src/llvm/lib/Target/X86/README-FPStack.txt

   1 //===---------------------------------------------------------------------===//
   2 // Random ideas for the X86 backend: FP stack related stuff
   3 //===---------------------------------------------------------------------===//
   4
   5 //===---------------------------------------------------------------------===//
   6
   7 Some targets (e.g. athlons) prefer freep to fstp ST(0):
   8 http://gcc.gnu.org/ml/gcc-patches/2004-04/msg00659.html
   9
  10 //===---------------------------------------------------------------------===//
  11
  12 This should use fiadd on chips where it is profitable:
  13 double foo(double P, int *I) { return P+*I; }
  14
  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.
  17
  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]
  22
  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]
  27
  28 //===---------------------------------------------------------------------===//
  29
  30 The FP stackifier should handle simple permutates to reduce number of shuffle
  31 instructions, e.g. turning:
  32
  33 fld P   ->              fld Q
  34 fld Q                   fld P
  35 fxch
  36
  37 or:
  38
  39 fxch    ->              fucomi
  40 fucomi                  jl X
  41 jg X
  42
  43 Ideas:
  44 http://gcc.gnu.org/ml/gcc-patches/2004-11/msg02410.html
  45
  46
  47 //===---------------------------------------------------------------------===//
  48
  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.
  51
  52 //===---------------------------------------------------------------------===//
  53
  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
  57
  58 Opencode the sincos[f] libcall.
  59
  60 //===---------------------------------------------------------------------===//
  61
  62 None of the FPStack instructions are handled in
  63 X86RegisterInfo::foldMemoryOperand, which prevents the spiller from
  64 folding spill code into the instructions.
  65
  66 //===---------------------------------------------------------------------===//
  67
  68 Currently the x86 codegen isn't very good at mixing SSE and FPStack
  69 code:
  70
  71 unsigned int foo(double x) { return x; }
  72
  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
  82
  83 This just requires being smarter when custom expanding fptoui.
  84
  85 //===---------------------------------------------------------------------===//