From 1e8a345a29ed034a6314e04b695e143ec1fff56f Mon Sep 17 00:00:00 2001 From: sewardj Date: Thu, 31 Jan 2002 18:01:34 +0000 Subject: [PATCH] [project @ 2002-01-31 18:01:34 by sewardj] Make a quite-large start on native code generator documentation. --- ghc/docs/comm/index.html | 1 + ghc/docs/comm/the-beast/ncg.html | 299 ++++++++++++++++++++++++++++++++++++++ 2 files changed, 300 insertions(+) create mode 100644 ghc/docs/comm/the-beast/ncg.html diff --git a/ghc/docs/comm/index.html b/ghc/docs/comm/index.html index dd381c9..f18713c 100644 --- a/ghc/docs/comm/index.html +++ b/ghc/docs/comm/index.html @@ -56,6 +56,7 @@

The Mighty Simplifier

The Evil Mangler

Alien Functions +

The Native Code Generator

RTS & Libraries

diff --git a/ghc/docs/comm/the-beast/ncg.html b/ghc/docs/comm/the-beast/ncg.html new file mode 100644 index 0000000..2f0d799 --- /dev/null +++ b/ghc/docs/comm/the-beast/ncg.html @@ -0,0 +1,299 @@ + + + + + The GHC Commentary - The Native Code Generator + + + +

The GHC Commentary - The Native Code Generator

+ On x86 and sparc platforms, GHC can generate assembly code + directly, without having to go via C. This can sometimes almost + halve compilation time, and avoids the fragility and + horribleness of the mangler. The NCG is enabled by default for + non-optimising compilation on x86 and sparc. For most programs + it generates code which runs only about 1% slower than that + created by gcc on x86s, so it is well worth using even with + optimised compilation. FP-intensive x86 programs see a bigger + slowdown, and all sparc code runs about 5% slower due to + us not filling branch delay slots. +

+ In the distant past + the NCG could also generate Alpha code, and that machinery + is still there, but will need extensive refurbishment to + get it going again, due to underlying infrastructural changes. + Budding hackers thinking of doing a PowerPC port would do well + to use the sparc bits as a starting point. +

+ The NCG has always been something of a second-class citizen + inside GHC, an unloved child, rather. This means that its + integration into the compiler as a whole is rather clumsy, which + brings some problems described below. That apart, the NCG + proper is fairly cleanly designed, as target-independent as it + reasonably can be, and so should not be difficult to retarget. +

+ The following details are correct as per the CVS head of end-Jan + 2002. + +

Overview

+ The top-level code generator fn is +

+ absCtoNat :: AbstractC -> UniqSM (SDoc, Pretty.Doc) +

+ The returned SDoc is for debugging, so is empty unless + you specify -ddump-stix. The Pretty.Doc + bit is the final assembly code. Translation involves three main + phases, the first and third of which are target-independent. +

Translation into the Stix representation. Stix + is a simple tree-like RTL-style language, in which you can + mention: +
+
- An infinite number of temporary, virtual registers. +
- The STG "magic" registers (MagicId), such as + the heap and stack pointers. +
- Literals and low-level machine ops (MachOp). +
- Simple address computations. +
- Reads and writes of: memory, virtual regs, and various STG + regs. +
- Labels and if ... goto ... style control-flow. +
+
+ Stix has two main associated types: +
+
- StixStmt -- trees executed for their side + effects: assignments, control transfers, and auxiliary junk + such as segment changes and literal data. +
- StixExpr -- trees which denote a value. +
+
+ Translation into Stix is almost completely + target-independent. Needed dependencies are knowledge of + word size and endianness, used when generating code to do + deal with half-word fields in info tables. This could be + abstracted out easily enough. Also, the Stix translation + needs to know which MagicIds map to registers + on the given target, and which are stored in offsets from + BaseReg. +
+ After initial Stix generation, the trees are cleaned up with + constant-folding and a little copy-propagation ("Stix + inlining", as the code misleadingly calls it). We take + the opportunity to translate MagicIds which are + stored in memory on the given target, into suitable memory + references. Those which are stored in registers are left + alone. There is also a half-hearted attempt to lift literal + strings to the top level in cases where nested strings have + been observed to give incorrect code in the past. +
+ Primitive machine-level operations will already be phrased in + terms of MachOps in the presented Abstract C, and + these are passed through unchanged. We comment only that the + MachOps have been chosen so as to be easy to + implement on all targets, and their meaning is intended to be + unambiguous, and the same on all targets, regardless of word + size or endianness. +
+
Instruction selection. This is the only majorly + target-specific phase. It turns Stix statements and + expressions into sequences of Instr, a data + type which is different for each architecture. + Instr, unsurprisingly, has various supporting + types, such as Reg, Operand, + Imm, etc. The generated instructions may refer + to specific machine registers, or to arbitrary virtual + registers, either those created within the instruction + selector, or those mentioned in the Stix passed to it. +
+ The instruction selectors live in MachCode.lhs. + The core functions, for each target, are: +
+ + getAmode :: StixExpr -> NatM Amode + getRegister :: StixExpr -> NatM Register + assignMem_IntCode :: PrimRep -> StixExpr -> StixExpr -> NatM InstrBlock + assignReg_IntCode :: PrimRep -> StixReg -> StixExpr -> NatM InstrBlock + +
+ The insn selectors use the "maximal munch" property. The + bizarrely-misnamed getRegister translates + expressions. A simplified version of its type is: +
+ getRegister :: StixExpr -> NatM (OrdList Instr, Reg) +
+ That is: it (monadically) turns a StixExpr into a + sequence of instructions, and a register, with the meaning + that after executing the (possibly empty) sequence of + instructions, the (possibly virtual) register will + hold the resulting value. The real situation is complicated + by the presence of fixed registers, and is detailed below. +
+ Maximal munch is a greedy algorithm and is known not to give + globally optimal code sequences, but it is good enough, and + fast and simple. Early incarnations of the NCG used something + more sophisticated, but that is long gone now. +
+ Similarly, getAmode translates a value, intended + to denote an address, into a sequence of insns leading up to + a (processor-specific) addressing mode. This stuff could be + done using the general getRegister selector, but + would necessarily generate poorer code, because the calculated + address would be forced into a register, which might be + unnecessary if it could partially or wholly be calculated + using an addressing mode. +
+ Finally, assignMem_IntCode and + assignReg_IntCode create instruction sequences to + calculate a value and store it in the given register, or at + the given address. Because these guys translate a statement, + not a value, they just return a sequence of insns and no + associated register. Floating-point and 64-bit integer + assignments have analogous selectors. +
+ Apart from the complexities of fixed vs floating registers, + discussed below, the instruction selector is as simple + as it can be. It looks long and scary but detailed + examination reveals it to be fairly straightforward. +
+
Register allocation. The register allocator, + AsmRegAlloc.lhs takes sequences of + Instrs which mention a mixture of real and + virtual registers, and returns a modified sequence referring + only to real ones. It is gloriously and entirely + target-independent. Well, not exactly true. Instead it + regards Instr (instructions) and Reg + (virtual and real registers) as abstract types, to which it has + the following interface: +
+ + insnFuture :: Instr -> InsnFuture + regUsage :: Instr -> RegUsage + patchRegs :: Instr -> (Reg -> Reg) -> Instr + +
+ insnFuture is used to (re)construct the graph of + all possible control transfers between the insns to be + allocated. regUsage returns the sets of registers + read and written by an instruction. And + patchRegs is used to apply the allocator's final + decision on virtual-to-real reg mapping to an instruction. +
+ Clearly these 3 fns have to be written anew for each + architecture. They are defined in + RegAllocInfo.lhs. Think twice, no, thrice, + before modifying them: making false claims about insn + behaviour will lead to hard-to-find register allocation + errors. +
+ AsmRegAlloc.lhs contains detailed comments about + how the allocator works. Here is a summary. The head honcho +
+ allocUsingTheseRegs :: [Instr] -> [Reg] -> (Bool, [Instr]) +
+ takes a list of instructions and a list of real registers + available for allocation, and maps as many of the virtual regs + in the input into real ones as it can. The returned + Bool indicates whether or not it was + successful. If so, that's the end of it. If not, the caller + of allocUsingTheseRegs will attempt spilling. + More of that later. What allocUsingTheseRegs + does is: +
+
- Implicitly number each instruction by its position in the + input list. +
  +
- Using insnFuture, create the set of all flow + edges -- possible control transfers -- within this set of + insns. +
  +
- Using regUsage and iterating around the flow + graph from the previous step, calculate, for each virtual + register, the set of flow edges on which it is live. +
  +
- Make a real-register committment map, which gives the set + of edges for which each real register is committed (in + use). These sets are initially empty. For each virtual + register, attempt to find a real register whose current + committment does not intersect that of the virtual + register -- ie, is uncommitted on all edges that the + virtual reg is live. If successful, this means the vreg + can be assigned to the realreg, so add the vreg's set to + the realreg's committment. +
  +
- If all the vregs were assigned to a realreg, use + patchInstr to apply the mapping to the insns themselves. +
+
+ Spilling +
+ If allocUsingTheseRegs fails, a baroque + mechanism comes into play. We now know that much simpler + schemes are available to do the same thing. Anyways: +
+ The logic above allocUsingTheseRegs, in + doGeneralAlloc and runRegAllocate, + observe that allocation has failed with some set R of real + registers. So they apply runRegAllocate a second + time to the code, but remove (typically) two registers from R + before doing so. This naturally fails too, but returns a + partially-allocated sequence. doGeneralAlloc + then inserts spill code into the sequence, and finally re-runs + allocUsingTheseRegs, but supplying the original, + unadulterated R. This is guaranteed to succeed since the two + registers previously removed from R are sufficient to allocate + all the spill/restore instructions added. +
+ Because x86 is very short of registers, and in the worst case + needs three removed from R, a softly-softly approach is used. + doGeneralAlloc first tries with zero regs removed + from R, then if that fails one, then two, etc. This means + allocUsingTheseRegs may get run several times + before a successful arrangement is arrived at. + findReservedRegs cooks up the sets of spill + registers to try with. +
+ The resulting machinery is complicated and the generated spill + code is appalling. The saving grace is that spills are very + rare so it doesn't matter much. I did not invent this -- I inherited it. +
+ Dealing with common cases fast +
+ The entire reg-alloc mechanism described so far is general and + correct, but expensive overkill for many simple code blocks. + So to begin with we use start off with + doSimpleAlloc, which attempts to do something + simple. It exploits the observation that if the total number + of virtual registers does not exceed the number of real ones + available, we can simply dole out a new realreg each time we + see mention of a new vreg, with no regard for control flow. + doSimpleAlloc therefore attempts this in a + single pass over the code. It gives up if it runs out of real + regs or sees any condition which renders the above observation + invalid (fixed reg uses, for example). +
+ This clever hack handles the majority of code blocks quickly. + It was copied from the previous reg-allocator (the + Mattson/Partain/Marlow/Gill one). +

+ +

Complications, observations, and possible improvements

+ +

How to debug the NCG without losing your sanity

+ +

How to debug the NCG without losing your sanity

+ + + + +

+ +Last modified: Fri Aug 10 11:47:41 EST 2001 + + + + -- 1.7.10.4