[PATCH] D41949: [RISCV] implement li pseudo instruction

Mario Werner via Phabricator via llvm-commits llvm-commits at lists.llvm.org
Mon Feb 26 02:32:26 PST 2018


niosHD added inline comments.


================
Comment at: lib/Target/RISCV/RISCVInstrInfo.td:401
+// PseudoLI is probably not the best idea anyway given that up to
+// 8 32-bit instructions are needed to generate an arbitrary 64-bit immediate.
+let hasSideEffects = 0, mayLoad = 0, mayStore = 0, Size = 8,
----------------
efriedma wrote:
> I'm not sure it's a good idea to make code generation use this pseudo-instruction; you'll miss optimization opportunities, like MachineCSE of lui instructions.
Indeed, me neither. I also raised this concern in one of our weekly sync up calls and the consensus was to go with the Pseudo instruction for now. However, I am definitely not opposed to expand the respective immediate loads early into machine instructions.


================
Comment at: test/MC/RISCV/rv64i-aliases-valid.s:94
+# CHECK-EXPAND: addi t4, t4, -272
+li t4, 0x123456789abcdef0
+# CHECK-EXPAND: addiw t5, zero, -1
----------------
efriedma wrote:
> This seems a little unfortunate... given you can load an arbitrary 32-bit immediate in two instructions, you should be able to load a 64-bit immediate in six instructions ("hi << 32 | lo").  But I guess that requires a second register?
Correct, with a second register, 6 instructions would be sufficient. Unfortunately, using a second register is, at least for the assembler, not an option. On the other hand, during codegen I think we should invest these two (virtual) registers. Additionally, in the long term, loading the constant from a constant pool should be evaluated given that it could be even more efficient. (assuming RV64I: 64-bit constant + 1 load + at most 2 instructions for the address calculation)


https://reviews.llvm.org/D41949





More information about the llvm-commits mailing list