diff options
author | Michal Terepeta <michal.terepeta@gmail.com> | 2018-10-04 13:56:59 -0400 |
---|---|---|
committer | Ben Gamari <ben@smart-cactus.org> | 2018-10-07 18:36:07 -0400 |
commit | 5d5307f943d7581d7013ffe20af22233273fba06 (patch) | |
tree | ec9ae993cfa44d2cfe797e0422eb388933277100 /compiler/cmm | |
parent | e4bec29cb475b7e1431dad41fb8d4438814641c9 (diff) | |
download | haskell-5d5307f943d7581d7013ffe20af22233273fba06.tar.gz |
Add Int8# and Word8#
This is the first step of implementing:
https://github.com/ghc-proposals/ghc-proposals/pull/74
The main highlights/changes:
- `primops.txt.pp` gets two new sections for two new primitive types
for signed and unsigned 8-bit integers (`Int8#` and `Word8`
respectively) along with basic arithmetic and comparison
operations. `PrimRep`/`RuntimeRep` get two new constructors for
them. All of the primops translate into the existing `MachOP`s.
- For `CmmCall`s the codegen will now zero-extend the values at call
site (so that they can be moved to the right register) and then
truncate them back their original width.
- x86 native codegen needed some updates, since it wasn't able to deal
with the new widths, but all the changes are quite localized. LLVM
backend seems to just work.
Bumps binary submodule.
Signed-off-by: Michal Terepeta <michal.terepeta@gmail.com>
Test Plan: ./validate with new tests
Reviewers: hvr, goldfire, bgamari, simonmar
Subscribers: Abhiroop, dfeuer, rwbarton, thomie, carter
Differential Revision: https://phabricator.haskell.org/D4475
Diffstat (limited to 'compiler/cmm')
-rw-r--r-- | compiler/cmm/CmmExpr.hs | 5 | ||||
-rw-r--r-- | compiler/cmm/CmmMachOp.hs | 10 | ||||
-rw-r--r-- | compiler/cmm/CmmUtils.hs | 6 | ||||
-rw-r--r-- | compiler/cmm/MkGraph.hs | 73 | ||||
-rw-r--r-- | compiler/cmm/PprC.hs | 3 |
5 files changed, 84 insertions, 13 deletions
diff --git a/compiler/cmm/CmmExpr.hs b/compiler/cmm/CmmExpr.hs index d129d601f4..601b1d9b85 100644 --- a/compiler/cmm/CmmExpr.hs +++ b/compiler/cmm/CmmExpr.hs @@ -6,7 +6,7 @@ module CmmExpr ( CmmExpr(..), cmmExprType, cmmExprWidth, maybeInvertCmmExpr - , CmmReg(..), cmmRegType + , CmmReg(..), cmmRegType, cmmRegWidth , CmmLit(..), cmmLitType , LocalReg(..), localRegType , GlobalReg(..), isArgReg, globalRegType @@ -273,6 +273,9 @@ cmmRegType :: DynFlags -> CmmReg -> CmmType cmmRegType _ (CmmLocal reg) = localRegType reg cmmRegType dflags (CmmGlobal reg) = globalRegType dflags reg +cmmRegWidth :: DynFlags -> CmmReg -> Width +cmmRegWidth dflags = typeWidth . cmmRegType dflags + localRegType :: LocalReg -> CmmType localRegType (LocalReg _ rep) = rep diff --git a/compiler/cmm/CmmMachOp.hs b/compiler/cmm/CmmMachOp.hs index c5e9d9bf27..70e53d2325 100644 --- a/compiler/cmm/CmmMachOp.hs +++ b/compiler/cmm/CmmMachOp.hs @@ -107,6 +107,14 @@ data MachOp | MO_FS_Conv Width Width -- Float -> Signed int | MO_SS_Conv Width Width -- Signed int -> Signed int | MO_UU_Conv Width Width -- unsigned int -> unsigned int + | MO_XX_Conv Width Width -- int -> int; puts no requirements on the + -- contents of upper bits when extending; + -- narrowing is simply truncation; the only + -- expectation is that we can recover the + -- original value by applying the opposite + -- MO_XX_Conv, e.g., + -- MO_XX_CONV W64 W8 (MO_XX_CONV W8 W64 x) + -- is equivalent to just x. | MO_FF_Conv Width Width -- Float -> Float -- Vector element insertion and extraction operations @@ -392,6 +400,7 @@ machOpResultType dflags mop tys = MO_SS_Conv _ to -> cmmBits to MO_UU_Conv _ to -> cmmBits to + MO_XX_Conv _ to -> cmmBits to MO_FS_Conv _ to -> cmmBits to MO_SF_Conv _ to -> cmmFloat to MO_FF_Conv _ to -> cmmFloat to @@ -483,6 +492,7 @@ machOpArgReps dflags op = MO_SS_Conv from _ -> [from] MO_UU_Conv from _ -> [from] + MO_XX_Conv from _ -> [from] MO_SF_Conv from _ -> [from] MO_FS_Conv from _ -> [from] MO_FF_Conv from _ -> [from] diff --git a/compiler/cmm/CmmUtils.hs b/compiler/cmm/CmmUtils.hs index 42d64842e2..11e4df5bf4 100644 --- a/compiler/cmm/CmmUtils.hs +++ b/compiler/cmm/CmmUtils.hs @@ -97,6 +97,8 @@ primRepCmmType dflags LiftedRep = gcWord dflags primRepCmmType dflags UnliftedRep = gcWord dflags primRepCmmType dflags IntRep = bWord dflags primRepCmmType dflags WordRep = bWord dflags +primRepCmmType _ Int8Rep = b8 +primRepCmmType _ Word8Rep = b8 primRepCmmType _ Int64Rep = b64 primRepCmmType _ Word64Rep = b64 primRepCmmType dflags AddrRep = bWord dflags @@ -131,8 +133,10 @@ primRepForeignHint VoidRep = panic "primRepForeignHint:VoidRep" primRepForeignHint LiftedRep = AddrHint primRepForeignHint UnliftedRep = AddrHint primRepForeignHint IntRep = SignedHint -primRepForeignHint WordRep = NoHint +primRepForeignHint Int8Rep = SignedHint primRepForeignHint Int64Rep = SignedHint +primRepForeignHint WordRep = NoHint +primRepForeignHint Word8Rep = NoHint primRepForeignHint Word64Rep = NoHint primRepForeignHint AddrRep = AddrHint -- NB! AddrHint, but NonPtrArg primRepForeignHint FloatRep = NoHint diff --git a/compiler/cmm/MkGraph.hs b/compiler/cmm/MkGraph.hs index 70229d067d..bcd03bfa67 100644 --- a/compiler/cmm/MkGraph.hs +++ b/compiler/cmm/MkGraph.hs @@ -38,6 +38,7 @@ import OrdList import SMRep (ByteOff) import UniqSupply import Util +import Panic ----------------------------------------------------------------------------- @@ -309,18 +310,33 @@ copyIn :: DynFlags -> Convention -> Area copyIn dflags conv area formals extra_stk = (stk_size, [r | (_, RegisterParam r) <- args], map ci (stk_args ++ args)) where - ci (reg, RegisterParam r) = - CmmAssign (CmmLocal reg) (CmmReg (CmmGlobal r)) - ci (reg, StackParam off) = - CmmAssign (CmmLocal reg) (CmmLoad (CmmStackSlot area off) ty) - where ty = localRegType reg + -- See Note [Width of parameters] + ci (reg, RegisterParam r@(VanillaReg {})) = + let local = CmmLocal reg + global = CmmReg (CmmGlobal r) + width = cmmRegWidth dflags local + expr + | width == wordWidth dflags = global + | width < wordWidth dflags = + CmmMachOp (MO_XX_Conv (wordWidth dflags) width) [global] + | otherwise = panic "Parameter width greater than word width" - init_offset = widthInBytes (wordWidth dflags) -- infotable + in CmmAssign local expr - (stk_off, stk_args) = assignStack dflags init_offset localRegType extra_stk + -- Non VanillaRegs + ci (reg, RegisterParam r) = + CmmAssign (CmmLocal reg) (CmmReg (CmmGlobal r)) - (stk_size, args) = assignArgumentsPos dflags stk_off conv - localRegType formals + ci (reg, StackParam off) = + CmmAssign (CmmLocal reg) (CmmLoad (CmmStackSlot area off) ty) + where ty = localRegType reg + + init_offset = widthInBytes (wordWidth dflags) -- infotable + + (stk_off, stk_args) = assignStack dflags init_offset localRegType extra_stk + + (stk_size, args) = assignArgumentsPos dflags stk_off conv + localRegType formals -- Factoring out the common parts of the copyout functions yielded something -- more complicated: @@ -346,8 +362,21 @@ copyOutOflow dflags conv transfer area actuals updfr_off extra_stack_stuff where (regs, graph) = foldr co ([], mkNop) (setRA ++ args ++ stack_params) - co (v, RegisterParam r) (rs, ms) - = (r:rs, mkAssign (CmmGlobal r) v <*> ms) + -- See Note [Width of parameters] + co (v, RegisterParam r@(VanillaReg {})) (rs, ms) = + let width = cmmExprWidth dflags v + value + | width == wordWidth dflags = v + | width < wordWidth dflags = + CmmMachOp (MO_XX_Conv width (wordWidth dflags)) [v] + | otherwise = panic "Parameter width greater than word width" + + in (r:rs, mkAssign (CmmGlobal r) value <*> ms) + + -- Non VanillaRegs + co (v, RegisterParam r) (rs, ms) = + (r:rs, mkAssign (CmmGlobal r) v <*> ms) + co (v, StackParam off) (rs, ms) = (rs, mkStore (CmmStackSlot area off) v <*> ms) @@ -374,6 +403,28 @@ copyOutOflow dflags conv transfer area actuals updfr_off extra_stack_stuff (cmmExprType dflags) actuals +-- Note [Width of parameters] +-- +-- Consider passing a small (< word width) primitive like Int8# to a function +-- through a register. It's actually non-trivial to do this without +-- extending/narrowing: +-- * Global registers are considered to have native word width (i.e., 64-bits on +-- x86-64), so CmmLint would complain if we assigne an 8-bit parameter to a +-- global register. +-- * Same problem exists with LLVM IR. +-- * Lowering gets harder since on x86-32 not every register exposes its lower +-- 8 bits (e.g., for %eax we can use %al, but there isn't a corresponding +-- 8-bit register for %edi). So we would either need to extend/narrow anyway, +-- or complicate the calling convention. +-- So instead, we always extend every parameter smaller than native word width +-- in copyOutOflow and then truncate it back to the expected width in copyIn. +-- Note that we do this in cmm using MO_XX_Conv to avoid requiring +-- zero-/sign-extending - it's up to a backend to handle this in a most +-- efficient way (e.g., a simple register move) +-- +-- There was some discussion about this on this PR: +-- https://github.com/ghc-proposals/ghc-proposals/pull/74 + mkCallEntry :: DynFlags -> Convention -> [CmmFormal] -> [CmmFormal] -> (Int, [GlobalReg], CmmAGraph) diff --git a/compiler/cmm/PprC.hs b/compiler/cmm/PprC.hs index a979d49501..17fef7fc97 100644 --- a/compiler/cmm/PprC.hs +++ b/compiler/cmm/PprC.hs @@ -646,6 +646,9 @@ pprMachOp_for_C mop = case mop of MO_SS_Conv from to | from == to -> empty MO_SS_Conv _from to -> parens (machRep_S_CType to) + MO_XX_Conv from to | from == to -> empty + MO_XX_Conv _from to -> parens (machRep_U_CType to) + MO_FF_Conv from to | from == to -> empty MO_FF_Conv _from to -> parens (machRep_F_CType to) |