1 //===---------------------------------------------------------------------===//
2 // Random notes about and ideas for the SystemZ backend.
3 //===---------------------------------------------------------------------===//
5 The initial backend is deliberately restricted to z10. We should add support
6 for later architectures at some point.
10 SystemZDAGToDAGISel::SelectInlineAsmMemoryOperand() is passed "m" for all
11 inline asm memory constraints; it doesn't get to see the original constraint.
12 This means that it must conservatively treat all inline asm constraints
13 as the most restricted type, "R".
17 If an inline asm ties an i32 "r" result to an i64 input, the input
18 will be treated as an i32, leaving the upper bits uninitialised.
21 define void @f4(i32 *%dst) {
22 %val = call i32 asm "blah $0", "=r,0" (i64 103)
23 store i32 %val, i32 *%dst
27 from CodeGen/SystemZ/asm-09.ll will use LHI rather than LGHI.
28 to load 103. This seems to be a general target-independent problem.
32 The tuning of the choice between LOAD ADDRESS (LA) and addition in
33 SystemZISelDAGToDAG.cpp is suspect. It should be tweaked based on
34 performance measurements.
38 We don't support tail calls at present.
42 We don't support prefetching yet.
46 There is no scheduling support.
50 We don't use the BRANCH ON COUNT or BRANCH ON INDEX families of instruction.
54 We might want to use BRANCH ON CONDITION for conditional indirect calls
55 and conditional returns.
59 We don't use the condition code results of anything except comparisons.
61 Implementing this may need something more finely grained than the z_cmp
62 and z_ucmp that we have now. It might (or might not) also be useful to
63 have a mask of "don't care" values in conditional branches. For example,
64 integer comparisons never set CC to 3, so the bottom bit of the CC mask
65 isn't particularly relevant. JNLH and JE are equally good for testing
66 equality after an integer comparison, etc.
70 We don't use the LOAD AND TEST or TEST DATA CLASS instructions.
74 We could use the generic floating-point forms of LOAD COMPLEMENT,
75 LOAD NEGATIVE and LOAD POSITIVE in cases where we don't need the
76 condition codes. For example, we could use LCDFR instead of LCDBR.
80 We don't optimize block memory operations.
82 It's definitely worth using things like MVC, CLC, NC, XC and OC with
83 constant lengths. MVCIN may be worthwhile too.
85 We should probably implement things like memcpy using MVC with EXECUTE.
86 Likewise memcmp and CLC. MVCLE and CLCLE could be useful too.
90 We don't optimize string operations.
92 MVST, CLST, SRST and CUSE could be useful here. Some of the TRANSLATE
93 family might be too, although they are probably more difficult to exploit.
97 We don't take full advantage of builtins like fabsl because the calling
98 conventions require f128s to be returned by invisible reference.
102 ADD LOGICAL WITH SIGNED IMMEDIATE could be useful when we need to
103 produce a carry. SUBTRACT LOGICAL IMMEDIATE could be useful when we
104 need to produce a borrow. (Note that there are no memory forms of
105 ADD LOGICAL WITH CARRY and SUBTRACT LOGICAL WITH BORROW, so the high
106 part of 128-bit memory operations would probably need to be done
111 We don't use the halfword forms of LOAD REVERSED and STORE REVERSED
116 We could take advantage of the various ... UNDER MASK instructions,
117 such as ICM and STCM.
121 DAGCombiner can detect integer absolute, but there's not yet an associated
122 ISD opcode. We could add one and implement it using LOAD POSITIVE.
123 Negated absolutes could use LOAD NEGATIVE.
127 DAGCombiner doesn't yet fold truncations of extended loads. Functions like:
129 unsigned long f (unsigned long x, unsigned short *y)
131 return (x << 32) | *y;
141 but truncating the load would give:
151 define i64 @f1(i64 %a) {
156 ought to be implemented as:
162 but two-address optimisations reverse the order of the AND and force:
169 CodeGen/SystemZ/and-04.ll has several examples of this.
173 Out-of-range displacements are usually handled by loading the full
174 address into a register. In many cases it would be better to create
175 an anchor point instead. E.g. for:
177 define void @f4a(i128 *%aptr, i64 %base) {
178 %addr = add i64 %base, 524288
179 %bptr = inttoptr i64 %addr to i128 *
180 %a = load volatile i128 *%aptr
181 %b = load i128 *%bptr
182 %add = add i128 %a, %b
183 store i128 %add, i128 *%aptr
187 (from CodeGen/SystemZ/int-add-08.ll) we load %base+524288 and %base+524296
188 into separate registers, rather than using %base+524288 as a base for both.
192 Dynamic stack allocations round the size to 8 bytes and then allocate
193 that rounded amount. It would be simpler to subtract the unrounded
194 size from the copy of the stack pointer and then align the result.
195 See CodeGen/SystemZ/alloca-01.ll for an example.
199 Atomic loads and stores use the default compare-and-swap based implementation.
200 This is much too conservative in practice, since the architecture guarantees
201 that 1-, 2-, 4- and 8-byte loads and stores to aligned addresses are
206 If needed, we can support 16-byte atomics using LPQ, STPQ and CSDG.
210 We might want to model all access registers and use them to spill