[X86][mem-fold] Use the generated memory folding table

[skan]

cmp -> diff

[skan]

Add comments

[RKSimon]

llvm/lib/Target/X86/CMakeLists.txt

21

Looks like the rest of the table is sorted?

[skan]

llvm/lib/Target/X86/CMakeLists.txt

21

Good question, the tables are sorted by enum of the instruction. I added a comment in https://github.com/llvm/llvm-project/blob/main/llvm/utils/TableGen/X86FoldTablesEmitter.cpp#L112 to explain this.

[yubing]

LGTM

[skan]

Refine comments

[thakis]

Per https://reviews.llvm.org/D38028#1324522, maybe @craig.topper should look at this too?

llvm/lib/Target/X86/CMakeLists.txt

 tablegen(LLVM X86GenFastISel.inc -gen-fast-isel)
 tablegen(LLVM X86GenGlobalISel.inc -gen-global-isel)
 tablegen(LLVM X86GenInstrInfo.inc -gen-instr-info
                                   -instr-info-expand-mi-operand-info=0)
 tablegen(LLVM X86GenMnemonicTables.inc -gen-x86-mnemonic-tables -asmwriternum=1)
 tablegen(LLVM X86GenRegisterBank.inc -gen-register-bank)
 tablegen(LLVM X86GenRegisterInfo.inc -gen-register-info)
 tablegen(LLVM X86GenSubtargetInfo.inc -gen-subtarget)
-
-if (X86_GEN_FOLD_TABLES)
-  tablegen(LLVM X86GenFoldTables.inc -gen-x86-fold-tables -asmwriternum=1)
-endif()
 
 add_public_tablegen_target(X86CommonTableGen)
 
 set(sources
   X86ArgumentStackSlotRebase.cpp
   X86AsmPrinter.cpp
   X86AvoidTrailingCall.cpp
   X86CallFrameOptimization.cpp

llvm/lib/Target/X86/X86InstrFoldTables.cpp

 #include <vector>
 
 using namespace llvm;
 
 // These tables are sorted by their RegOp value allowing them to be binary
 // searched at runtime without the need for additional storage. The enum values
 // are currently emitted in X86GenInstrInfo.inc in alphabetical order. Which
 // makes sorting these tables a simple matter of alphabetizing the table.
-//
-// We also have a tablegen emitter that tries to autogenerate these tables
-// by comparing encoding information. This can be enabled by passing
-// X86_GEN_FOLD_TABLES=ON to cmake which fill produce X86GenFoldTables.inc
-// in the build area. There are currently some bugs in the autogenerated table
-// that require a manual review to copy them from the autogenerated table into
-// this table. It is unclear if we will ever be able to fully automate this
-// because as new instruction are added into holes in the X86 opcode map they
-// potentially pair up with old instructions and create new entries in the
-// tables that would be incorrect. The manual review process allows us a chance
-// to catch these before they become observable bugs.
-#include "X86MemFoldTables.inc"
+#include "X86GenFoldTables.inc"
 static const X86MemoryFoldTableEntry BroadcastFoldTable2[] = {
   { X86::VADDPDZ128rr,   X86::VADDPDZ128rmb,   TB_BCAST_SD },
   { X86::VADDPDZ256rr,   X86::VADDPDZ256rmb,   TB_BCAST_SD },
   { X86::VADDPDZrr,      X86::VADDPDZrmb,      TB_BCAST_SD },
   { X86::VADDPSZ128rr,   X86::VADDPSZ128rmb,   TB_BCAST_SS },
   { X86::VADDPSZ256rr,   X86::VADDPSZ256rmb,   TB_BCAST_SS },
   { X86::VADDPSZrr,      X86::VADDPSZrmb,      TB_BCAST_SS },
   { X86::VCMPPDZ128rri,  X86::VCMPPDZ128rmbi,  TB_BCAST_SD },

llvm/test/TableGen/x86-fold-tables.td

+// Steps to fix this test if it fails:
+//   1. Update X86ManualFoldTables.def to manually correct the entries in fold tables, or
+//      fix the vulnerable rules in X86FoldTablesEmitter.cpp until the diff is reasonable
+//   2. cp <generated_file> x86-fold-tables.inc
+
+// RUN: llvm-tblgen -gen-x86-fold-tables -asmwriternum=1 %p/../../lib/Target/X86/X86.td -I %p/../../lib/Target/X86 -I %p/../../include -o %t
+// RUN: diff %p/x86-fold-tables.inc %t

llvm/test/TableGen/x86-memory-folding.td

-// RUN: llvm-tblgen -gen-x86-fold-tables -asmwriternum=1 %p/../../lib/Target/X86/X86.td -I %p/../../lib/Target/X86 -I %p/../../include -o %t1
-// RUN: cmp %p/../../lib/Target/X86/X86MemFoldTables.inc %t1