1 -----------------------------------------------------------------------------
3 -- Machine-dependent assembly language
5 -- (c) The University of Glasgow 1993-2004
7 -----------------------------------------------------------------------------
9 #include "HsVersions.h"
10 #include "nativeGen/NCG.h"
21 = ALWAYS -- What's really used? ToDo
41 -- -----------------------------------------------------------------------------
42 -- Intel x86 instructions
45 Intel, in their infinite wisdom, selected a stack model for floating
46 point registers on x86. That might have made sense back in 1979 --
47 nowadays we can see it for the nonsense it really is. A stack model
48 fits poorly with the existing nativeGen infrastructure, which assumes
49 flat integer and FP register sets. Prior to this commit, nativeGen
50 could not generate correct x86 FP code -- to do so would have meant
51 somehow working the register-stack paradigm into the register
52 allocator and spiller, which sounds very difficult.
54 We have decided to cheat, and go for a simple fix which requires no
55 infrastructure modifications, at the expense of generating ropey but
56 correct FP code. All notions of the x86 FP stack and its insns have
57 been removed. Instead, we pretend (to the instruction selector and
58 register allocator) that x86 has six floating point registers, %fake0
59 .. %fake5, which can be used in the usual flat manner. We further
60 claim that x86 has floating point instructions very similar to SPARC
61 and Alpha, that is, a simple 3-operand register-register arrangement.
62 Code generation and register allocation proceed on this basis.
64 When we come to print out the final assembly, our convenient fiction
65 is converted to dismal reality. Each fake instruction is
66 independently converted to a series of real x86 instructions.
67 %fake0 .. %fake5 are mapped to %st(0) .. %st(5). To do reg-reg
68 arithmetic operations, the two operands are pushed onto the top of the
69 FP stack, the operation done, and the result copied back into the
70 relevant register. There are only six %fake registers because 2 are
71 needed for the translation, and x86 has 8 in total.
73 The translation is inefficient but is simple and it works. A cleverer
74 translation would handle a sequence of insns, simulating the FP stack
75 contents, would not impose a fixed mapping from %fake to %st regs, and
76 hopefully could avoid most of the redundant reg-reg moves of the
79 We might as well make use of whatever unique FP facilities Intel have
80 chosen to bless us with (let's not be churlish, after all).
81 Hence GLDZ and GLD1. Bwahahahahahahaha!
85 MORE FLOATING POINT MUSINGS...
87 Intel's internal floating point registers are by default 80 bit
88 extended precision. This means that all operations done on values in
89 registers are done at 80 bits, and unless the intermediate values are
90 truncated to the appropriate size (32 or 64 bits) by storing in
91 memory, calculations in registers will give different results from
92 calculations which pass intermediate values in memory (eg. via
95 One solution is to set the FPU into 64 bit precision mode. Some OSs
96 do this (eg. FreeBSD) and some don't (eg. Linux). The problem here is
97 that this will only affect 64-bit precision arithmetic; 32-bit
98 calculations will still be done at 64-bit precision in registers. So
99 it doesn't solve the whole problem.
101 There's also the issue of what the C library is expecting in terms of
102 precision. It seems to be the case that glibc on Linux expects the
103 FPU to be set to 80 bit precision, so setting it to 64 bit could have
104 unexpected effects. Changing the default could have undesirable
105 effects on other 3rd-party library code too, so the right thing would
106 be to save/restore the FPU control word across Haskell code if we were
109 gcc's -ffloat-store gives consistent results by always storing the
110 results of floating-point calculations in memory, which works for both
111 32 and 64-bit precision. However, it only affects the values of
112 user-declared floating point variables in C, not intermediate results.
113 GHC in -fvia-C mode uses -ffloat-store (see the -fexcess-precision
116 Another problem is how to spill floating point registers in the
117 register allocator. Should we spill the whole 80 bits, or just 64?
118 On an OS which is set to 64 bit precision, spilling 64 is fine. On
119 Linux, spilling 64 bits will round the results of some operations.
120 This is what gcc does. Spilling at 80 bits requires taking up a full
121 128 bit slot (so we get alignment). We spill at 80-bits and ignore
122 the alignment problems.
124 In the future, we'll use the SSE registers for floating point. This
125 requires a CPU that supports SSE2 (ordinary SSE only supports 32 bit
126 precision float ops), which means P4 or Xeon and above. Using SSE
127 will solve all these problems, because the SSE registers use fixed 32
128 bit or 64 bit precision.
138 -- some static data spat out during code
139 -- generation. Will be extracted before
141 | LDATA Section [CmmStatic]
143 -- start a new basic block. Useful during
144 -- codegen, removed later. Preceding
145 -- instruction should be a jump, as per the
146 -- invariants for a BasicBlock (see Cmm).
149 -- specify current stack offset for
150 -- benefit of subsequent passes
153 -- | spill this reg to a stack slot
156 -- | reload this reg from a stack slot
161 | MOV Size Operand Operand
162 | MOVZxL Size Operand Operand -- size is the size of operand 1
163 | MOVSxL Size Operand Operand -- size is the size of operand 1
164 -- x86_64 note: plain mov into a 32-bit register always zero-extends
165 -- into the 64-bit reg, in contrast to the 8 and 16-bit movs which
166 -- don't affect the high bits of the register.
168 -- Load effective address (also a very useful three-operand add instruction :-)
169 | LEA Size Operand Operand
172 | ADD Size Operand Operand
173 | ADC Size Operand Operand
174 | SUB Size Operand Operand
176 | MUL Size Operand Operand
177 | IMUL Size Operand Operand -- signed int mul
178 | IMUL2 Size Operand -- %edx:%eax = operand * %eax
180 | DIV Size Operand -- eax := eax:edx/op, edx := eax:edx%op
181 | IDIV Size Operand -- ditto, but signed
183 -- Simple bit-twiddling.
184 | AND Size Operand Operand
185 | OR Size Operand Operand
186 | XOR Size Operand Operand
188 | NEGI Size Operand -- NEG instruction (name clash with Cond)
190 -- Shifts (amount may be immediate or %cl only)
191 | SHL Size Operand{-amount-} Operand
192 | SAR Size Operand{-amount-} Operand
193 | SHR Size Operand{-amount-} Operand
195 | BT Size Imm Operand
201 -- Note that we cheat by treating G{ABS,MOV,NEG} of doubles
202 -- as single instructions right up until we spit them out.
203 -- all the 3-operand fake fp insns are src1 src2 dst
204 -- and furthermore are constrained to be fp regs only.
205 -- IMPORTANT: keep is_G_insn up to date with any changes here
206 | GMOV Reg Reg -- src(fpreg), dst(fpreg)
207 | GLD Size AddrMode Reg -- src, dst(fpreg)
208 | GST Size Reg AddrMode -- src(fpreg), dst
210 | GLDZ Reg -- dst(fpreg)
211 | GLD1 Reg -- dst(fpreg)
213 | GFTOI Reg Reg -- src(fpreg), dst(intreg)
214 | GDTOI Reg Reg -- src(fpreg), dst(intreg)
216 | GITOF Reg Reg -- src(intreg), dst(fpreg)
217 | GITOD Reg Reg -- src(intreg), dst(fpreg)
219 | GADD Size Reg Reg Reg -- src1, src2, dst
220 | GDIV Size Reg Reg Reg -- src1, src2, dst
221 | GSUB Size Reg Reg Reg -- src1, src2, dst
222 | GMUL Size Reg Reg Reg -- src1, src2, dst
224 -- FP compare. Cond must be `elem` [EQQ, NE, LE, LTT, GE, GTT]
225 -- Compare src1 with src2; set the Zero flag iff the numbers are
226 -- comparable and the comparison is True. Subsequent code must
227 -- test the %eflags zero flag regardless of the supplied Cond.
228 | GCMP Cond Reg Reg -- src1, src2
230 | GABS Size Reg Reg -- src, dst
231 | GNEG Size Reg Reg -- src, dst
232 | GSQRT Size Reg Reg -- src, dst
233 | GSIN Size CLabel CLabel Reg Reg -- src, dst
234 | GCOS Size CLabel CLabel Reg Reg -- src, dst
235 | GTAN Size CLabel CLabel Reg Reg -- src, dst
237 | GFREE -- do ffree on all x86 regs; an ugly hack
240 #if x86_64_TARGET_ARCH
241 -- SSE2 floating point: we use a restricted set of the available SSE2
242 -- instructions for floating-point.
244 -- use MOV for moving (either movss or movsd (movlpd better?))
246 | CVTSS2SD Reg Reg -- F32 to F64
247 | CVTSD2SS Reg Reg -- F64 to F32
248 | CVTTSS2SIQ Operand Reg -- F32 to I32/I64 (with truncation)
249 | CVTTSD2SIQ Operand Reg -- F64 to I32/I64 (with truncation)
250 | CVTSI2SS Operand Reg -- I32/I64 to F32
251 | CVTSI2SD Operand Reg -- I32/I64 to F64
253 -- use ADD & SUB for arithmetic. In both cases, operands
256 -- SSE2 floating-point division:
257 | FDIV Size Operand Operand -- divisor, dividend(dst)
259 -- use CMP for comparisons. ucomiss and ucomisd instructions
260 -- compare single/double prec floating point respectively.
262 | SQRT Size Operand Reg -- src, dst
266 | TEST Size Operand Operand
267 | CMP Size Operand Operand
273 -- both unused (SDM):
279 | JXX Cond BlockId -- includes unconditional branches
280 | JXX_GBL Cond Imm -- non-local version of JXX
281 | JMP_TBL Operand [BlockId] -- table jump
282 | CALL (Either Imm Reg) [Reg]
285 | CLTD Size -- sign extend %eax into %edx:%eax
287 | FETCHGOT Reg -- pseudo-insn for ELF position-independent code
291 -- addl __GLOBAL_OFFSET_TABLE__+.-1b, %reg
292 | FETCHPC Reg -- pseudo-insn for Darwin position-independent code
299 = OpReg Reg -- register
300 | OpImm Imm -- immediate value
301 | OpAddr AddrMode -- memory reference
305 i386_insert_ffrees :: [GenBasicBlock Instr] -> [GenBasicBlock Instr]
306 i386_insert_ffrees blocks
307 | or (map (any is_G_instr) [ instrs | BasicBlock id instrs <- blocks ])
308 = map ffree_before_nonlocal_transfers blocks
312 ffree_before_nonlocal_transfers (BasicBlock id insns)
313 = BasicBlock id (foldr p [] insns)
314 where p insn r = case insn of
315 CALL _ _ -> GFREE : insn : r
316 JMP _ -> GFREE : insn : r
319 -- if you ever add a new FP insn to the fake x86 FP insn set,
320 -- you must update this too
321 is_G_instr :: Instr -> Bool
324 GMOV _ _ -> True; GLD _ _ _ -> True; GST _ _ _ -> True
325 GLDZ _ -> True; GLD1 _ -> True
326 GFTOI _ _ -> True; GDTOI _ _ -> True
327 GITOF _ _ -> True; GITOD _ _ -> True
328 GADD _ _ _ _ -> True; GDIV _ _ _ _ -> True
329 GSUB _ _ _ _ -> True; GMUL _ _ _ _ -> True
330 GCMP _ _ _ -> True; GABS _ _ _ -> True
331 GNEG _ _ _ -> True; GSQRT _ _ _ -> True
332 GSIN _ _ _ _ _ -> True; GCOS _ _ _ _ _ -> True; GTAN _ _ _ _ _ -> True
333 GFREE -> panic "is_G_instr: GFREE (!)"
335 #endif /* i386_TARGET_ARCH */