[llvm] [X86][CodeGen] Support long instruction fixup for APX NDD instructions (PR #83578)
Shengchen Kan via llvm-commits
llvm-commits at lists.llvm.org
Fri Mar 1 18:36:07 PST 2024
================
@@ -613,6 +613,87 @@ bool X86ExpandPseudo::expandMI(MachineBasicBlock &MBB,
case X86::CALL64m_RVMARKER:
expandCALL_RVMARKER(MBB, MBBI);
return true;
+ case X86::ADD32mi_ND:
+ case X86::ADD64mi32_ND:
+ case X86::SUB32mi_ND:
+ case X86::SUB64mi32_ND:
+ case X86::AND32mi_ND:
+ case X86::AND64mi32_ND:
+ case X86::OR32mi_ND:
+ case X86::OR64mi32_ND:
+ case X86::XOR32mi_ND:
+ case X86::XOR64mi32_ND:
+ case X86::ADC32mi_ND:
+ case X86::ADC64mi32_ND:
+ case X86::SBB32mi_ND:
+ case X86::SBB64mi32_ND: {
+ // It's possible for an EVEX-encoded legacy instruction to reach the 15-byte
+ // instruction length limit: 4 bytes of EVEX prefix + 1 byte of opcode + 1
+ // byte of ModRM + 1 byte of SIB + 4 bytes of displacement + 4 bytes of
+ // immediate = 15 bytes in total, e.g.
+ //
+ // addq $184, -96, %rax
+ //
+ // In such a case, no additional segment override prefix can be used. To
+ // resolve the issue, we split the “long” instruction into 2 instructions:
+ //
+ // subq $184, %fs:257(%rbx, %rcx), %rax
+ //
+ // ->
+ //
+ // movq %fs:257(%rbx, %rcx),%rax
+ // subq $184, %rax
+ int MemOpNo = X86::getFirstAddrOperandIdx(MI);
+ Register Segment = MI.getOperand(MemOpNo + X86::AddrSegmentReg).getReg();
+ if (Segment == X86::NoRegister)
----------------
KanRobert wrote:
We explained this in https://discourse.llvm.org/t/rfc-support-long-instruction-fixup-for-x86/76539
APX is supported only in 64-bit mode, where compiler only use 64-bit memory addressing (namely base and index register are 64-bit, no ADSIZE)
https://github.com/llvm/llvm-project/pull/83578
More information about the llvm-commits
mailing list