1919#include " llvm/ADT/SetVector.h"
2020#include " llvm/ADT/Triple.h"
2121#include " llvm/Bitcode/BitcodeWriterPass.h"
22+ #include " llvm/GenXIntrinsics/GenXSPIRVWriterAdaptor.h"
2223#include " llvm/IR/IRPrintingPasses.h"
2324#include " llvm/IR/InstIterator.h"
2425#include " llvm/IR/Instructions.h"
2526#include " llvm/IR/LLVMContext.h"
2627#include " llvm/IR/LegacyPassManager.h"
2728#include " llvm/IR/Module.h"
2829#include " llvm/IRReader/IRReader.h"
30+ #include " llvm/SYCLLowerIR/LowerESIMD.h"
2931#include " llvm/Support/CommandLine.h"
3032#include " llvm/Support/InitLLVM.h"
3133#include " llvm/Support/Path.h"
3537#include " llvm/Support/WithColor.h"
3638#include " llvm/Transforms/IPO.h"
3739#include " llvm/Transforms/IPO/GlobalDCE.h"
40+ #include " llvm/Transforms/InstCombine/InstCombine.h"
41+ #include " llvm/Transforms/Scalar.h"
3842#include " llvm/Transforms/Utils/Cloning.h"
3943
4044#include < memory>
@@ -92,6 +96,45 @@ static cl::opt<bool> SplitEsimd{"split-esimd",
9296 cl::desc (" Split SYCL and ESIMD kernels"
9397 cl::cat (PostLinkCat)};
9498
99+ // TODO Design note: sycl-post-link should probably separate different kinds of
100+ // its functionality on logical and source level:
101+ // - LLVM IR module splitting
102+ // - Running LLVM IR passes on resulting modules
103+ // - Generating additional files (like spec constants, dead arg info,...)
104+ // The tool itself could be just a "driver" creating needed pipelines from the
105+ // above actions. This could help make the tool structure clearer and more
106+ // maintainable.
107+
108+ static cl::opt<bool > LowerEsimd{
109+ " lower-esimd" cl::desc (" Lower ESIMD constructs" cl::cat (PostLinkCat)};
110+
111+ static cl::opt<bool >
112+ OptLevelO0 (" O0" " Optimization level 0. Similar to clang -O0"
113+ cl::cat(PostLinkCat));
114+
115+ static cl::opt<bool >
116+ OptLevelO1 (" O1" " Optimization level 1. Similar to clang -O1"
117+ cl::cat(PostLinkCat));
118+
119+ static cl::opt<bool >
120+ OptLevelO2 (" O2" " Optimization level 2. Similar to clang -O2"
121+ cl::cat(PostLinkCat));
122+
123+ static cl::opt<bool > OptLevelOs (
124+ " Os"
125+ cl::desc (
126+ " Like -O2 with extra optimizations for size. Similar to clang -Os"
127+ cl::cat(PostLinkCat));
128+
129+ static cl::opt<bool > OptLevelOz (
130+ " Oz"
131+ cl::desc (" Like -Os but reduces code size further. Similar to clang -Oz"
132+ cl::cat(PostLinkCat));
133+
134+ static cl::opt<bool >
135+ OptLevelO3 (" O3" " Optimization level 3. Similar to clang -O3"
136+ cl::cat(PostLinkCat));
137+
95138enum IRSplitMode {
96139 SPLIT_PER_TU, // one module per translation unit
97140 SPLIT_PER_KERNEL, // one module per kernel
@@ -625,6 +668,59 @@ static string_vector saveResultSymbolsLists(string_vector &ResSymbolsLists,
625668 } \
626669 }
627670
671+ // Helper function for creating Inliner pass.
672+ // The approach is taken from opt tool.
673+ static Pass *createFunctionInliningPassHelper () {
674+ if (OptLevelO0)
675+ return createFunctionInliningPass (0 , 0 , false );
676+
677+ if (OptLevelO1)
678+ return createFunctionInliningPass (1 , 0 , false );
679+
680+ if (OptLevelO2)
681+ return createFunctionInliningPass (2 , 0 , false );
682+
683+ if (OptLevelOs)
684+ return createFunctionInliningPass (2 , 1 , false );
685+
686+ if (OptLevelOz)
687+ return createFunctionInliningPass (2 , 2 , false );
688+
689+ if (OptLevelO3)
690+ return createFunctionInliningPass (3 , 0 , false );
691+
692+ return createFunctionInliningPass ();
693+ }
694+
695+ // When ESIMD code was separated from the regular SYCL code,
696+ // we can safely process ESIMD part.
697+ // TODO: support options like -debug-pass, -print-[before|after], and others
698+ static void LowerEsimdConstructs (Module &M) {
699+ legacy::PassManager MPM;
700+ MPM.add (createSYCLLowerESIMDPass ());
701+ if (!OptLevelO0) {
702+ // Inlining and SROA passes are required to make
703+ // ESIMD/accessor_gather_scatter.cpp test work.
704+ MPM.add (createFunctionInliningPassHelper ());
705+ MPM.add (createSROAPass ());
706+ }
707+ MPM.add (createESIMDLowerVecArgPass ());
708+ MPM.add (createESIMDLowerLoadStorePass ());
709+ if (!OptLevelO0) {
710+ MPM.add (createSROAPass ());
711+ MPM.add (createEarlyCSEPass (true ));
712+ MPM.add (createInstructionCombiningPass ());
713+ MPM.add (createDeadCodeEliminationPass ());
714+ MPM.add (createFunctionInliningPassHelper ());
715+ MPM.add (createSROAPass ());
716+ MPM.add (createEarlyCSEPass (true ));
717+ MPM.add (createInstructionCombiningPass ());
718+ MPM.add (createDeadCodeEliminationPass ());
719+ }
720+ MPM.add (createGenXSPIRVWriterAdaptorPass ());
721+ MPM.run (M);
722+ }
723+
628724using TableFiles = std::map<StringRef, string_vector>;
629725
630726static TableFiles processOneModule (std::unique_ptr<Module> M, bool IsEsimd,
@@ -633,6 +729,9 @@ static TableFiles processOneModule(std::unique_ptr<Module> M, bool IsEsimd,
633729 if (!M)
634730 return TblFiles;
635731
732+ if (IsEsimd && LowerEsimd)
733+ LowerEsimdConstructs (*M);
734+
636735 std::map<StringRef, std::vector<Function *>> GlobalsSet;
637736
638737 bool DoSplit = SplitMode.getNumOccurrences () > 0 ;
@@ -685,11 +784,16 @@ static TableFiles processOneModule(std::unique_ptr<Module> M, bool IsEsimd,
685784 ResultModules.push_back (std::move (M));
686785
687786 {
688- // reuse input module if there were no spec constants and no splitting
787+ // Reuse input module with only regular SYCL kernels if there were
788+ // no spec constants and no splitting.
789+ // We cannot reuse input module for ESIMD code since it was transformed.
790+ bool CanReuseInputModule = !SpecConstsMet && (ResultModules.size () == 1 ) &&
791+ !SyclAndEsimdKernels && !IsEsimd;
689792 string_vector Files =
690- SpecConstsMet || (ResultModules.size () > 1 ) || SyclAndEsimdKernels
691- ? saveResultModules (ResultModules, IsEsimd ? " esimd_" " "
692- : string_vector{InputFilename};
793+ CanReuseInputModule
794+ ? string_vector{InputFilename}
795+ : saveResultModules (ResultModules, IsEsimd ? " esimd_" " "
796+
693797 // "Code" column is always output
694798 std::copy (Files.begin (), Files.end (),
695799 std::back_inserter (TblFiles[COL_CODE]));
@@ -816,6 +920,10 @@ int main(int argc, char **argv) {
816920 " - Specialization constant intrinsic transformer. Replaces symbolic\n "
817921 " ID-based intrinsics to integer ID-based ones to make them friendly\n "
818922 " for the SPIRV translator\n "
923+ " When the tool splits input module into regular SYCL and ESIMD kernels,\n "
924+ " it performs a set of specific lowering and transformation passes on\n "
925+ " ESIMD module, which is enabled by the '-lower-esimd' option. Regular\n "
926+ " optimization level options are supported, e.g. -O[0|1|2|3|s|z].\n "
819927 " Normally, the tool generates a number of files and \" file table\"\n "
820928 " file listing all generated files in a table manner. For example, if\n "
821929 " the input file 'example.bc' contains two kernels, then the command\n "
0 commit comments