.load. It traverses phis plus call/return edges encoded by shadow.mem.arg.* and shadow.mem.in/out.
IPRedundantLoadElimination (pass name ip-rle) removes repeated loads inside a basic block when both the pointer operand and the MemorySSA TLVar are identical and no intervening non-ShadowMem memory write occurs.
IPStoreSinking (pass name ip-sink) moves a store and its adjacent shadow.mem.store marker forward within a basic block to just before the earliest observable user when all crossed instructions are side-effect free.
IPStoreToLoadForwarding (pass name ip-forward) walks MemorySSA edges across phis and direct-call boundaries to replace a load with a unique reaching stored value.

Configuration

All four passes default to singleton-only regions and expose hidden tuning flags:

IPDeadStoreElimination: -ipdse-only-singleton, -ipdse-max-def-use
IPRedundantLoadElimination: -ip-rle-only-singleton
IPStoreSinking: -ip-sink-only-singleton
IPStoreToLoadForwarding: -ip-forward-only-singleton

These passes are exposed through the interprocedural optimization front-end.