+#elif sparc_TARGET_ARCH
+
+ = let -- At least for sparc V8
+ bytes_per_word = 4
+
+ -- speaks for itself
+ w32_to_w8s_bigEndian :: Word32 -> [Word8]
+ w32_to_w8s_bigEndian w
+ = [fromIntegral (0xFF .&. (w `shiftR` 24)),
+ fromIntegral (0xFF .&. (w `shiftR` 16)),
+ fromIntegral (0xFF .&. (w `shiftR` 8)),
+ fromIntegral (0xFF .&. w)]
+
+ offsets_to_pushW
+ = concat
+ [ [a_offW .. a_offW + cgRepSizeW a_rep - 1]
+
+ | (a_offW, a_rep) <- arg_offs_n_reps
+ ]
+
+ total_argWs = length offsets_to_pushW
+ argWs_on_stack = if total_argWs > 6 then total_argWs - 6
+ else 0
+
+ -- The stack pointer must be kept 8-byte aligned, which means
+ -- we need to calculate this quantity too
+ argWs_on_stack_ROUNDED_UP
+ | odd argWs_on_stack = 1 + argWs_on_stack
+ | otherwise = argWs_on_stack
+
+ -- some helpers to assemble sparc insns.
+ -- REGS
+ iReg, oReg, gReg, fReg :: Int -> Word32
+ iReg = fromIntegral . (+ 24)
+ oReg = fromIntegral . (+ 8)
+ gReg = fromIntegral . (+ 0)
+ fReg = fromIntegral
+
+ sp = oReg 6
+ i0 = iReg 0
+ i7 = iReg 7
+ o0 = oReg 0
+ o1 = oReg 1
+ o7 = oReg 7
+ g0 = gReg 0
+ g1 = gReg 1
+ f0 = fReg 0
+ f1 = fReg 1
+
+ -- INSN templates
+ insn_r_r_i :: Word32 -> Word32 -> Word32 -> Int -> Word32
+ insn_r_r_i op3 rs1 rd imm13
+ = (3 `shiftL` 30)
+ .|. (rs1 `shiftL` 25)
+ .|. (op3 `shiftL` 19)
+ .|. (rd `shiftL` 14)
+ .|. (1 `shiftL` 13)
+ .|. mkSimm13 imm13
+
+ insn_r_i_r :: Word32 -> Word32 -> Int -> Word32 -> Word32
+ insn_r_i_r op3 rs1 imm13 rd
+ = (2 `shiftL` 30)
+ .|. (rd `shiftL` 25)
+ .|. (op3 `shiftL` 19)
+ .|. (rs1 `shiftL` 14)
+ .|. (1 `shiftL` 13)
+ .|. mkSimm13 imm13
+
+ mkSimm13 :: Int -> Word32
+ mkSimm13 imm13
+ = let imm13w = (fromIntegral imm13) :: Word32
+ in imm13w .&. 0x1FFF
+
+ -- REAL (non-synthetic) insns
+ -- or %rs1, %rs2, %rd
+ mkOR :: Word32 -> Word32 -> Word32 -> Word32
+ mkOR rs1 rs2 rd
+ = (2 `shiftL` 30)
+ .|. (rd `shiftL` 25)
+ .|. (op3_OR `shiftL` 19)
+ .|. (rs1 `shiftL` 14)
+ .|. (0 `shiftL` 13)
+ .|. rs2
+ where op3_OR = 2 :: Word32
+
+ -- ld(int) [%rs + imm13], %rd
+ mkLD rs1 imm13 rd = insn_r_r_i 0x00{-op3_LD-} rd rs1 imm13
+
+ -- st(int) %rs, [%rd + imm13]
+ mkST = insn_r_r_i 0x04 -- op3_ST
+
+ -- st(float) %rs, [%rd + imm13]
+ mkSTF = insn_r_r_i 0x24 -- op3_STF
+
+ -- jmpl %rs + imm13, %rd
+ mkJMPL = insn_r_i_r 0x38 -- op3_JMPL
+
+ -- save %rs + imm13, %rd
+ mkSAVE = insn_r_i_r 0x3C -- op3_SAVE
+
+ -- restore %rs + imm13, %rd
+ mkRESTORE = insn_r_i_r 0x3D -- op3_RESTORE
+
+ -- SYNTHETIC insns
+ mkNOP = mkOR g0 g0 g0
+ mkCALL reg = mkJMPL reg 0 o7
+ mkRET = mkJMPL i7 8 g0
+ mkRESTORE_TRIVIAL = mkRESTORE g0 0 g0
+ in
+ --trace (show (map fst arg_offs_n_reps))
+ concatMap w32_to_w8s_bigEndian (
+
+ {- On entry, %o0 is the arg passed from the interpreter. After
+ the initial save insn, it will be in %i0. Studying the sparc
+ docs one would have thought that the minimum frame size is 92
+ bytes, but gcc always uses at least 112, and indeed there are
+ segfaults a-plenty with 92. So I use 112 here as well. I
+ don't understand why, tho.
+ -}
+ [mkSAVE sp (- ({-92-}112 + 4*argWs_on_stack_ROUNDED_UP)) sp]
+
+ {- For each arg in args_offs_n_reps, examine the associated
+ CgRep to determine how many words there are. This gives a
+ bunch of offsets on the H stack. Move the first 6 words into
+ %o0 .. %o5 and the rest on the stack, starting at [%sp+92].
+ Use %g1 as a temp.
+ -}
+ ++ let doArgW (offW, wordNo)
+ | wordNo < 6
+ = [mkLD i0 (bytes_per_word * offW) (oReg wordNo)]
+ | otherwise
+ = [mkLD i0 (bytes_per_word * offW) g1,
+ mkST g1 sp (92 + bytes_per_word * (wordNo - 6))]
+ in
+ concatMap doArgW (zip offsets_to_pushW [0 ..])
+
+ {- Get the addr to call into %g1, bearing in mind that there's
+ an Addr# tag at the indicated location, and do the call:
+
+ ld [4*(1 /*tag*/ +addr_offW) + %i0], %g1
+ call %g1
+ -}
+ ++ [mkLD i0 (bytes_per_word * addr_offW) g1,
+ mkCALL g1,
+ mkNOP]
+
+ {- Depending on what the return type is, get the result
+ from %o0 or %o1:%o0 or %f0 or %f1:%f0.
+
+ st %o0, [%i0 + 4] -- 32 bit int
+ or
+ st %o0, [%i0 + 4] -- 64 bit int
+ st %o1, [%i0 + 8] -- or the other way round?
+ or
+ st %f0, [%i0 + 4] -- 32 bit float
+ or
+ st %f0, [%i0 + 4] -- 64 bit float
+ st %f1, [%i0 + 8] -- or the other way round?
+
+ -}
+ ++ let i32 = [mkST o0 i0 0]
+ i64 = [mkST o0 i0 0, mkST o1 i0 4]
+ f32 = [mkSTF f0 i0 0]
+ f64 = [mkSTF f0 i0 0, mkSTF f1 i0 4]
+ in
+ case r_rep of
+ NonPtrArg -> i32
+ DoubleArg -> f64
+ FloatArg -> f32
+ VoidArg -> []
+ other -> moan64 "ByteCodeFFI.mkMarshalCode_wrk(sparc)"
+ (ppr r_rep)
+
+ ++ [mkRET,
+ mkRESTORE_TRIVIAL] -- this is in the delay slot of the RET
+ )
+#elif powerpc_TARGET_ARCH
+
+ = let
+ bytes_per_word = 4
+
+ -- speaks for itself
+ w32_to_w8s_bigEndian :: Word32 -> [Word8]
+ w32_to_w8s_bigEndian w
+ = [fromIntegral (0xFF .&. (w `shiftR` 24)),
+ fromIntegral (0xFF .&. (w `shiftR` 16)),
+ fromIntegral (0xFF .&. (w `shiftR` 8)),
+ fromIntegral (0xFF .&. w)]
+
+ -- addr and result bits offsetsW
+ a_off = addr_offW * bytes_per_word
+ result_off = r_offW * bytes_per_word
+
+ linkageArea = 24
+ parameterArea = sum [ cgRepSizeW a_rep * bytes_per_word
+ | (_, a_rep) <- arg_offs_n_reps ]
+ savedRegisterArea = 4
+ frameSize = padTo16 (linkageArea + max parameterArea 32 + savedRegisterArea)
+ padTo16 x = case x `mod` 16 of
+ 0 -> x
+ y -> x - y + 16
+
+ pass_parameters [] _ _ = []
+ pass_parameters ((a_offW, a_rep):args) nextFPR offsetW =
+ let
+ haskellArgOffset = a_offW * bytes_per_word
+ offsetW' = offsetW + cgRepSizeW a_rep
+
+ pass_word w
+ | offsetW + w < 8 =
+ [0x801f0000 -- lwz rX, src(r31)
+ .|. (fromIntegral src .&. 0xFFFF)
+ .|. (fromIntegral (offsetW+w+3) `shiftL` 21)]
+ | otherwise =
+ [0x801f0000 -- lwz r0, src(r31)
+ .|. (fromIntegral src .&. 0xFFFF),
+ 0x90010000 -- stw r0, dst(r1)
+ .|. (fromIntegral dst .&. 0xFFFF)]
+ where
+ src = haskellArgOffset + w*bytes_per_word
+ dst = linkageArea + (offsetW+w) * bytes_per_word
+ in
+ case a_rep of
+ FloatArg | nextFPR < 14 ->
+ (0xc01f0000 -- lfs fX, haskellArgOffset(r31)
+ .|. (fromIntegral haskellArgOffset .&. 0xFFFF)
+ .|. (fromIntegral nextFPR `shiftL` 21))
+ : pass_parameters args (nextFPR+1) offsetW'
+ DoubleArg | nextFPR < 14 ->
+ (0xc81f0000 -- lfd fX, haskellArgOffset(r31)
+ .|. (fromIntegral haskellArgOffset .&. 0xFFFF)
+ .|. (fromIntegral nextFPR `shiftL` 21))
+ : pass_parameters args (nextFPR+1) offsetW'
+ _ ->
+ concatMap pass_word [0 .. cgRepSizeW a_rep - 1]
+ ++ pass_parameters args nextFPR offsetW'
+
+ gather_result = case r_rep of
+ VoidArg -> []
+ FloatArg ->
+ [0xd03f0000 .|. (fromIntegral result_off .&. 0xFFFF)]
+ -- stfs f1, result_off(r31)
+ DoubleArg ->
+ [0xd83f0000 .|. (fromIntegral result_off .&. 0xFFFF)]
+ -- stfs f1, result_off(r31)
+ _ | cgRepSizeW r_rep == 2 ->
+ [0x907f0000 .|. (fromIntegral result_off .&. 0xFFFF),
+ 0x909f0000 .|. (fromIntegral (result_off+4) .&. 0xFFFF)]
+ -- stw r3, result_off(r31)
+ -- stw r4, result_off+4(r31)
+ _ | cgRepSizeW r_rep == 1 ->
+ [0x907f0000 .|. (fromIntegral result_off .&. 0xFFFF)]
+ -- stw r3, result_off(r31)
+ in
+ concatMap w32_to_w8s_bigEndian $ [
+ 0x7c0802a6, -- mflr r0
+ 0x93e1fffc, -- stw r31,-4(r1)
+ 0x90010008, -- stw r0,8(r1)
+ 0x94210000 .|. (fromIntegral (-frameSize) .&. 0xFFFF),
+ -- stwu r1, -frameSize(r1)
+ 0x7c7f1b78 -- mr r31, r3
+ ] ++ pass_parameters arg_offs_n_reps 1 0 ++ [
+ 0x819f0000 .|. (fromIntegral a_off .&. 0xFFFF),
+ -- lwz r12, a_off(r31)
+ 0x7d8903a6, -- mtctr r12
+ 0x4e800421 -- bctrl
+ ] ++ gather_result ++ [
+ 0x80210000, -- lwz r1, 0(r1)
+ 0x83e1fffc, -- lwz r31, -4(r1)
+ 0x80010008, -- lwz r0, 8(r1)
+ 0x7c0803a6, -- mtlr r0
+ 0x4e800020 -- blr
+ ]
+#else
+
+ = undefined
+
+#endif