GHC 9.14.1 · lts/ghc-9.14.x · 7a6c33e · 2026-10-12
Moduleghc-9.14.1GHC2021
Stack frame header size in bytes.
The stack frame header is made of the values that are always saved (regardless of the context.) It consists of the saved return address and a pointer to the previous frame. Thus, its size is two stack frame slots which equals two addresses/words (2 * 8 byte).
All registers are 8 byte wide.
The number of bytes that the stack pointer should be aligned to.
The number of spill slots available without allocating more.
Convert a spill slot number to a *byte* offset.
Get the registers that are being used by this instruction. regUsage doesn't need to do any trickery for jumps and such. Just state precisely the regs read and written by that insn. The consequences of control flow transfers, as far as register allocation goes, are taken care of by the register allocator.
Caller-saved registers (according to calling convention)
These registers may be clobbered after a jump.
Apply a given mapping to all the register references in this instruction.
Checks whether this instruction is a jump/branch instruction.
One that can change the flow of control in a way that the register allocator needs to worry about.
Get the BlockIds of the jump destinations (if any)
Change the destination of this (potential) jump instruction.
Used in the linear allocator when adding fixup blocks for join points.
mkSpillInstr :: HasCallStack=> NCGConfig-> RegWithFormatregister to spill
-> Intcurrent stack delta
-> Intspill slot to use
-> [Instr]Generate instructions to spill a register into a spill slot.
mkLoadInstr :: NCGConfig-> RegWithFormatregister to load
-> Intcurrent stack delta
-> Intspill slot to use
-> [Instr]Generate instructions to load a register from a spill slot.
See if this instruction is telling us the current C stack delta
Not real instructions. Just meta data
Copy the value in a register to another one.
Must work for all register classes.
Take the source and destination from this (potential) reg -> reg move instruction
We have to be a bit careful here: A MOV can also mean an implicit conversion. This case is filtered out.
Make an unconditional jump instruction.
Decrement sp to allocate stack space.
The stack grows downwards, so we decrement the stack pointer by n (bytes).
This is dual to mkStackDeallocInstr. sp is the RISCV stack pointer, not
to be confused with the STG stack pointer.
Increment SP to deallocate stack space.
The stack grows downwards, so we increment the stack pointer by n (bytes).
This is dual to mkStackAllocInstr. sp is the RISCV stack pointer, not to
be confused with the STG stack pointer.
COMMENT SDocComment pseudo-op
MULTILINE_COMMENT SDocMulti-line comment pseudo-op
ANN SDoc InstrLOCATION Int Int Int LexicalFastStringLocation pseudo-op .loc (file, line, col, name)
LDATA Section RawCmmStaticsStatic data spat out during code generation.
NEWBLOCK BlockIdStart a new basic block.
Useful during codegen, removed later. Preceding instruction should be a jump, as per the invariants for a BasicBlock (see Cmm).
DELTA IntSpecify current stack offset for benefit of subsequent passes
PUSH_STACK_FRAMEPush a minimal stack frame consisting of the return address (RA) and the frame pointer (FP).
POP_STACK_FRAMEPop the minimal stack frame of prior PUSH_STACK_FRAME.
ADD Operand Operand OperandArithmetic addition (both integer and floating point)
rd = rs1 + rs2SUB Operand Operand OperandArithmetic subtraction (both integer and floating point)
rd = rs1 - rs2AND Operand Operand OperandLogical AND (integer only)
rd = rs1 & rs2OR Operand Operand OperandLogical OR (integer only)
rd = rs1 | rs2SLL Operand Operand OperandLogical left shift (zero extened, integer only)
rd = rs1 << rs2SRL Operand Operand OperandLogical right shift (zero extened, integer only)
rd = rs1 >> rs2SRA Operand Operand OperandArithmetic right shift (sign-extened, integer only)
rd = rs1 >> rs2STR Format Operand OperandStore to memory (both, integer and floating point)
LDR Format Operand OperandLoad from memory (sign-extended, integer and floating point)
LDRU Format Operand OperandLoad from memory (unsigned, integer and floating point)
MUL Operand Operand OperandArithmetic multiplication (both, integer and floating point)
rd = rn × rmNEG Operand OperandNegation (both, integer and floating point)
rd = -op2DIV Operand Operand OperandDivision (both, integer and floating point)
rd = rn ÷ rmREM Operand Operand OperandRemainder (integer only, signed)
rd = rn % rmREMU Operand Operand OperandRemainder (integer only, unsigned)
rd = |rn % rm|MULH Operand Operand OperandHigh part of a multiplication that doesn't fit into 64bits (integer only)
E.g. for a multiplication with 64bits width: rd = (rs1 * rs2) >> 64.
DIVU Operand Operand OperandUnsigned division (integer only)
rd = |rn ÷ rm|XOR Operand Operand OperandXOR (integer only)
rd = rn ⊕ op2ORI Operand Operand OperandORI with immediate (integer only)
rd = rn | op2XORI Operand Operand OperandOR with immediate (integer only)
rd = rn ⊕ op2MOV Operand OperandMove to register (integer and floating point)
rd = rn or rd = #imm
CSET Operand Operand Operand CondPseudo-op for conditional setting of a register.
if(o2 cond o3) op <- 1 else op <- 0J TargetLike B, but only used for non-local jumps. Used to distinguish genJumps from others.
J_TBL [Maybe BlockId] (Maybe CLabel) RegA jump instruction with data for switch/jump tables
B TargetUnconditional jump (no linking)
BL Reg [Reg]Unconditional jump, links return address (sets ra/x1)
BCOND Cond Operand Operand Targetbranch with condition (integer only)
FENCE FenceType FenceTypeFence instruction
Memory barrier.
FCVT FcvtVariant Operand OperandFloating point conversion
FABS Operand OperandFloating point ABSolute value
FMIN Operand Operand OperandMin dest = min(r1)
FMAX Operand Operand OperandMax
FMA FMASign Operand Operand Operand OperandFloating-point fused multiply-add instructions
fmadd : d = r1 * r2 + r3
fnmsub: d = r1 * r2 - r3
fmsub : d = - r1 * r2 + r3
fnmadd: d = - r1 * r2 - r3
Instruction InstrDefined in ghc-9.14.1 · GHC.CmmToAsm.RV64 · orphanInstruction instance for RV64
Operand of a FENCE instruction (r, w or rw)
Variant of a floating point conversion instruction