en Copyright 2015 Java http://172.16.0.5:8080/history/llvm-project-15.0.7/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll#4622afa9 AMDGPU: Convert AMDGPUResourceUsageAnalysis to a Module passThis is more precise in the face of indirect calls and aliases, stillassuming the call target is defined somewhere in the current module.This sometimes changes the order the functions are printed, and alsochanges the point where context errors are printed relative tostdout. This also likely has negative consequences for compile timeand memory usage. List of files: /llvm-project-15.0.7/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll Mon, 17 Jan 2022 15:55:36 +0000 Matt Arsenault <Matthew.Arsenault@amd.com> http://172.16.0.5:8080/history/llvm-project-15.0.7/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll#935abab6 AMDGPU: Use module level register maximums for unknown calleesCompute the theoretical register budget based on the IR functionsignature/attributes, and use the global maximum register budgets forunknown callees.This should fix the kernel reported register usage in the presence ofindirect calls. The previous fix in2b08f6af62afbf32e89a6a392dbafa92c62f7bdf was incorrect becauset it wasonly taking the maximum in the known call graph, and missing somethingthat was either outside of it or codegened later.This fixes a second case I discovered where calls to aliases also didnot work as expected. CallGraphAnalysis misses these, so functionscalled through aliases were not codegened ahead of callers asexpected. CallGraphAnalysis should probably be fixed to understandthis case, and there's likely a bug with IPRA here. This fixesnumerous failures in the conformance test at -O0. List of files: /llvm-project-15.0.7/llvm/test/CodeGen/AMDGPU/call-alias-register-usage3.ll Fri, 14 Jan 2022 01:13:29 +0000 Matt Arsenault <Matthew.Arsenault@amd.com>