[ARM] MVE integer min and max
[llvm-core.git] / lib / Target / X86 / README-FPStack.txt
blob39efd2dbcf1adc2424434730fded2e60cd6759bf
1 //===---------------------------------------------------------------------===//
2 // Random ideas for the X86 backend: FP stack related stuff
3 //===---------------------------------------------------------------------===//
5 //===---------------------------------------------------------------------===//
7 Some targets (e.g. athlons) prefer freep to fstp ST(0):
8 http://gcc.gnu.org/ml/gcc-patches/2004-04/msg00659.html
10 //===---------------------------------------------------------------------===//
12 This should use fiadd on chips where it is profitable:
13 double foo(double P, int *I) { return P+*I; }
15 We have fiadd patterns now but the followings have the same cost and
16 complexity. We need a way to specify the later is more profitable.
18 def FpADD32m  : FpI<(ops RFP:$dst, RFP:$src1, f32mem:$src2), OneArgFPRW,
19                     [(set RFP:$dst, (fadd RFP:$src1,
20                                      (extloadf64f32 addr:$src2)))]>;
21                 // ST(0) = ST(0) + [mem32]
23 def FpIADD32m : FpI<(ops RFP:$dst, RFP:$src1, i32mem:$src2), OneArgFPRW,
24                     [(set RFP:$dst, (fadd RFP:$src1,
25                                      (X86fild addr:$src2, i32)))]>;
26                 // ST(0) = ST(0) + [mem32int]
28 //===---------------------------------------------------------------------===//
30 The FP stackifier should handle simple permutates to reduce number of shuffle
31 instructions, e.g. turning:
33 fld P   ->              fld Q
34 fld Q                   fld P
35 fxch
37 or:
39 fxch    ->              fucomi
40 fucomi                  jl X
41 jg X
43 Ideas:
44 http://gcc.gnu.org/ml/gcc-patches/2004-11/msg02410.html
47 //===---------------------------------------------------------------------===//
49 Add a target specific hook to DAG combiner to handle SINT_TO_FP and
50 FP_TO_SINT when the source operand is already in memory.
52 //===---------------------------------------------------------------------===//
54 Open code rint,floor,ceil,trunc:
55 http://gcc.gnu.org/ml/gcc-patches/2004-08/msg02006.html
56 http://gcc.gnu.org/ml/gcc-patches/2004-08/msg02011.html
58 Opencode the sincos[f] libcall.
60 //===---------------------------------------------------------------------===//
62 None of the FPStack instructions are handled in
63 X86RegisterInfo::foldMemoryOperand, which prevents the spiller from
64 folding spill code into the instructions.
66 //===---------------------------------------------------------------------===//
68 Currently the x86 codegen isn't very good at mixing SSE and FPStack
69 code:
71 unsigned int foo(double x) { return x; }
73 foo:
74         subl $20, %esp
75         movsd 24(%esp), %xmm0
76         movsd %xmm0, 8(%esp)
77         fldl 8(%esp)
78         fisttpll (%esp)
79         movl (%esp), %eax
80         addl $20, %esp
81         ret
83 This just requires being smarter when custom expanding fptoui.
85 //===---------------------------------------------------------------------===//