Please sign in to comment.
IMPALA-4705, IMPALA-4779, IMPALA-4780: Fix some Expr bugs with codegen
This change fixes expr-test.cc to work with codegen as it's originally intended. Fixing it uncovers a couple of bugs fixed in this patch: IMPALA-4705: When an IR function is materialized, its function body is parsed to find all its callee functions to be materialized too. However, the old code doesn't detect callee fnctions referenced indirectly (e.g. a callee function passed as argument to another function). This change fixes the problem above inspecting the use lists of llvm::Function objects. When parsing the bitcode module into memory, LLVM already establishes a use list for each llvm::Value object which llvm::Function is a subclass of. A use list contains all the locations in the module in which the Value is referenced. For a llvm::Function object, that would be its call sites and constant expressions referencing the functions. By using the use lists of llvm::Function in the module, a global map is established at Impala initialization time to map functions to their corresponding callee functions. This map is then used when materializing a function to ensure all its callee functions are also materialized recursively. IMPALA-4779: conditional function isfalse(), istrue(), isnotfalse(), isnotrue() aren't cross-compiled so they will lead to unexpected query failure when codegen is enabled. This change will cross-compile these functions. IMPALA-4780: next_day() always returns NULL when codegen is enabled. The bound checks for next_day() use some class static variables initialized in the global constructors (@llvm.global_ctors). However, we never execute the global constructors before calling the JIT compiled functions. This causes these variables to remain as zero, causing all executions of next_day() to fail the bound checks. The reason why these class static variables aren't compiled as global constants in LLVM IR is that TimestampFunctions::MIN_YEAR is not a compile time constant. This change fixes the problem above by setting TimestampFunctions::MIN_YEAR to a known constant value. A DCHECK is added to verify that it matches the value defined in the boost library. Change-Id: I40fdb035a565ae2f9c9fbf4db48a548653ef7608 Reviewed-on: http://gerrit.cloudera.org:8080/5732 Reviewed-by: Michael Ho <email@example.com> Tested-by: Impala Public Jenkins
- Loading branch information...
Showing with 225 additions and 234 deletions.
- +2 −1 be/src/codegen/llvm-codegen-test.cc
- +75 −113 be/src/codegen/llvm-codegen.cc
- +34 −32 be/src/codegen/llvm-codegen.h
- +20 −0 be/src/exprs/conditional-functions-ir.cc
- +0 −19 be/src/exprs/conditional-functions.cc
- +7 −2 be/src/exprs/expr-codegen-test.cc
- +6 −7 be/src/exprs/expr-test.cc
- +3 −42 be/src/exprs/timestamp-functions-ir.cc
- +16 −0 be/src/exprs/timestamp-functions.cc
- +22 −13 be/src/exprs/timestamp-functions.h
- +7 −3 be/src/service/fe-support.cc
- +1 −1 be/src/service/fe-support.h
- +1 −1 be/src/service/impalad-main.cc
- +20 −0 be/src/testutil/test-udfs.cc
- +7 −0 testdata/workloads/functional-query/queries/QueryTest/udf.test
- +4 −0 tests/query_test/test_udfs.py
Oops, something went wrong.