do some serious surgery on CellSPU to get it back into a world

where it uses types consistently.

llvm-svn: 98532
This commit is contained in:
Chris Lattner 2010-03-15 05:53:47 +00:00
parent 26e6273772
commit eb319f36b9
4 changed files with 50 additions and 38 deletions

View File

@ -205,10 +205,9 @@ def CellSDKnand:
// Shift/rotate intrinsics:
//===----------------------------------------------------------------------===//
/* FIXME: These have (currently unenforced) type conflicts. */
def CellSDKshli:
Pat<(int_spu_si_shli (v4i32 VECREG:$rA), uimm7:$val),
(SHLIv4i32 VECREG:$rA, uimm7:$val)>;
(SHLIv4i32 VECREG:$rA, (TO_IMM32 imm:$val))>;
def CellSDKshlqbi:
Pat<(int_spu_si_shlqbi VECREG:$rA, R32C:$rB),
@ -216,7 +215,7 @@ def CellSDKshlqbi:
def CellSDKshlqii:
Pat<(int_spu_si_shlqbii VECREG:$rA, uimm7:$val),
(SHLQBIIv16i8 VECREG:$rA, uimm7:$val)>;
(SHLQBIIv16i8 VECREG:$rA, (TO_IMM32 imm:$val))>;
def CellSDKshlqby:
Pat<(int_spu_si_shlqby VECREG:$rA, R32C:$rB),
@ -224,7 +223,8 @@ def CellSDKshlqby:
def CellSDKshlqbyi:
Pat<(int_spu_si_shlqbyi VECREG:$rA, uimm7:$val),
(SHLQBYIv16i8 VECREG:$rA, uimm7:$val)>;
(SHLQBYIv16i8 VECREG:$rA, (TO_IMM32 imm:$val))>;
//===----------------------------------------------------------------------===//
// Branch/compare intrinsics:

View File

@ -155,8 +155,9 @@ multiclass CompareLogicalGreaterEqual64 {
defm I64LGE: CompareLogicalGreaterEqual64;
def : Pat<(setuge R64C:$rA, R64C:$rB), I64LGEr64.Fragment>;
def : Pat<(setuge (v2i64 VECREG:$rA), (v2i64 VECREG:$rB)),
I64LGEv2i64.Fragment>;
def : Pat<(v2i64 (setuge (v2i64 VECREG:$rA), (v2i64 VECREG:$rB))),
I64LGEv2i64.Fragment>;
// i64 setult:
def : I64SETCCNegCond<setult, I64LGEr64>;
@ -233,8 +234,8 @@ multiclass CompareGreaterEqual64 {
defm I64GE: CompareGreaterEqual64;
def : Pat<(setge R64C:$rA, R64C:$rB), I64GEr64.Fragment>;
def : Pat<(setge (v2i64 VECREG:$rA), (v2i64 VECREG:$rB)),
I64GEv2i64.Fragment>;
def : Pat<(v2i64 (setge (v2i64 VECREG:$rA), (v2i64 VECREG:$rB))),
I64GEv2i64.Fragment>;
// i64 setult:
def : I64SETCCNegCond<setlt, I64GEr64>;

View File

@ -2179,10 +2179,10 @@ multiclass ShiftLeftHalfwordImm
defm SHLHI : ShiftLeftHalfwordImm;
def : Pat<(SPUvec_shl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
(SHLHIv8i16 VECREG:$rA, uimm7:$val)>;
(SHLHIv8i16 VECREG:$rA, (TO_IMM16 uimm7:$val))>;
def : Pat<(shl R16C:$rA, (i32 uimm7:$val)),
(SHLHIr16 R16C:$rA, uimm7:$val)>;
(SHLHIr16 R16C:$rA, (TO_IMM16 uimm7:$val))>;
//===----------------------------------------------------------------------===//
@ -2410,8 +2410,8 @@ multiclass RotateLeftHalfwordImm
defm ROTHI: RotateLeftHalfwordImm;
def : Pat<(SPUvec_rotl VECREG:$rA, (i32 uimm7:$val)),
(ROTHIv8i16 VECREG:$rA, imm:$val)>;
def : Pat<(SPUvec_rotl (v8i16 VECREG:$rA), (i32 uimm7:$val)),
(ROTHIv8i16 VECREG:$rA, (TO_IMM16 imm:$val))>;
//-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
// Rotate word:
@ -2682,10 +2682,10 @@ def : Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i32 imm:$val)),
(ROTHMIv8i16 VECREG:$rA, imm:$val)>;
def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i16 imm:$val)),
(ROTHMIv8i16 VECREG:$rA, imm:$val)>;
(ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
def: Pat<(SPUvec_srl (v8i16 VECREG:$rA), (i8 imm:$val)),
(ROTHMIv8i16 VECREG:$rA, imm:$val)>;
(ROTHMIv8i16 VECREG:$rA, (TO_IMM32 imm:$val))>;
def ROTHMIr16:
ROTHMIInst<(outs R16C:$rT), (ins R16C:$rA, rothNeg7imm:$val),
@ -2695,10 +2695,10 @@ def: Pat<(srl R16C:$rA, (i32 uimm7:$val)),
(ROTHMIr16 R16C:$rA, uimm7:$val)>;
def: Pat<(srl R16C:$rA, (i16 uimm7:$val)),
(ROTHMIr16 R16C:$rA, uimm7:$val)>;
(ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
def: Pat<(srl R16C:$rA, (i8 uimm7:$val)),
(ROTHMIr16 R16C:$rA, uimm7:$val)>;
(ROTHMIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
// ROTM v4i32 form: See the ROTHM v8i16 comments.
class ROTMInst<dag OOL, dag IOL, list<dag> pattern>:
@ -2709,14 +2709,14 @@ def ROTMv4i32:
ROTMInst<(outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
[/* see patterns below - $rB must be negated */]>;
def : Pat<(SPUvec_srl VECREG:$rA, R32C:$rB),
def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R32C:$rB),
(ROTMv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
def : Pat<(SPUvec_srl VECREG:$rA, R16C:$rB),
def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R16C:$rB),
(ROTMv4i32 VECREG:$rA,
(SFIr32 (XSHWr16 R16C:$rB), 0))>;
def : Pat<(SPUvec_srl VECREG:$rA, R8C:$rB),
def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), R8C:$rB),
(ROTMv4i32 VECREG:$rA,
(SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
@ -2742,11 +2742,11 @@ def ROTMIv4i32:
[(set (v4i32 VECREG:$rT),
(SPUvec_srl VECREG:$rA, (i32 uimm7:$val)))]>;
def : Pat<(SPUvec_srl VECREG:$rA, (i16 uimm7:$val)),
(ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i16 uimm7:$val)),
(ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
def : Pat<(SPUvec_srl VECREG:$rA, (i8 uimm7:$val)),
(ROTMIv4i32 VECREG:$rA, uimm7:$val)>;
def : Pat<(SPUvec_srl (v4i32 VECREG:$rA), (i8 uimm7:$val)),
(ROTMIv4i32 VECREG:$rA, (TO_IMM32 uimm7:$val))>;
// ROTMI r32 form: know how to complement the immediate value.
def ROTMIr32:
@ -2755,10 +2755,10 @@ def ROTMIr32:
[(set R32C:$rT, (srl R32C:$rA, (i32 uimm7:$val)))]>;
def : Pat<(srl R32C:$rA, (i16 imm:$val)),
(ROTMIr32 R32C:$rA, uimm7:$val)>;
(ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
def : Pat<(srl R32C:$rA, (i8 imm:$val)),
(ROTMIr32 R32C:$rA, uimm7:$val)>;
(ROTMIr32 R32C:$rA, (TO_IMM32 uimm7:$val))>;
//-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~-~
// ROTQMBY: This is a vector form merely so that when used in an
@ -2916,14 +2916,14 @@ def ROTMAHv8i16:
"rotmah\t$rT, $rA, $rB", RotateShift,
[/* see patterns below - $rB must be negated */]>;
def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R32C:$rB),
(ROTMAHv8i16 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R16C:$rB),
(ROTMAHv8i16 VECREG:$rA,
(SFIr32 (XSHWr16 R16C:$rB), 0))>;
def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), R8C:$rB),
(ROTMAHv8i16 VECREG:$rA,
(SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
@ -2950,10 +2950,10 @@ def ROTMAHIv8i16:
(SPUvec_sra (v8i16 VECREG:$rA), (i32 uimm7:$val)))]>;
def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i16 uimm7:$val)),
(ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
(ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
def : Pat<(SPUvec_sra (v8i16 VECREG:$rA), (i8 uimm7:$val)),
(ROTMAHIv8i16 (v8i16 VECREG:$rA), (i32 uimm7:$val))>;
(ROTMAHIv8i16 (v8i16 VECREG:$rA), (TO_IMM32 uimm7:$val))>;
def ROTMAHIr16:
RRForm<0b01111110000, (outs R16C:$rT), (ins R16C:$rA, rothNeg7imm_i16:$val),
@ -2961,25 +2961,25 @@ def ROTMAHIr16:
[(set R16C:$rT, (sra R16C:$rA, (i16 uimm7:$val)))]>;
def : Pat<(sra R16C:$rA, (i32 imm:$val)),
(ROTMAHIr16 R16C:$rA, uimm7:$val)>;
(ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
def : Pat<(sra R16C:$rA, (i8 imm:$val)),
(ROTMAHIr16 R16C:$rA, uimm7:$val)>;
(ROTMAHIr16 R16C:$rA, (TO_IMM32 uimm7:$val))>;
def ROTMAv4i32:
RRForm<0b01011010000, (outs VECREG:$rT), (ins VECREG:$rA, R32C:$rB),
"rotma\t$rT, $rA, $rB", RotateShift,
[/* see patterns below - $rB must be negated */]>;
def : Pat<(SPUvec_sra VECREG:$rA, R32C:$rB),
(ROTMAv4i32 (v4i32 VECREG:$rA), (SFIr32 R32C:$rB, 0))>;
def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R32C:$rB),
(ROTMAv4i32 VECREG:$rA, (SFIr32 R32C:$rB, 0))>;
def : Pat<(SPUvec_sra VECREG:$rA, R16C:$rB),
(ROTMAv4i32 (v4i32 VECREG:$rA),
def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R16C:$rB),
(ROTMAv4i32 VECREG:$rA,
(SFIr32 (XSHWr16 R16C:$rB), 0))>;
def : Pat<(SPUvec_sra VECREG:$rA, R8C:$rB),
(ROTMAv4i32 (v4i32 VECREG:$rA),
def : Pat<(SPUvec_sra (v4i32 VECREG:$rA), R8C:$rB),
(ROTMAv4i32 VECREG:$rA,
(SFIr32 (XSHWr16 (XSBHr8 R8C:$rB)), 0))>;
def ROTMAr32:

View File

@ -9,6 +9,17 @@
// Cell SPU Instruction Operands:
//===----------------------------------------------------------------------===//
// TO_IMM32 - Convert an i8/i16 to i32.
def TO_IMM32 : SDNodeXForm<imm, [{
return getI32Imm(N->getZExtValue());
}]>;
// TO_IMM16 - Convert an i8/i32 to i16.
def TO_IMM16 : SDNodeXForm<imm, [{
return CurDAG->getTargetConstant(N->getZExtValue(), MVT::i16);
}]>;
def LO16 : SDNodeXForm<imm, [{
unsigned val = N->getZExtValue();
// Transformation function: get the low 16 bits.