2 % (c) The AQUA Project, Glasgow University, 1993-1998
6 module AsmCodeGen ( nativeCodeGen ) where
8 #include "HsVersions.h"
9 #include "nativeGen/NCG.h"
11 import List ( intersperse )
18 import AbsCStixGen ( genCodeAbstractC )
19 import AbsCSyn ( AbstractC, MagicId )
20 import AbsCUtils ( mkAbsCStmtList )
21 import AsmRegAlloc ( runRegAllocate )
22 import PrimOp ( commutableOp, PrimOp(..) )
23 import RegAllocInfo ( findReservedRegs )
24 import Stix ( StixTree(..), StixReg(..),
25 pprStixTrees, pprStixTree, CodeSegment(..),
26 stixCountTempUses, stixSubst,
27 NatM, initNat, mapNat,
28 NatM_State, mkNatM_State,
29 uniqOfNatM_State, deltaOfNatM_State )
30 import UniqSupply ( returnUs, thenUs, mapUs, initUs,
31 initUs_, UniqSM, UniqSupply,
32 lazyThenUs, lazyMapUs )
33 import MachMisc ( IF_ARCH_i386(i386_insert_ffrees,) )
35 import OrdList ( fromOL, concatOL )
40 The 96/03 native-code generator has machine-independent and
41 machine-dependent modules (those \tr{#include}'ing \tr{NCG.h}).
43 This module (@AsmCodeGen@) is the top-level machine-independent
44 module. It uses @AbsCStixGen.genCodeAbstractC@ to produce @StixTree@s
45 (defined in module @Stix@), using support code from @StixInfo@ (info
46 tables), @StixPrim@ (primitive operations), @StixMacro@ (Abstract C
47 macros), and @StixInteger@ (GMP arbitrary-precision operations).
49 Before entering machine-dependent land, we do some machine-independent
50 @genericOpt@imisations (defined below) on the @StixTree@s.
52 We convert to the machine-specific @Instr@ datatype with
53 @stmt2Instrs@, assuming an ``infinite'' supply of registers. We then
54 use a machine-independent register allocator (@runRegAllocate@) to
55 rejoin reality. Obviously, @runRegAllocate@ has machine-specific
56 helper functions (see about @RegAllocInfo@ below).
58 The machine-dependent bits break down as follows:
60 \item[@MachRegs@:] Everything about the target platform's machine
61 registers (and immediate operands, and addresses, which tend to
62 intermingle/interact with registers).
64 \item[@MachMisc@:] Includes the @Instr@ datatype (possibly should
65 have a module of its own), plus a miscellany of other things
66 (e.g., @targetDoubleSize@, @smStablePtrTable@, ...)
68 \item[@MachCode@:] @stmt2Instrs@ is where @Stix@ stuff turns into
71 \item[@PprMach@:] @pprInstr@ turns an @Instr@ into text (well, really
74 \item[@RegAllocInfo@:] In the register allocator, we manipulate
75 @MRegsState@s, which are @BitSet@s, one bit per machine register.
76 When we want to say something about a specific machine register
77 (e.g., ``it gets clobbered by this instruction''), we set/unset
78 its bit. Obviously, we do this @BitSet@ thing for efficiency
81 The @RegAllocInfo@ module collects together the machine-specific
82 info needed to do register allocation.
88 nativeCodeGen :: AbstractC -> UniqSupply -> (SDoc, SDoc)
90 = let absCstmts = mkAbsCStmtList absC
91 (sdoc_pairs, us1) = initUs us (lazyMapUs absCtoNat absCstmts)
92 stix_sdocs = map fst sdoc_pairs
93 insn_sdocs = map snd sdoc_pairs
95 insn_sdoc = my_vcat insn_sdocs
96 stix_sdoc = vcat stix_sdocs
99 my_trace m x = trace m x
100 my_vcat sds = vcat (intersperse (char ' '
101 $$ ptext SLIT("# ___ncg_debug_marker")
105 my_vcat sds = vcat sds
109 my_trace "nativeGen: begin"
110 (stix_sdoc, insn_sdoc)
113 absCtoNat :: AbstractC -> UniqSM (SDoc, SDoc)
115 = genCodeAbstractC absC `thenUs` \ stixRaw ->
116 genericOpt stixRaw `bind` \ stixOpt ->
117 genMachCode stixOpt `thenUs` \ pre_regalloc ->
118 regAlloc pre_regalloc `bind` \ almost_final ->
119 x86fp_kludge almost_final `bind` \ final_mach_code ->
120 vcat (map pprInstr final_mach_code) `bind` \ final_sdoc ->
121 pprStixTrees stixOpt `bind` \ stix_sdoc ->
122 returnUs (stix_sdoc, final_sdoc)
126 x86fp_kludge :: [Instr] -> [Instr]
127 x86fp_kludge = IF_ARCH_i386(i386_insert_ffrees,id)
129 regAlloc :: InstrBlock -> [Instr]
130 regAlloc = runRegAllocate allocatableRegs findReservedRegs
133 Top level code generator for a chunk of stix code. For this part of
134 the computation, we switch from the UniqSM monad to the NatM monad.
135 The latter carries not only a Unique, but also an Int denoting the
136 current C stack pointer offset in the generated code; this is needed
137 for creating correct spill offsets on architectures which don't offer,
138 or for which it would be prohibitively expensive to employ, a frame
139 pointer register. Viz, x86.
141 The offset is measured in bytes, and indicates the difference between
142 the current (simulated) C stack-ptr and the value it was at the
143 beginning of the block. For stacks which grow down, this value should
144 be either zero or negative.
146 Switching between the two monads whilst carrying along the same Unique
147 supply breaks abstraction. Is that bad?
150 genMachCode :: [StixTree] -> UniqSM InstrBlock
152 genMachCode stmts initial_us
153 = let initial_st = mkNatM_State initial_us 0
154 (blocks, final_st) = initNat initial_st
155 (mapNat stmt2Instrs stmts)
156 instr_list = concatOL blocks
157 final_us = uniqOfNatM_State final_st
158 final_delta = deltaOfNatM_State final_st
161 then (instr_list, final_us)
162 else pprPanic "genMachCode: nonzero final delta"
166 %************************************************************************
168 \subsection[NCOpt]{The Generic Optimiser}
170 %************************************************************************
172 This is called between translating Abstract C to its Tree and actually
173 using the Native Code Generator to generate the annotations. It's a
174 chance to do some strength reductions.
176 ** Remember these all have to be machine independent ***
178 Note that constant-folding should have already happened, but we might
179 have introduced some new opportunities for constant-folding wrt
180 address manipulations.
183 genericOpt :: [StixTree] -> [StixTree]
184 genericOpt = map stixConFold . stixPeep
188 stixPeep :: [StixTree] -> [StixTree]
190 -- This transformation assumes that the temp assigned to in t1
191 -- is not assigned to in t2; for otherwise the target of the
192 -- second assignment would be substituted for, giving nonsense
193 -- code. As far as I can see, StixTemps are only ever assigned
194 -- to once. It would be nice to be sure!
196 stixPeep ( t1@(StAssign pka (StReg (StixTemp u pk)) rhs)
199 | stixCountTempUses u t2 == 1
200 && sum (map (stixCountTempUses u) ts) == 0
203 trace ("nativeGen: inlining " ++ showSDoc (pprStixTree rhs))
205 (stixPeep (stixSubst u rhs t2 : ts))
207 stixPeep (t1:t2:ts) = t1 : stixPeep (t2:ts)
211 -- disable stix inlining until we figure out how to fix the
212 -- latent bugs in the register allocator which are exposed by
217 For most nodes, just optimize the children.
220 stixConFold :: StixTree -> StixTree
222 stixConFold (StInd pk addr) = StInd pk (stixConFold addr)
224 stixConFold (StAssign pk dst src)
225 = StAssign pk (stixConFold dst) (stixConFold src)
227 stixConFold (StJump addr) = StJump (stixConFold addr)
229 stixConFold (StCondJump addr test)
230 = StCondJump addr (stixConFold test)
232 stixConFold (StCall fn cconv pk args)
233 = StCall fn cconv pk (map stixConFold args)
236 Fold indices together when the types match:
238 stixConFold (StIndex pk (StIndex pk' base off) off')
240 = StIndex pk (stixConFold base)
241 (stixConFold (StPrim IntAddOp [off, off']))
243 stixConFold (StIndex pk base off)
244 = StIndex pk (stixConFold base) (stixConFold off)
247 For PrimOps, we first optimize the children, and then we try our hand
248 at some constant-folding.
251 stixConFold (StPrim op args) = stixPrimFold op (map stixConFold args)
254 Replace register leaves with appropriate StixTrees for the given
258 stixConFold leaf@(StReg (StixMagicId id))
259 = case (stgReg id) of
260 Always tree -> stixConFold tree
263 stixConFold other = other
266 Now, try to constant-fold the PrimOps. The arguments have already
267 been optimized and folded.
271 :: PrimOp -- The operation from an StPrim
272 -> [StixTree] -- The optimized arguments
275 stixPrimFold op arg@[StInt x]
277 IntNegOp -> StInt (-x)
280 stixPrimFold op args@[StInt x, StInt y]
282 CharGtOp -> StInt (if x > y then 1 else 0)
283 CharGeOp -> StInt (if x >= y then 1 else 0)
284 CharEqOp -> StInt (if x == y then 1 else 0)
285 CharNeOp -> StInt (if x /= y then 1 else 0)
286 CharLtOp -> StInt (if x < y then 1 else 0)
287 CharLeOp -> StInt (if x <= y then 1 else 0)
288 IntAddOp -> StInt (x + y)
289 IntSubOp -> StInt (x - y)
290 IntMulOp -> StInt (x * y)
291 IntQuotOp -> StInt (x `quot` y)
292 IntRemOp -> StInt (x `rem` y)
293 IntGtOp -> StInt (if x > y then 1 else 0)
294 IntGeOp -> StInt (if x >= y then 1 else 0)
295 IntEqOp -> StInt (if x == y then 1 else 0)
296 IntNeOp -> StInt (if x /= y then 1 else 0)
297 IntLtOp -> StInt (if x < y then 1 else 0)
298 IntLeOp -> StInt (if x <= y then 1 else 0)
299 -- ToDo: WordQuotOp, WordRemOp.
303 When possible, shift the constants to the right-hand side, so that we
304 can match for strength reductions. Note that the code generator will
305 also assume that constants have been shifted to the right when
309 stixPrimFold op [x@(StInt _), y] | commutableOp op = stixPrimFold op [y, x]
312 We can often do something with constants of 0 and 1 ...
315 stixPrimFold op args@[x, y@(StInt 0)]
328 IntNeOp | is_comparison -> x
333 StPrim opp [_, _] -> opp `elem` comparison_ops
336 stixPrimFold op args@[x, y@(StInt 1)]
344 Now look for multiplication/division by powers of 2 (integers).
347 stixPrimFold op args@[x, y@(StInt n)]
349 IntMulOp -> case exactLog2 n of
350 Nothing -> StPrim op args
351 Just p -> StPrim ISllOp [x, StInt p]
352 IntQuotOp -> case exactLog2 n of
353 Nothing -> StPrim op args
354 Just p -> StPrim ISrlOp [x, StInt p]
358 Anything else is just too hard.
361 stixPrimFold op args = StPrim op args
366 = [ CharGtOp , CharGeOp , CharEqOp , CharNeOp , CharLtOp , CharLeOp,
367 IntGtOp , IntGeOp , IntEqOp , IntNeOp , IntLtOp , IntLeOp,
368 WordGtOp , WordGeOp , WordEqOp , WordNeOp , WordLtOp , WordLeOp,
369 AddrGtOp , AddrGeOp , AddrEqOp , AddrNeOp , AddrLtOp , AddrLeOp,
370 FloatGtOp , FloatGeOp , FloatEqOp , FloatNeOp , FloatLtOp , FloatLeOp,
371 DoubleGtOp, DoubleGeOp, DoubleEqOp, DoubleNeOp, DoubleLtOp, DoubleLeOp