def brtarget : Operand<OtherVT>;
+// CVT conversion modes
+// These must match the enum in NVPTX.h
+def CvtNONE : PatLeaf<(i32 0x0)>;
+def CvtRNI : PatLeaf<(i32 0x1)>;
+def CvtRZI : PatLeaf<(i32 0x2)>;
+def CvtRMI : PatLeaf<(i32 0x3)>;
+def CvtRPI : PatLeaf<(i32 0x4)>;
+def CvtRN : PatLeaf<(i32 0x5)>;
+def CvtRZ : PatLeaf<(i32 0x6)>;
+def CvtRM : PatLeaf<(i32 0x7)>;
+def CvtRP : PatLeaf<(i32 0x8)>;
+
+def CvtNONE_FTZ : PatLeaf<(i32 0x10)>;
+def CvtRNI_FTZ : PatLeaf<(i32 0x11)>;
+def CvtRZI_FTZ : PatLeaf<(i32 0x12)>;
+def CvtRMI_FTZ : PatLeaf<(i32 0x13)>;
+def CvtRPI_FTZ : PatLeaf<(i32 0x14)>;
+def CvtRN_FTZ : PatLeaf<(i32 0x15)>;
+def CvtRZ_FTZ : PatLeaf<(i32 0x16)>;
+def CvtRM_FTZ : PatLeaf<(i32 0x17)>;
+def CvtRP_FTZ : PatLeaf<(i32 0x18)>;
+
+def CvtSAT : PatLeaf<(i32 0x20)>;
+def CvtSAT_FTZ : PatLeaf<(i32 0x30)>;
+
+def CvtMode : Operand<i32> {
+ let PrintMethod = "printCvtMode";
+}
+
+// Compare modes
+// These must match the enum in NVPTX.h
+def CmpEQ : PatLeaf<(i32 0)>;
+def CmpNE : PatLeaf<(i32 1)>;
+def CmpLT : PatLeaf<(i32 2)>;
+def CmpLE : PatLeaf<(i32 3)>;
+def CmpGT : PatLeaf<(i32 4)>;
+def CmpGE : PatLeaf<(i32 5)>;
+def CmpLO : PatLeaf<(i32 6)>;
+def CmpLS : PatLeaf<(i32 7)>;
+def CmpHI : PatLeaf<(i32 8)>;
+def CmpHS : PatLeaf<(i32 9)>;
+def CmpEQU : PatLeaf<(i32 10)>;
+def CmpNEU : PatLeaf<(i32 11)>;
+def CmpLTU : PatLeaf<(i32 12)>;
+def CmpLEU : PatLeaf<(i32 13)>;
+def CmpGTU : PatLeaf<(i32 14)>;
+def CmpGEU : PatLeaf<(i32 15)>;
+def CmpNUM : PatLeaf<(i32 16)>;
+def CmpNAN : PatLeaf<(i32 17)>;
+
+def CmpEQ_FTZ : PatLeaf<(i32 0x100)>;
+def CmpNE_FTZ : PatLeaf<(i32 0x101)>;
+def CmpLT_FTZ : PatLeaf<(i32 0x102)>;
+def CmpLE_FTZ : PatLeaf<(i32 0x103)>;
+def CmpGT_FTZ : PatLeaf<(i32 0x104)>;
+def CmpGE_FTZ : PatLeaf<(i32 0x105)>;
+def CmpLO_FTZ : PatLeaf<(i32 0x106)>;
+def CmpLS_FTZ : PatLeaf<(i32 0x107)>;
+def CmpHI_FTZ : PatLeaf<(i32 0x108)>;
+def CmpHS_FTZ : PatLeaf<(i32 0x109)>;
+def CmpEQU_FTZ : PatLeaf<(i32 0x10A)>;
+def CmpNEU_FTZ : PatLeaf<(i32 0x10B)>;
+def CmpLTU_FTZ : PatLeaf<(i32 0x10C)>;
+def CmpLEU_FTZ : PatLeaf<(i32 0x10D)>;
+def CmpGTU_FTZ : PatLeaf<(i32 0x10E)>;
+def CmpGEU_FTZ : PatLeaf<(i32 0x10F)>;
+def CmpNUM_FTZ : PatLeaf<(i32 0x110)>;
+def CmpNAN_FTZ : PatLeaf<(i32 0x111)>;
+
+def CmpMode : Operand<i32> {
+ let PrintMethod = "printCmpMode";
+}
+
+def F32ConstZero : Operand<f32>, PatLeaf<(f32 fpimm)>, SDNodeXForm<fpimm, [{
+ return CurDAG->getTargetConstantFP(0.0, MVT::f32);
+ }]>;
+def F32ConstOne : Operand<f32>, PatLeaf<(f32 fpimm)>, SDNodeXForm<fpimm, [{
+ return CurDAG->getTargetConstantFP(1.0, MVT::f32);
+ }]>;
+
//===----------------------------------------------------------------------===//
// NVPTX Instruction Predicate Definitions
//===----------------------------------------------------------------------===//
-def hasAtomRedG32 : Predicate<"Subtarget.hasAtomRedG32()">;
-def hasAtomRedS32 : Predicate<"Subtarget.hasAtomRedS32()">;
-def hasAtomRedGen32 : Predicate<"Subtarget.hasAtomRedGen32()">;
+def hasAtomRedG32 : Predicate<"Subtarget->hasAtomRedG32()">;
+def hasAtomRedS32 : Predicate<"Subtarget->hasAtomRedS32()">;
+def hasAtomRedGen32 : Predicate<"Subtarget->hasAtomRedGen32()">;
def useAtomRedG32forGen32 :
- Predicate<"!Subtarget.hasAtomRedGen32() && Subtarget.hasAtomRedG32()">;
-def hasBrkPt : Predicate<"Subtarget.hasBrkPt()">;
-def hasAtomRedG64 : Predicate<"Subtarget.hasAtomRedG64()">;
-def hasAtomRedS64 : Predicate<"Subtarget.hasAtomRedS64()">;
-def hasAtomRedGen64 : Predicate<"Subtarget.hasAtomRedGen64()">;
+ Predicate<"!Subtarget->hasAtomRedGen32() && Subtarget->hasAtomRedG32()">;
+def hasBrkPt : Predicate<"Subtarget->hasBrkPt()">;
+def hasAtomRedG64 : Predicate<"Subtarget->hasAtomRedG64()">;
+def hasAtomRedS64 : Predicate<"Subtarget->hasAtomRedS64()">;
+def hasAtomRedGen64 : Predicate<"Subtarget->hasAtomRedGen64()">;
def useAtomRedG64forGen64 :
- Predicate<"!Subtarget.hasAtomRedGen64() && Subtarget.hasAtomRedG64()">;
-def hasAtomAddF32 : Predicate<"Subtarget.hasAtomAddF32()">;
-def hasVote : Predicate<"Subtarget.hasVote()">;
-def hasDouble : Predicate<"Subtarget.hasDouble()">;
-def reqPTX20 : Predicate<"Subtarget.reqPTX20()">;
-def hasLDU : Predicate<"Subtarget.hasLDU()">;
-def hasGenericLdSt : Predicate<"Subtarget.hasGenericLdSt()">;
-
-def doF32FTZ : Predicate<"UseF32FTZ">;
-
-def doFMAF32 : Predicate<"doFMAF32">;
-def doFMAF32_ftz : Predicate<"(doFMAF32 && UseF32FTZ)">;
-def doFMAF32AGG : Predicate<"doFMAF32AGG">;
-def doFMAF32AGG_ftz : Predicate<"(doFMAF32AGG && UseF32FTZ)">;
-def doFMAF64 : Predicate<"doFMAF64">;
-def doFMAF64AGG : Predicate<"doFMAF64AGG">;
-def doFMADF32 : Predicate<"doFMADF32">;
-def doFMADF32_ftz : Predicate<"(doFMADF32 && UseF32FTZ)">;
+ Predicate<"!Subtarget->hasAtomRedGen64() && Subtarget->hasAtomRedG64()">;
+def hasAtomAddF32 : Predicate<"Subtarget->hasAtomAddF32()">;
+def hasVote : Predicate<"Subtarget->hasVote()">;
+def hasDouble : Predicate<"Subtarget->hasDouble()">;
+def reqPTX20 : Predicate<"Subtarget->reqPTX20()">;
+def hasLDG : Predicate<"Subtarget->hasLDG()">;
+def hasLDU : Predicate<"Subtarget->hasLDU()">;
+def hasGenericLdSt : Predicate<"Subtarget->hasGenericLdSt()">;
+
+def doF32FTZ : Predicate<"useF32FTZ()">;
+def doNoF32FTZ : Predicate<"!useF32FTZ()">;
def doMulWide : Predicate<"doMulWide">;
-def allowFMA : Predicate<"allowFMA">;
-def allowFMA_ftz : Predicate<"(allowFMA && UseF32FTZ)">;
+def allowFMA : Predicate<"allowFMA()">;
+def noFMA : Predicate<"!allowFMA()">;
-def do_DIVF32_APPROX : Predicate<"do_DIVF32_PREC==0">;
-def do_DIVF32_FULL : Predicate<"do_DIVF32_PREC==1">;
+def do_DIVF32_APPROX : Predicate<"getDivF32Level()==0">;
+def do_DIVF32_FULL : Predicate<"getDivF32Level()==1">;
-def hasHWROT32 : Predicate<"Subtarget.hasHWROT32()">;
-
-def true : Predicate<"1">;
-
-//===----------------------------------------------------------------------===//
-// Special Handling for 8-bit Operands and Operations
-//
-// PTX supports 8-bit signed and unsigned types, but does not support 8-bit
-// operations (like add, shift, etc) except for ld/st/cvt. SASS does not have
-// 8-bit registers.
-//
-// PTX ld, st and cvt instructions permit source and destination data operands
-// to be wider than the instruction-type size, so that narrow values may be
-// loaded, stored, and converted using regular-width registers.
-//
-// So in PTX generation, we
-// - always use 16-bit registers in place in 8-bit registers.
-// (8-bit variables should stay as 8-bit as they represent memory layout.)
-// - for the following 8-bit operations, we sign-ext/zero-ext the 8-bit values
-// before operation
-// . div
-// . rem
-// . neg (sign)
-// . set, setp
-// . shr
-//
-// We are patching the operations by inserting the cvt instructions in the
-// asm strings of the affected instructions.
-//
-// Since vector operations, except for ld/st, are eventually elementized. We
-// do not need to special-hand the vector 8-bit operations.
-//
-//
-//===----------------------------------------------------------------------===//
+def do_SQRTF32_APPROX : Predicate<"!usePrecSqrtF32()">;
+def do_SQRTF32_RN : Predicate<"usePrecSqrtF32()">;
-// Generate string block like
-// {
-// .reg .s16 %temp1;
-// .reg .s16 %temp2;
-// cvt.s16.s8 %temp1, %a;
-// cvt.s16.s8 %temp2, %b;
-// opc.s16 %dst, %temp1, %temp2;
-// }
-// when OpcStr=opc.s TypeStr=s16 CVTStr=cvt.s16.s8
-class Handle_i8rr<string OpcStr, string TypeStr, string CVTStr> {
- string s = !strconcat("{{\n\t",
- !strconcat(".reg .", !strconcat(TypeStr,
- !strconcat(" \t%temp1;\n\t",
- !strconcat(".reg .", !strconcat(TypeStr,
- !strconcat(" \t%temp2;\n\t",
- !strconcat(CVTStr, !strconcat(" \t%temp1, $a;\n\t",
- !strconcat(CVTStr, !strconcat(" \t%temp2, $b;\n\t",
- !strconcat(OpcStr, "16 \t$dst, %temp1, %temp2;\n\t}}"))))))))))));
-}
+def hasHWROT32 : Predicate<"Subtarget->hasHWROT32()">;
+def noHWROT32 : Predicate<"!Subtarget->hasHWROT32()">;
-// Generate string block like
-// {
-// .reg .s16 %temp1;
-// .reg .s16 %temp2;
-// cvt.s16.s8 %temp1, %a;
-// mov.b16 %temp2, %b;
-// cvt.s16.s8 %temp2, %temp2;
-// opc.s16 %dst, %temp1, %temp2;
-// }
-// when OpcStr=opc.s TypeStr=s16 CVTStr=cvt.s16.s8
-class Handle_i8ri<string OpcStr, string TypeStr, string CVTStr> {
- string s = !strconcat("{{\n\t",
- !strconcat(".reg .", !strconcat(TypeStr,
- !strconcat(" \t%temp1;\n\t",
- !strconcat(".reg .",
- !strconcat(TypeStr, !strconcat(" \t%temp2;\n\t",
- !strconcat(CVTStr, !strconcat(" \t%temp1, $a;\n\t",
- !strconcat("mov.b16 \t%temp2, $b;\n\t",
- !strconcat(CVTStr, !strconcat(" \t%temp2, %temp2;\n\t",
- !strconcat(OpcStr, "16 \t$dst, %temp1, %temp2;\n\t}}")))))))))))));
-}
+def true : Predicate<"1">;
-// Generate string block like
-// {
-// .reg .s16 %temp1;
-// .reg .s16 %temp2;
-// mov.b16 %temp1, %b;
-// cvt.s16.s8 %temp1, %temp1;
-// cvt.s16.s8 %temp2, %a;
-// opc.s16 %dst, %temp1, %temp2;
-// }
-// when OpcStr=opc.s TypeStr=s16 CVTStr=cvt.s16.s8
-class Handle_i8ir<string OpcStr, string TypeStr, string CVTStr> {
- string s = !strconcat("{{\n\t",
- !strconcat(".reg .", !strconcat(TypeStr,
- !strconcat(" \t%temp1;\n\t",
- !strconcat(".reg .", !strconcat(TypeStr,
- !strconcat(" \t%temp2;\n\t",
- !strconcat("mov.b16 \t%temp1, $a;\n\t",
- !strconcat(CVTStr, !strconcat(" \t%temp1, %temp1;\n\t",
- !strconcat(CVTStr, !strconcat(" \t%temp2, $b;\n\t",
- !strconcat(OpcStr, "16 \t$dst, %temp1, %temp2;\n\t}}")))))))))))));
-}
+def hasPTX31 : Predicate<"Subtarget->getPTXVersion() >= 31">;
//===----------------------------------------------------------------------===//
def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
!strconcat(OpcStr, "16 \t$dst, $a, $b;"),
[(set Int16Regs:$dst, (OpNode Int16Regs:$a, (imm):$b))]>;
- def i8rr : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a, Int8Regs:$b))]>;
- def i8ri : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a, (imm):$b))]>;
-}
-
-multiclass I3_i8<string OpcStr, SDNode OpNode, string TypeStr, string CVTStr> {
- def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
- Int64Regs:$b))]>;
- def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int64Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
- def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
- Int32Regs:$b))]>;
- def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
- def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
- Int16Regs:$b))]>;
- def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int16Regs:$dst, (OpNode Int16Regs:$a, (imm):$b))]>;
- def i8rr : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- Handle_i8rr<OpcStr, TypeStr, CVTStr>.s,
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a, Int8Regs:$b))]>;
- def i8ri : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- Handle_i8ri<OpcStr, TypeStr, CVTStr>.s,
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a, (imm):$b))]>;
-}
-
-multiclass I3_noi8<string OpcStr, SDNode OpNode> {
- def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int64Regs:$dst, (OpNode Int64Regs:$a,
- Int64Regs:$b))]>;
- def i64ri : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int64Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
- def i32rr : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int32Regs:$a,
- Int32Regs:$b))]>;
- def i32ri : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
- def i16rr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int16Regs:$dst, (OpNode Int16Regs:$a,
- Int16Regs:$b))]>;
- def i16ri : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int16Regs:$dst, (OpNode Int16Regs:$a, (imm):$b))]>;
}
multiclass ADD_SUB_INT_32<string OpcStr, SDNode OpNode> {
!strconcat(OpcStr, ".ftz.f32 \t$dst, $a, $b;"),
[(set Float32Regs:$dst,
(OpNode Float32Regs:$a, Float32Regs:$b))]>,
- Requires<[allowFMA_ftz]>;
+ Requires<[allowFMA, doF32FTZ]>;
def f32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, f32imm:$b),
!strconcat(OpcStr, ".ftz.f32 \t$dst, $a, $b;"),
[(set Float32Regs:$dst,
(OpNode Float32Regs:$a, fpimm:$b))]>,
- Requires<[allowFMA_ftz]>;
+ Requires<[allowFMA, doF32FTZ]>;
def f32rr : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, Float32Regs:$b),
!strconcat(OpcStr, ".f32 \t$dst, $a, $b;"),
(ins Float64Regs:$a, Float64Regs:$b),
!strconcat(OpcStr, ".rn.f64 \t$dst, $a, $b;"),
[(set Float64Regs:$dst,
- (OpNode Float64Regs:$a, Float64Regs:$b))]>;
+ (OpNode Float64Regs:$a, Float64Regs:$b))]>,
+ Requires<[noFMA]>;
def f64ri : NVPTXInst<(outs Float64Regs:$dst),
(ins Float64Regs:$a, f64imm:$b),
!strconcat(OpcStr, ".rn.f64 \t$dst, $a, $b;"),
[(set Float64Regs:$dst,
- (OpNode Float64Regs:$a, fpimm:$b))]>;
+ (OpNode Float64Regs:$a, fpimm:$b))]>,
+ Requires<[noFMA]>;
def f32rr_ftz : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, Float32Regs:$b),
!strconcat(OpcStr, ".rn.ftz.f32 \t$dst, $a, $b;"),
[(set Float32Regs:$dst,
(OpNode Float32Regs:$a, Float32Regs:$b))]>,
- Requires<[doF32FTZ]>;
+ Requires<[noFMA, doF32FTZ]>;
def f32ri_ftz : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, f32imm:$b),
!strconcat(OpcStr, ".rn.ftz.f32 \t$dst, $a, $b;"),
[(set Float32Regs:$dst,
(OpNode Float32Regs:$a, fpimm:$b))]>,
- Requires<[doF32FTZ]>;
+ Requires<[noFMA, doF32FTZ]>;
def f32rr : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, Float32Regs:$b),
!strconcat(OpcStr, ".rn.f32 \t$dst, $a, $b;"),
[(set Float32Regs:$dst,
- (OpNode Float32Regs:$a, Float32Regs:$b))]>;
+ (OpNode Float32Regs:$a, Float32Regs:$b))]>,
+ Requires<[noFMA]>;
def f32ri : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, f32imm:$b),
!strconcat(OpcStr, ".rn.f32 \t$dst, $a, $b;"),
[(set Float32Regs:$dst,
- (OpNode Float32Regs:$a, fpimm:$b))]>;
+ (OpNode Float32Regs:$a, fpimm:$b))]>,
+ Requires<[noFMA]>;
}
multiclass F2<string OpcStr, SDNode OpNode> {
// NVPTX Instructions.
//===----------------------------------------------------------------------===//
+//-----------------------------------
+// General Type Conversion
+//-----------------------------------
+
+let hasSideEffects = 0 in {
+// Generate a cvt to the given type from all possible types.
+// Each instance takes a CvtMode immediate that defines the conversion mode to
+// use. It can be CvtNONE to omit a conversion mode.
+multiclass CVT_FROM_ALL<string FromName, RegisterClass RC> {
+ def _s16 : NVPTXInst<(outs RC:$dst),
+ (ins Int16Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".s16\t$dst, $src;"),
+ []>;
+ def _u16 : NVPTXInst<(outs RC:$dst),
+ (ins Int16Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".u16\t$dst, $src;"),
+ []>;
+ def _f16 : NVPTXInst<(outs RC:$dst),
+ (ins Int16Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".f16\t$dst, $src;"),
+ []>;
+ def _s32 : NVPTXInst<(outs RC:$dst),
+ (ins Int32Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".s32\t$dst, $src;"),
+ []>;
+ def _u32 : NVPTXInst<(outs RC:$dst),
+ (ins Int32Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".u32\t$dst, $src;"),
+ []>;
+ def _s64 : NVPTXInst<(outs RC:$dst),
+ (ins Int64Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".s64\t$dst, $src;"),
+ []>;
+ def _u64 : NVPTXInst<(outs RC:$dst),
+ (ins Int64Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".u64\t$dst, $src;"),
+ []>;
+ def _f32 : NVPTXInst<(outs RC:$dst),
+ (ins Float32Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".f32\t$dst, $src;"),
+ []>;
+ def _f64 : NVPTXInst<(outs RC:$dst),
+ (ins Float64Regs:$src, CvtMode:$mode),
+ !strconcat("cvt${mode:base}${mode:ftz}${mode:sat}.",
+ FromName, ".f64\t$dst, $src;"),
+ []>;
+}
+
+// Generate a cvt to all possible types.
+defm CVT_s16 : CVT_FROM_ALL<"s16", Int16Regs>;
+defm CVT_u16 : CVT_FROM_ALL<"u16", Int16Regs>;
+defm CVT_f16 : CVT_FROM_ALL<"f16", Int16Regs>;
+defm CVT_s32 : CVT_FROM_ALL<"s32", Int32Regs>;
+defm CVT_u32 : CVT_FROM_ALL<"u32", Int32Regs>;
+defm CVT_s64 : CVT_FROM_ALL<"s64", Int64Regs>;
+defm CVT_u64 : CVT_FROM_ALL<"u64", Int64Regs>;
+defm CVT_f32 : CVT_FROM_ALL<"f32", Float32Regs>;
+defm CVT_f64 : CVT_FROM_ALL<"f64", Float64Regs>;
+
+// This set of cvt is different from the above. The type of the source
+// and target are the same.
+//
+def CVT_INREG_s16_s8 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
+ "cvt.s16.s8 \t$dst, $src;", []>;
+def CVT_INREG_s32_s8 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
+ "cvt.s32.s8 \t$dst, $src;", []>;
+def CVT_INREG_s32_s16 : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src),
+ "cvt.s32.s16 \t$dst, $src;", []>;
+def CVT_INREG_s64_s8 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
+ "cvt.s64.s8 \t$dst, $src;", []>;
+def CVT_INREG_s64_s16 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
+ "cvt.s64.s16 \t$dst, $src;", []>;
+def CVT_INREG_s64_s32 : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src),
+ "cvt.s64.s32 \t$dst, $src;", []>;
+}
+
//-----------------------------------
// Integer Arithmetic
//-----------------------------------
return CurDAG->getTargetConstant(temp.shl(v), MVT::i16);
}]>;
-def MULWIDES64 : NVPTXInst<(outs Int64Regs:$dst),
- (ins Int32Regs:$a, Int32Regs:$b),
+def MULWIDES64
+ : NVPTXInst<(outs Int64Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
+ "mul.wide.s32 \t$dst, $a, $b;", []>;
+def MULWIDES64Imm
+ : NVPTXInst<(outs Int64Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
"mul.wide.s32 \t$dst, $a, $b;", []>;
-def MULWIDES64Imm : NVPTXInst<(outs Int64Regs:$dst),
- (ins Int32Regs:$a, i64imm:$b),
+def MULWIDES64Imm64
+ : NVPTXInst<(outs Int64Regs:$dst), (ins Int32Regs:$a, i64imm:$b),
"mul.wide.s32 \t$dst, $a, $b;", []>;
-def MULWIDEU64 : NVPTXInst<(outs Int64Regs:$dst),
- (ins Int32Regs:$a, Int32Regs:$b),
+def MULWIDEU64
+ : NVPTXInst<(outs Int64Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
+ "mul.wide.u32 \t$dst, $a, $b;", []>;
+def MULWIDEU64Imm
+ : NVPTXInst<(outs Int64Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
"mul.wide.u32 \t$dst, $a, $b;", []>;
-def MULWIDEU64Imm : NVPTXInst<(outs Int64Regs:$dst),
- (ins Int32Regs:$a, i64imm:$b),
+def MULWIDEU64Imm64
+ : NVPTXInst<(outs Int64Regs:$dst), (ins Int32Regs:$a, i64imm:$b),
"mul.wide.u32 \t$dst, $a, $b;", []>;
-def MULWIDES32 : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int16Regs:$a, Int16Regs:$b),
+def MULWIDES32
+ : NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
"mul.wide.s16 \t$dst, $a, $b;", []>;
-def MULWIDES32Imm : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int16Regs:$a, i32imm:$b),
+def MULWIDES32Imm
+ : NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
+ "mul.wide.s16 \t$dst, $a, $b;", []>;
+def MULWIDES32Imm32
+ : NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, i32imm:$b),
"mul.wide.s16 \t$dst, $a, $b;", []>;
-def MULWIDEU32 : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int16Regs:$a, Int16Regs:$b),
- "mul.wide.u16 \t$dst, $a, $b;", []>;
-def MULWIDEU32Imm : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int16Regs:$a, i32imm:$b),
+def MULWIDEU32
+ : NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
+ "mul.wide.u16 \t$dst, $a, $b;", []>;
+def MULWIDEU32Imm
+ : NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
"mul.wide.u16 \t$dst, $a, $b;", []>;
+def MULWIDEU32Imm32
+ : NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, i32imm:$b),
+ "mul.wide.u16 \t$dst, $a, $b;", []>;
def : Pat<(shl (sext Int32Regs:$a), (i32 Int5Const:$b)),
(MULWIDES64Imm Int32Regs:$a, (SHL2MUL32 node:$b))>,
(MULWIDES64 Int32Regs:$a, Int32Regs:$b)>,
Requires<[doMulWide]>;
def : Pat<(mul (sext Int32Regs:$a), (i64 SInt32Const:$b)),
- (MULWIDES64Imm Int32Regs:$a, (i64 SInt32Const:$b))>,
+ (MULWIDES64Imm64 Int32Regs:$a, (i64 SInt32Const:$b))>,
Requires<[doMulWide]>;
def : Pat<(mul (zext Int32Regs:$a), (zext Int32Regs:$b)),
- (MULWIDEU64 Int32Regs:$a, Int32Regs:$b)>, Requires<[doMulWide]>;
+ (MULWIDEU64 Int32Regs:$a, Int32Regs:$b)>,
+ Requires<[doMulWide]>;
def : Pat<(mul (zext Int32Regs:$a), (i64 UInt32Const:$b)),
- (MULWIDEU64Imm Int32Regs:$a, (i64 UInt32Const:$b))>,
+ (MULWIDEU64Imm64 Int32Regs:$a, (i64 UInt32Const:$b))>,
Requires<[doMulWide]>;
def : Pat<(mul (sext Int16Regs:$a), (sext Int16Regs:$b)),
- (MULWIDES32 Int16Regs:$a, Int16Regs:$b)>, Requires<[doMulWide]>;
+ (MULWIDES32 Int16Regs:$a, Int16Regs:$b)>,
+ Requires<[doMulWide]>;
def : Pat<(mul (sext Int16Regs:$a), (i32 SInt16Const:$b)),
- (MULWIDES32Imm Int16Regs:$a, (i32 SInt16Const:$b))>,
+ (MULWIDES32Imm32 Int16Regs:$a, (i32 SInt16Const:$b))>,
Requires<[doMulWide]>;
def : Pat<(mul (zext Int16Regs:$a), (zext Int16Regs:$b)),
- (MULWIDEU32 Int16Regs:$a, Int16Regs:$b)>, Requires<[doMulWide]>;
+ (MULWIDEU32 Int16Regs:$a, Int16Regs:$b)>,
+ Requires<[doMulWide]>;
def : Pat<(mul (zext Int16Regs:$a), (i32 UInt16Const:$b)),
- (MULWIDEU32Imm Int16Regs:$a, (i32 UInt16Const:$b))>,
+ (MULWIDEU32Imm32 Int16Regs:$a, (i32 UInt16Const:$b))>,
+ Requires<[doMulWide]>;
+
+
+def SDTMulWide
+ : SDTypeProfile<1, 2, [SDTCisSameAs<1, 2>]>;
+def mul_wide_signed
+ : SDNode<"NVPTXISD::MUL_WIDE_SIGNED", SDTMulWide>;
+def mul_wide_unsigned
+ : SDNode<"NVPTXISD::MUL_WIDE_UNSIGNED", SDTMulWide>;
+
+def : Pat<(i32 (mul_wide_signed Int16Regs:$a, Int16Regs:$b)),
+ (MULWIDES32 Int16Regs:$a, Int16Regs:$b)>,
+ Requires<[doMulWide]>;
+def : Pat<(i32 (mul_wide_signed Int16Regs:$a, imm:$b)),
+ (MULWIDES32Imm Int16Regs:$a, imm:$b)>,
+ Requires<[doMulWide]>;
+def : Pat<(i32 (mul_wide_unsigned Int16Regs:$a, Int16Regs:$b)),
+ (MULWIDEU32 Int16Regs:$a, Int16Regs:$b)>,
+ Requires<[doMulWide]>;
+def : Pat<(i32 (mul_wide_unsigned Int16Regs:$a, imm:$b)),
+ (MULWIDEU32Imm Int16Regs:$a, imm:$b)>,
+ Requires<[doMulWide]>;
+
+
+def : Pat<(i64 (mul_wide_signed Int32Regs:$a, Int32Regs:$b)),
+ (MULWIDES64 Int32Regs:$a, Int32Regs:$b)>,
+ Requires<[doMulWide]>;
+def : Pat<(i64 (mul_wide_signed Int32Regs:$a, imm:$b)),
+ (MULWIDES64Imm Int32Regs:$a, imm:$b)>,
+ Requires<[doMulWide]>;
+def : Pat<(i64 (mul_wide_unsigned Int32Regs:$a, Int32Regs:$b)),
+ (MULWIDEU64 Int32Regs:$a, Int32Regs:$b)>,
+ Requires<[doMulWide]>;
+def : Pat<(i64 (mul_wide_unsigned Int32Regs:$a, imm:$b)),
+ (MULWIDEU64Imm Int32Regs:$a, imm:$b)>,
Requires<[doMulWide]>;
defm MULT : I3<"mul.lo.s", mul>;
-defm MULTHS : I3_noi8<"mul.hi.s", mulhs>;
-defm MULTHU : I3_noi8<"mul.hi.u", mulhu>;
-def MULTHSi8rr : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- !strconcat("{{ \n\t",
- !strconcat(".reg \t.s16 temp1; \n\t",
- !strconcat(".reg \t.s16 temp2; \n\t",
- !strconcat("cvt.s16.s8 \ttemp1, $a; \n\t",
- !strconcat("cvt.s16.s8 \ttemp2, $b; \n\t",
- !strconcat("mul.lo.s16 \t$dst, temp1, temp2; \n\t",
- !strconcat("shr.s16 \t$dst, $dst, 8; \n\t",
- !strconcat("}}", "")))))))),
- [(set Int8Regs:$dst, (mulhs Int8Regs:$a, Int8Regs:$b))]>;
-def MULTHSi8ri : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- !strconcat("{{ \n\t",
- !strconcat(".reg \t.s16 temp1; \n\t",
- !strconcat(".reg \t.s16 temp2; \n\t",
- !strconcat("cvt.s16.s8 \ttemp1, $a; \n\t",
- !strconcat("mov.b16 \ttemp2, $b; \n\t",
- !strconcat("cvt.s16.s8 \ttemp2, temp2; \n\t",
- !strconcat("mul.lo.s16 \t$dst, temp1, temp2; \n\t",
- !strconcat("shr.s16 \t$dst, $dst, 8; \n\t",
- !strconcat("}}", ""))))))))),
- [(set Int8Regs:$dst, (mulhs Int8Regs:$a, imm:$b))]>;
-def MULTHUi8rr : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- !strconcat("{{ \n\t",
- !strconcat(".reg \t.u16 temp1; \n\t",
- !strconcat(".reg \t.u16 temp2; \n\t",
- !strconcat("cvt.u16.u8 \ttemp1, $a; \n\t",
- !strconcat("cvt.u16.u8 \ttemp2, $b; \n\t",
- !strconcat("mul.lo.u16 \t$dst, temp1, temp2; \n\t",
- !strconcat("shr.u16 \t$dst, $dst, 8; \n\t",
- !strconcat("}}", "")))))))),
- [(set Int8Regs:$dst, (mulhu Int8Regs:$a, Int8Regs:$b))]>;
-def MULTHUi8ri : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- !strconcat("{{ \n\t",
- !strconcat(".reg \t.u16 temp1; \n\t",
- !strconcat(".reg \t.u16 temp2; \n\t",
- !strconcat("cvt.u16.u8 \ttemp1, $a; \n\t",
- !strconcat("mov.b16 \ttemp2, $b; \n\t",
- !strconcat("cvt.u16.u8 \ttemp2, temp2; \n\t",
- !strconcat("mul.lo.u16 \t$dst, temp1, temp2; \n\t",
- !strconcat("shr.u16 \t$dst, $dst, 8; \n\t",
- !strconcat("}}", ""))))))))),
- [(set Int8Regs:$dst, (mulhu Int8Regs:$a, imm:$b))]>;
-
-
-defm SDIV : I3_i8<"div.s", sdiv, "s16", "cvt.s16.s8">;
-defm UDIV : I3_i8<"div.u", udiv, "u16", "cvt.u16.u8">;
-
-defm SREM : I3_i8<"rem.s", srem, "s16", "cvt.s16.s8">;
+defm MULTHS : I3<"mul.hi.s", mulhs>;
+defm MULTHU : I3<"mul.hi.u", mulhu>;
+
+defm SDIV : I3<"div.s", sdiv>;
+defm UDIV : I3<"div.u", udiv>;
+
+defm SREM : I3<"rem.s", srem>;
// The ri version will not be selected as DAGCombiner::visitSREM will lower it.
-defm UREM : I3_i8<"rem.u", urem, "u16", "cvt.u16.u8">;
+defm UREM : I3<"rem.u", urem>;
// The ri version will not be selected as DAGCombiner::visitUREM will lower it.
-def MAD8rrr : NVPTXInst<(outs Int8Regs:$dst),
- (ins Int8Regs:$a, Int8Regs:$b, Int8Regs:$c),
- "mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int8Regs:$dst, (add (mul Int8Regs:$a, Int8Regs:$b),
- Int8Regs:$c))]>;
-def MAD8rri : NVPTXInst<(outs Int8Regs:$dst),
- (ins Int8Regs:$a, Int8Regs:$b, i8imm:$c),
- "mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int8Regs:$dst, (add (mul Int8Regs:$a, Int8Regs:$b),
- imm:$c))]>;
-def MAD8rir : NVPTXInst<(outs Int8Regs:$dst),
- (ins Int8Regs:$a, i8imm:$b, Int8Regs:$c),
- "mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int8Regs:$dst, (add (mul Int8Regs:$a, imm:$b),
- Int8Regs:$c))]>;
-def MAD8rii : NVPTXInst<(outs Int8Regs:$dst),
- (ins Int8Regs:$a, i8imm:$b, i8imm:$c),
- "mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int8Regs:$dst, (add (mul Int8Regs:$a, imm:$b),
- imm:$c))]>;
+def SDTIMAD
+ : SDTypeProfile<1, 3, [SDTCisSameAs<0, 1>, SDTCisInt<0>,
+ SDTCisInt<2>, SDTCisSameAs<0, 2>,
+ SDTCisSameAs<0, 3>]>;
+def imad
+ : SDNode<"NVPTXISD::IMAD", SDTIMAD>;
def MAD16rrr : NVPTXInst<(outs Int16Regs:$dst),
(ins Int16Regs:$a, Int16Regs:$b, Int16Regs:$c),
"mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int16Regs:$dst, (add
- (mul Int16Regs:$a, Int16Regs:$b), Int16Regs:$c))]>;
+ [(set Int16Regs:$dst,
+ (imad Int16Regs:$a, Int16Regs:$b, Int16Regs:$c))]>;
def MAD16rri : NVPTXInst<(outs Int16Regs:$dst),
(ins Int16Regs:$a, Int16Regs:$b, i16imm:$c),
"mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int16Regs:$dst, (add
- (mul Int16Regs:$a, Int16Regs:$b), imm:$c))]>;
+ [(set Int16Regs:$dst,
+ (imad Int16Regs:$a, Int16Regs:$b, imm:$c))]>;
def MAD16rir : NVPTXInst<(outs Int16Regs:$dst),
(ins Int16Regs:$a, i16imm:$b, Int16Regs:$c),
"mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int16Regs:$dst, (add
- (mul Int16Regs:$a, imm:$b), Int16Regs:$c))]>;
+ [(set Int16Regs:$dst,
+ (imad Int16Regs:$a, imm:$b, Int16Regs:$c))]>;
def MAD16rii : NVPTXInst<(outs Int16Regs:$dst),
(ins Int16Regs:$a, i16imm:$b, i16imm:$c),
"mad.lo.s16 \t$dst, $a, $b, $c;",
- [(set Int16Regs:$dst, (add (mul Int16Regs:$a, imm:$b),
- imm:$c))]>;
+ [(set Int16Regs:$dst,
+ (imad Int16Regs:$a, imm:$b, imm:$c))]>;
def MAD32rrr : NVPTXInst<(outs Int32Regs:$dst),
(ins Int32Regs:$a, Int32Regs:$b, Int32Regs:$c),
"mad.lo.s32 \t$dst, $a, $b, $c;",
- [(set Int32Regs:$dst, (add
- (mul Int32Regs:$a, Int32Regs:$b), Int32Regs:$c))]>;
+ [(set Int32Regs:$dst,
+ (imad Int32Regs:$a, Int32Regs:$b, Int32Regs:$c))]>;
def MAD32rri : NVPTXInst<(outs Int32Regs:$dst),
(ins Int32Regs:$a, Int32Regs:$b, i32imm:$c),
"mad.lo.s32 \t$dst, $a, $b, $c;",
- [(set Int32Regs:$dst, (add
- (mul Int32Regs:$a, Int32Regs:$b), imm:$c))]>;
+ [(set Int32Regs:$dst,
+ (imad Int32Regs:$a, Int32Regs:$b, imm:$c))]>;
def MAD32rir : NVPTXInst<(outs Int32Regs:$dst),
(ins Int32Regs:$a, i32imm:$b, Int32Regs:$c),
"mad.lo.s32 \t$dst, $a, $b, $c;",
- [(set Int32Regs:$dst, (add
- (mul Int32Regs:$a, imm:$b), Int32Regs:$c))]>;
+ [(set Int32Regs:$dst,
+ (imad Int32Regs:$a, imm:$b, Int32Regs:$c))]>;
def MAD32rii : NVPTXInst<(outs Int32Regs:$dst),
(ins Int32Regs:$a, i32imm:$b, i32imm:$c),
"mad.lo.s32 \t$dst, $a, $b, $c;",
- [(set Int32Regs:$dst, (add
- (mul Int32Regs:$a, imm:$b), imm:$c))]>;
+ [(set Int32Regs:$dst,
+ (imad Int32Regs:$a, imm:$b, imm:$c))]>;
def MAD64rrr : NVPTXInst<(outs Int64Regs:$dst),
(ins Int64Regs:$a, Int64Regs:$b, Int64Regs:$c),
"mad.lo.s64 \t$dst, $a, $b, $c;",
- [(set Int64Regs:$dst, (add
- (mul Int64Regs:$a, Int64Regs:$b), Int64Regs:$c))]>;
+ [(set Int64Regs:$dst,
+ (imad Int64Regs:$a, Int64Regs:$b, Int64Regs:$c))]>;
def MAD64rri : NVPTXInst<(outs Int64Regs:$dst),
(ins Int64Regs:$a, Int64Regs:$b, i64imm:$c),
"mad.lo.s64 \t$dst, $a, $b, $c;",
- [(set Int64Regs:$dst, (add
- (mul Int64Regs:$a, Int64Regs:$b), imm:$c))]>;
+ [(set Int64Regs:$dst,
+ (imad Int64Regs:$a, Int64Regs:$b, imm:$c))]>;
def MAD64rir : NVPTXInst<(outs Int64Regs:$dst),
(ins Int64Regs:$a, i64imm:$b, Int64Regs:$c),
"mad.lo.s64 \t$dst, $a, $b, $c;",
- [(set Int64Regs:$dst, (add
- (mul Int64Regs:$a, imm:$b), Int64Regs:$c))]>;
+ [(set Int64Regs:$dst,
+ (imad Int64Regs:$a, imm:$b, Int64Regs:$c))]>;
def MAD64rii : NVPTXInst<(outs Int64Regs:$dst),
(ins Int64Regs:$a, i64imm:$b, i64imm:$c),
"mad.lo.s64 \t$dst, $a, $b, $c;",
- [(set Int64Regs:$dst, (add
- (mul Int64Regs:$a, imm:$b), imm:$c))]>;
-
+ [(set Int64Regs:$dst,
+ (imad Int64Regs:$a, imm:$b, imm:$c))]>;
-def INEG8 : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$src),
- !strconcat("cvt.s16.s8 \t$dst, $src;\n\t",
- "neg.s16 \t$dst, $dst;"),
- [(set Int8Regs:$dst, (ineg Int8Regs:$src))]>;
def INEG16 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
"neg.s16 \t$dst, $src;",
[(set Int16Regs:$dst, (ineg Int16Regs:$src))]>;
[(set Float32Regs:$dst,
(fdiv Float32Regs:$a, Float32Regs:$b))]>,
Requires<[do_DIVF32_APPROX, doF32FTZ]>;
+def FDIV32approxri_ftz : NVPTXInst<(outs Float32Regs:$dst),
+ (ins Float32Regs:$a, f32imm:$b),
+ "div.approx.ftz.f32 \t$dst, $a, $b;",
+ [(set Float32Regs:$dst,
+ (fdiv Float32Regs:$a, fpimm:$b))]>,
+ Requires<[do_DIVF32_APPROX, doF32FTZ]>;
def FDIV32approxrr : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, Float32Regs:$b),
"div.approx.f32 \t$dst, $a, $b;",
[(set Float32Regs:$dst,
(fdiv Float32Regs:$a, Float32Regs:$b))]>,
Requires<[do_DIVF32_APPROX]>;
+def FDIV32approxri : NVPTXInst<(outs Float32Regs:$dst),
+ (ins Float32Regs:$a, f32imm:$b),
+ "div.approx.f32 \t$dst, $a, $b;",
+ [(set Float32Regs:$dst,
+ (fdiv Float32Regs:$a, fpimm:$b))]>,
+ Requires<[do_DIVF32_APPROX]>;
//
// F32 Semi-accurate reciprocal
//
(fdiv Float32Regs:$a, fpimm:$b))]>,
Requires<[reqPTX20]>;
+//
+// F32 rsqrt
+//
+
+def RSQRTF32approx1r : NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$b),
+ "rsqrt.approx.f32 \t$dst, $b;", []>;
+
+def: Pat<(fdiv FloatConst1, (int_nvvm_sqrt_f Float32Regs:$b)),
+ (RSQRTF32approx1r Float32Regs:$b)>,
+ Requires<[do_DIVF32_FULL, do_SQRTF32_APPROX, doNoF32FTZ]>;
multiclass FPCONTRACT32<string OpcStr, Predicate Pred> {
def rrr : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, Float32Regs:$b, Float32Regs:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float32Regs:$dst, (fadd
- (fmul Float32Regs:$a, Float32Regs:$b),
- Float32Regs:$c))]>, Requires<[Pred]>;
- // This is to WAR a weird bug in Tablegen that does not automatically
- // generate the following permutated rule rrr2 from the above rrr.
- // So we explicitly add it here. This happens to FMA32 only.
- // See the comments at FMAD32 and FMA32 for more information.
- def rrr2 : NVPTXInst<(outs Float32Regs:$dst),
- (ins Float32Regs:$a, Float32Regs:$b, Float32Regs:$c),
- !strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float32Regs:$dst, (fadd Float32Regs:$c,
- (fmul Float32Regs:$a, Float32Regs:$b)))]>,
+ [(set Float32Regs:$dst,
+ (fma Float32Regs:$a, Float32Regs:$b, Float32Regs:$c))]>,
Requires<[Pred]>;
def rri : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, Float32Regs:$b, f32imm:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float32Regs:$dst, (fadd
- (fmul Float32Regs:$a, Float32Regs:$b), fpimm:$c))]>,
+ [(set Float32Regs:$dst,
+ (fma Float32Regs:$a, Float32Regs:$b, fpimm:$c))]>,
Requires<[Pred]>;
def rir : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, f32imm:$b, Float32Regs:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float32Regs:$dst, (fadd
- (fmul Float32Regs:$a, fpimm:$b), Float32Regs:$c))]>,
+ [(set Float32Regs:$dst,
+ (fma Float32Regs:$a, fpimm:$b, Float32Regs:$c))]>,
Requires<[Pred]>;
def rii : NVPTXInst<(outs Float32Regs:$dst),
(ins Float32Regs:$a, f32imm:$b, f32imm:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float32Regs:$dst, (fadd
- (fmul Float32Regs:$a, fpimm:$b), fpimm:$c))]>,
+ [(set Float32Regs:$dst,
+ (fma Float32Regs:$a, fpimm:$b, fpimm:$c))]>,
Requires<[Pred]>;
}
def rrr : NVPTXInst<(outs Float64Regs:$dst),
(ins Float64Regs:$a, Float64Regs:$b, Float64Regs:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float64Regs:$dst, (fadd
- (fmul Float64Regs:$a, Float64Regs:$b),
- Float64Regs:$c))]>, Requires<[Pred]>;
+ [(set Float64Regs:$dst,
+ (fma Float64Regs:$a, Float64Regs:$b, Float64Regs:$c))]>,
+ Requires<[Pred]>;
def rri : NVPTXInst<(outs Float64Regs:$dst),
(ins Float64Regs:$a, Float64Regs:$b, f64imm:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float64Regs:$dst, (fadd (fmul Float64Regs:$a,
- Float64Regs:$b), fpimm:$c))]>, Requires<[Pred]>;
+ [(set Float64Regs:$dst,
+ (fma Float64Regs:$a, Float64Regs:$b, fpimm:$c))]>,
+ Requires<[Pred]>;
def rir : NVPTXInst<(outs Float64Regs:$dst),
(ins Float64Regs:$a, f64imm:$b, Float64Regs:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float64Regs:$dst, (fadd
- (fmul Float64Regs:$a, fpimm:$b), Float64Regs:$c))]>,
+ [(set Float64Regs:$dst,
+ (fma Float64Regs:$a, fpimm:$b, Float64Regs:$c))]>,
Requires<[Pred]>;
def rii : NVPTXInst<(outs Float64Regs:$dst),
(ins Float64Regs:$a, f64imm:$b, f64imm:$c),
!strconcat(OpcStr, " \t$dst, $a, $b, $c;"),
- [(set Float64Regs:$dst, (fadd
- (fmul Float64Regs:$a, fpimm:$b), fpimm:$c))]>,
+ [(set Float64Regs:$dst,
+ (fma Float64Regs:$a, fpimm:$b, fpimm:$c))]>,
Requires<[Pred]>;
}
-// Due to a unknown reason (most likely a bug in tablegen), tablegen does not
-// automatically generate the rrr2 rule from
-// the rrr rule (see FPCONTRACT32) for FMA32, though it does for FMAD32.
-// If we reverse the order of the following two lines, then rrr2 rule will be
-// generated for FMA32, but not for rrr.
-// Therefore, we manually write the rrr2 rule in FPCONTRACT32.
-defm FMAD32_ftz : FPCONTRACT32<"mad.ftz.f32", doFMADF32_ftz>;
-defm FMAD32 : FPCONTRACT32<"mad.f32", doFMADF32>;
-defm FMA32_ftz : FPCONTRACT32<"fma.rn.ftz.f32", doFMAF32_ftz>;
-defm FMA32 : FPCONTRACT32<"fma.rn.f32", doFMAF32>;
-defm FMA64 : FPCONTRACT64<"fma.rn.f64", doFMAF64>;
-
-// b*c-a => fmad(b, c, -a)
-multiclass FPCONTRACT32_SUB_PAT_MAD<NVPTXInst Inst, Predicate Pred> {
- def : Pat<(fsub (fmul Float32Regs:$b, Float32Regs:$c), Float32Regs:$a),
- (Inst Float32Regs:$b, Float32Regs:$c, (FNEGf32 Float32Regs:$a))>,
- Requires<[Pred]>;
-}
-
-// a-b*c => fmad(-b,c, a)
-// - legal because a-b*c <=> a+(-b*c) <=> a+(-b)*c
-// b*c-a => fmad(b, c, -a)
-// - legal because b*c-a <=> b*c+(-a)
-multiclass FPCONTRACT32_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
- def : Pat<(fsub Float32Regs:$a, (fmul Float32Regs:$b, Float32Regs:$c)),
- (Inst (FNEGf32 Float32Regs:$b), Float32Regs:$c, Float32Regs:$a)>,
- Requires<[Pred]>;
- def : Pat<(fsub (fmul Float32Regs:$b, Float32Regs:$c), Float32Regs:$a),
- (Inst Float32Regs:$b, Float32Regs:$c, (FNEGf32 Float32Regs:$a))>,
- Requires<[Pred]>;
-}
-
-// a-b*c => fmad(-b,c, a)
-// b*c-a => fmad(b, c, -a)
-multiclass FPCONTRACT64_SUB_PAT<NVPTXInst Inst, Predicate Pred> {
- def : Pat<(fsub Float64Regs:$a, (fmul Float64Regs:$b, Float64Regs:$c)),
- (Inst (FNEGf64 Float64Regs:$b), Float64Regs:$c, Float64Regs:$a)>,
- Requires<[Pred]>;
-
- def : Pat<(fsub (fmul Float64Regs:$b, Float64Regs:$c), Float64Regs:$a),
- (Inst Float64Regs:$b, Float64Regs:$c, (FNEGf64 Float64Regs:$a))>,
- Requires<[Pred]>;
-}
-
-defm FMAF32ext_ftz : FPCONTRACT32_SUB_PAT<FMA32_ftzrrr, doFMAF32AGG_ftz>;
-defm FMAF32ext : FPCONTRACT32_SUB_PAT<FMA32rrr, doFMAF32AGG>;
-defm FMADF32ext_ftz : FPCONTRACT32_SUB_PAT_MAD<FMAD32_ftzrrr, doFMADF32_ftz>;
-defm FMADF32ext : FPCONTRACT32_SUB_PAT_MAD<FMAD32rrr, doFMADF32>;
-defm FMAF64ext : FPCONTRACT64_SUB_PAT<FMA64rrr, doFMAF64AGG>;
+defm FMA32_ftz : FPCONTRACT32<"fma.rn.ftz.f32", doF32FTZ>;
+defm FMA32 : FPCONTRACT32<"fma.rn.f32", true>;
+defm FMA64 : FPCONTRACT64<"fma.rn.f64", true>;
def SINF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
"sin.approx.f32 \t$dst, $src;",
"cos.approx.f32 \t$dst, $src;",
[(set Float32Regs:$dst, (fcos Float32Regs:$src))]>;
+// Lower (frem x, y) into (sub x, (mul (floor (div x, y)) y))
+// e.g. "poor man's fmod()"
+
+// frem - f32 FTZ
+def : Pat<(frem Float32Regs:$x, Float32Regs:$y),
+ (FSUBf32rr_ftz Float32Regs:$x, (FMULf32rr_ftz (CVT_f32_f32
+ (FDIV32rr_prec_ftz Float32Regs:$x, Float32Regs:$y), CvtRMI_FTZ),
+ Float32Regs:$y))>,
+ Requires<[doF32FTZ]>;
+def : Pat<(frem Float32Regs:$x, fpimm:$y),
+ (FSUBf32rr_ftz Float32Regs:$x, (FMULf32ri_ftz (CVT_f32_f32
+ (FDIV32ri_prec_ftz Float32Regs:$x, fpimm:$y), CvtRMI_FTZ),
+ fpimm:$y))>,
+ Requires<[doF32FTZ]>;
+
+// frem - f32
+def : Pat<(frem Float32Regs:$x, Float32Regs:$y),
+ (FSUBf32rr Float32Regs:$x, (FMULf32rr (CVT_f32_f32
+ (FDIV32rr_prec Float32Regs:$x, Float32Regs:$y), CvtRMI),
+ Float32Regs:$y))>;
+def : Pat<(frem Float32Regs:$x, fpimm:$y),
+ (FSUBf32rr Float32Regs:$x, (FMULf32ri (CVT_f32_f32
+ (FDIV32ri_prec Float32Regs:$x, fpimm:$y), CvtRMI),
+ fpimm:$y))>;
+
+// frem - f64
+def : Pat<(frem Float64Regs:$x, Float64Regs:$y),
+ (FSUBf64rr Float64Regs:$x, (FMULf64rr (CVT_f64_f64
+ (FDIV64rr Float64Regs:$x, Float64Regs:$y), CvtRMI),
+ Float64Regs:$y))>;
+def : Pat<(frem Float64Regs:$x, fpimm:$y),
+ (FSUBf64rr Float64Regs:$x, (FMULf64ri (CVT_f64_f64
+ (FDIV64ri Float64Regs:$x, fpimm:$y), CvtRMI),
+ fpimm:$y))>;
+
//-----------------------------------
// Logical Arithmetic
//-----------------------------------
def b1ri: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$a, i1imm:$b),
!strconcat(OpcStr, ".pred \t$dst, $a, $b;"),
[(set Int1Regs:$dst, (OpNode Int1Regs:$a, imm:$b))]>;
- def b8rr: NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- !strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a, Int8Regs:$b))]>;
- def b8ri: NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- !strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a, imm:$b))]>;
def b16rr: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
!strconcat(OpcStr, ".b16 \t$dst, $a, $b;"),
[(set Int16Regs:$dst, (OpNode Int16Regs:$a,
def NOT1: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$src),
"not.pred \t$dst, $src;",
[(set Int1Regs:$dst, (not Int1Regs:$src))]>;
-def NOT8: NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$src),
- "not.b16 \t$dst, $src;",
- [(set Int8Regs:$dst, (not Int8Regs:$src))]>;
def NOT16: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
"not.b16 \t$dst, $src;",
[(set Int16Regs:$dst, (not Int16Regs:$src))]>;
!strconcat(OpcStr, "16 \t$dst, $a, $b;"),
[(set Int16Regs:$dst, (OpNode Int16Regs:$a,
(i32 imm:$b)))]>;
- def i8rr : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int32Regs:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a,
- Int32Regs:$b))]>;
- def i8ri : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i32imm:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a,
- (i32 imm:$b)))]>;
}
defm SHL : LSHIFT_FORMAT<"shl.b", shl>;
// For shifts, the second src operand must be 32-bit value
// Need to add cvt for the 8-bits.
-multiclass RSHIFT_FORMAT<string OpcStr, SDNode OpNode, string CVTStr> {
+multiclass RSHIFT_FORMAT<string OpcStr, SDNode OpNode> {
def i64rr : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a,
Int32Regs:$b),
!strconcat(OpcStr, "64 \t$dst, $a, $b;"),
!strconcat(OpcStr, "16 \t$dst, $a, $b;"),
[(set Int16Regs:$dst, (OpNode Int16Regs:$a,
(i32 imm:$b)))]>;
- def i8rr : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int32Regs:$b),
- !strconcat(CVTStr, !strconcat(" \t$dst, $a;\n\t",
- !strconcat(OpcStr, "16 \t$dst, $dst, $b;"))),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a,
- Int32Regs:$b))]>;
- def i8ri : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, i32imm:$b),
- !strconcat(CVTStr, !strconcat(" \t$dst, $a;\n\t",
- !strconcat(OpcStr, "16 \t$dst, $dst, $b;"))),
- [(set Int8Regs:$dst, (OpNode Int8Regs:$a,
- (i32 imm:$b)))]>;
}
-defm SRA : RSHIFT_FORMAT<"shr.s", sra, "cvt.s16.s8">;
-defm SRL : RSHIFT_FORMAT<"shr.u", srl, "cvt.u16.u8">;
+defm SRA : RSHIFT_FORMAT<"shr.s", sra>;
+defm SRL : RSHIFT_FORMAT<"shr.u", srl>;
+
+//
+// Rotate: use ptx shf instruction if available.
+//
+
+// 32 bit r2 = rotl r1, n
+// =>
+// r2 = shf.l r1, r1, n
+def ROTL32imm_hw : NVPTXInst<(outs Int32Regs:$dst),
+ (ins Int32Regs:$src, i32imm:$amt),
+ "shf.l.wrap.b32 \t$dst, $src, $src, $amt;",
+ [(set Int32Regs:$dst, (rotl Int32Regs:$src, (i32 imm:$amt)))]>,
+ Requires<[hasHWROT32]> ;
+
+def ROTL32reg_hw : NVPTXInst<(outs Int32Regs:$dst),
+ (ins Int32Regs:$src, Int32Regs:$amt),
+ "shf.l.wrap.b32 \t$dst, $src, $src, $amt;",
+ [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>,
+ Requires<[hasHWROT32]>;
+
+// 32 bit r2 = rotr r1, n
+// =>
+// r2 = shf.r r1, r1, n
+def ROTR32imm_hw : NVPTXInst<(outs Int32Regs:$dst),
+ (ins Int32Regs:$src, i32imm:$amt),
+ "shf.r.wrap.b32 \t$dst, $src, $src, $amt;",
+ [(set Int32Regs:$dst, (rotr Int32Regs:$src, (i32 imm:$amt)))]>,
+ Requires<[hasHWROT32]>;
+
+def ROTR32reg_hw : NVPTXInst<(outs Int32Regs:$dst),
+ (ins Int32Regs:$src, Int32Regs:$amt),
+ "shf.r.wrap.b32 \t$dst, $src, $src, $amt;",
+ [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>,
+ Requires<[hasHWROT32]>;
+//
+// Rotate: if ptx shf instruction is not available, then use shift+add
+//
// 32bit
def ROT32imm_sw : NVPTXInst<(outs Int32Regs:$dst),
(ins Int32Regs:$src, i32imm:$amt1, i32imm:$amt2),
}]>;
def : Pat<(rotl Int32Regs:$src, (i32 imm:$amt)),
- (ROT32imm_sw Int32Regs:$src, imm:$amt, (SUB_FRM_32 node:$amt))>;
+ (ROT32imm_sw Int32Regs:$src, imm:$amt, (SUB_FRM_32 node:$amt))>,
+ Requires<[noHWROT32]>;
def : Pat<(rotr Int32Regs:$src, (i32 imm:$amt)),
- (ROT32imm_sw Int32Regs:$src, (SUB_FRM_32 node:$amt), imm:$amt)>;
+ (ROT32imm_sw Int32Regs:$src, (SUB_FRM_32 node:$amt), imm:$amt)>,
+ Requires<[noHWROT32]>;
def ROTL32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src,
Int32Regs:$amt),
!strconcat("shr.b32 \t%rhs, $src, %amt2;\n\t",
!strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
!strconcat("}}", ""))))))))),
- [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>;
+ [(set Int32Regs:$dst, (rotl Int32Regs:$src, Int32Regs:$amt))]>,
+ Requires<[noHWROT32]>;
def ROTR32reg_sw : NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$src,
Int32Regs:$amt),
!strconcat("shl.b32 \t%rhs, $src, %amt2;\n\t",
!strconcat("add.u32 \t$dst, %lhs, %rhs;\n\t",
!strconcat("}}", ""))))))))),
- [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>;
+ [(set Int32Regs:$dst, (rotr Int32Regs:$src, Int32Regs:$amt))]>,
+ Requires<[noHWROT32]>;
// 64bit
def ROT64imm_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src,
!strconcat("}}", ""))))))))),
[(set Int64Regs:$dst, (rotr Int64Regs:$src, Int32Regs:$amt))]>;
+// BFE - bit-field extract
+
+multiclass BFE<string TyStr, RegisterClass RC> {
+ // BFE supports both 32-bit and 64-bit values, but the start and length
+ // operands are always 32-bit
+ def rrr
+ : NVPTXInst<(outs RC:$d),
+ (ins RC:$a, Int32Regs:$b, Int32Regs:$c),
+ !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
+ def rri
+ : NVPTXInst<(outs RC:$d),
+ (ins RC:$a, Int32Regs:$b, i32imm:$c),
+ !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
+ def rii
+ : NVPTXInst<(outs RC:$d),
+ (ins RC:$a, i32imm:$b, i32imm:$c),
+ !strconcat("bfe.", TyStr, " \t$d, $a, $b, $c;"), []>;
+}
+
+defm BFE_S32 : BFE<"s32", Int32Regs>;
+defm BFE_U32 : BFE<"u32", Int32Regs>;
+defm BFE_S64 : BFE<"s64", Int64Regs>;
+defm BFE_U64 : BFE<"u64", Int64Regs>;
+
+//-----------------------------------
+// General Comparison
+//-----------------------------------
+
+// General setp instructions
+multiclass SETP<string TypeStr, RegisterClass RC, Operand ImmCls> {
+ def rr : NVPTXInst<(outs Int1Regs:$dst),
+ (ins RC:$a, RC:$b, CmpMode:$cmp),
+ !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
+ []>;
+ def ri : NVPTXInst<(outs Int1Regs:$dst),
+ (ins RC:$a, ImmCls:$b, CmpMode:$cmp),
+ !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
+ []>;
+ def ir : NVPTXInst<(outs Int1Regs:$dst),
+ (ins ImmCls:$a, RC:$b, CmpMode:$cmp),
+ !strconcat("setp${cmp:base}${cmp:ftz}.", TypeStr, "\t$dst, $a, $b;"),
+ []>;
+}
+
+defm SETP_b16 : SETP<"b16", Int16Regs, i16imm>;
+defm SETP_s16 : SETP<"s16", Int16Regs, i16imm>;
+defm SETP_u16 : SETP<"u16", Int16Regs, i16imm>;
+defm SETP_b32 : SETP<"b32", Int32Regs, i32imm>;
+defm SETP_s32 : SETP<"s32", Int32Regs, i32imm>;
+defm SETP_u32 : SETP<"u32", Int32Regs, i32imm>;
+defm SETP_b64 : SETP<"b64", Int64Regs, i64imm>;
+defm SETP_s64 : SETP<"s64", Int64Regs, i64imm>;
+defm SETP_u64 : SETP<"u64", Int64Regs, i64imm>;
+defm SETP_f32 : SETP<"f32", Float32Regs, f32imm>;
+defm SETP_f64 : SETP<"f64", Float64Regs, f64imm>;
+
+// General set instructions
+multiclass SET<string TypeStr, RegisterClass RC, Operand ImmCls> {
+ def rr : NVPTXInst<(outs Int32Regs:$dst),
+ (ins RC:$a, RC:$b, CmpMode:$cmp),
+ !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
+ def ri : NVPTXInst<(outs Int32Regs:$dst),
+ (ins RC:$a, ImmCls:$b, CmpMode:$cmp),
+ !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
+ def ir : NVPTXInst<(outs Int32Regs:$dst),
+ (ins ImmCls:$a, RC:$b, CmpMode:$cmp),
+ !strconcat("set$cmp.", TypeStr, "\t$dst, $a, $b;"), []>;
+}
+
+defm SET_b16 : SET<"b16", Int16Regs, i16imm>;
+defm SET_s16 : SET<"s16", Int16Regs, i16imm>;
+defm SET_u16 : SET<"u16", Int16Regs, i16imm>;
+defm SET_b32 : SET<"b32", Int32Regs, i32imm>;
+defm SET_s32 : SET<"s32", Int32Regs, i32imm>;
+defm SET_u32 : SET<"u32", Int32Regs, i32imm>;
+defm SET_b64 : SET<"b64", Int64Regs, i64imm>;
+defm SET_s64 : SET<"s64", Int64Regs, i64imm>;
+defm SET_u64 : SET<"u64", Int64Regs, i64imm>;
+defm SET_f32 : SET<"f32", Float32Regs, f32imm>;
+defm SET_f64 : SET<"f64", Float64Regs, f64imm>;
+
+//-----------------------------------
+// General Selection
+//-----------------------------------
+
+// General selp instructions
+multiclass SELP<string TypeStr, RegisterClass RC, Operand ImmCls> {
+ def rr : NVPTXInst<(outs RC:$dst),
+ (ins RC:$a, RC:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
+ def ri : NVPTXInst<(outs RC:$dst),
+ (ins RC:$a, ImmCls:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
+ def ir : NVPTXInst<(outs RC:$dst),
+ (ins ImmCls:$a, RC:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
+ def ii : NVPTXInst<(outs RC:$dst),
+ (ins ImmCls:$a, ImmCls:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"), []>;
+}
+
+multiclass SELP_PATTERN<string TypeStr, RegisterClass RC, Operand ImmCls,
+ SDNode ImmNode> {
+ def rr : NVPTXInst<(outs RC:$dst),
+ (ins RC:$a, RC:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
+ [(set RC:$dst, (select Int1Regs:$p, RC:$a, RC:$b))]>;
+ def ri : NVPTXInst<(outs RC:$dst),
+ (ins RC:$a, ImmCls:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
+ [(set RC:$dst, (select Int1Regs:$p, RC:$a, ImmNode:$b))]>;
+ def ir : NVPTXInst<(outs RC:$dst),
+ (ins ImmCls:$a, RC:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
+ [(set RC:$dst, (select Int1Regs:$p, ImmNode:$a, RC:$b))]>;
+ def ii : NVPTXInst<(outs RC:$dst),
+ (ins ImmCls:$a, ImmCls:$b, Int1Regs:$p),
+ !strconcat("selp.", TypeStr, "\t$dst, $a, $b, $p;"),
+ [(set RC:$dst, (select Int1Regs:$p, ImmNode:$a, ImmNode:$b))]>;
+}
+
+defm SELP_b16 : SELP_PATTERN<"b16", Int16Regs, i16imm, imm>;
+defm SELP_s16 : SELP<"s16", Int16Regs, i16imm>;
+defm SELP_u16 : SELP<"u16", Int16Regs, i16imm>;
+defm SELP_b32 : SELP_PATTERN<"b32", Int32Regs, i32imm, imm>;
+defm SELP_s32 : SELP<"s32", Int32Regs, i32imm>;
+defm SELP_u32 : SELP<"u32", Int32Regs, i32imm>;
+defm SELP_b64 : SELP_PATTERN<"b64", Int64Regs, i64imm, imm>;
+defm SELP_s64 : SELP<"s64", Int64Regs, i64imm>;
+defm SELP_u64 : SELP<"u64", Int64Regs, i64imm>;
+defm SELP_f32 : SELP_PATTERN<"f32", Float32Regs, f32imm, fpimm>;
+defm SELP_f64 : SELP_PATTERN<"f64", Float64Regs, f64imm, fpimm>;
+
+//
+// Funnnel shift in clamp mode
+//
+// - SDNodes are created so they can be used in the DAG code,
+// e.g. NVPTXISelLowering (LowerShiftLeftParts and LowerShiftRightParts)
+//
+def SDTIntShiftDOp: SDTypeProfile<1, 3,
+ [SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>,
+ SDTCisInt<0>, SDTCisInt<3>]>;
+def FUN_SHFL_CLAMP : SDNode<"NVPTXISD::FUN_SHFL_CLAMP", SDTIntShiftDOp, []>;
+def FUN_SHFR_CLAMP : SDNode<"NVPTXISD::FUN_SHFR_CLAMP", SDTIntShiftDOp, []>;
+
+def FUNSHFLCLAMP : NVPTXInst<(outs Int32Regs:$dst),
+ (ins Int32Regs:$lo, Int32Regs:$hi, Int32Regs:$amt),
+ "shf.l.clamp.b32 \t$dst, $lo, $hi, $amt;",
+ [(set Int32Regs:$dst,
+ (FUN_SHFL_CLAMP Int32Regs:$lo,
+ Int32Regs:$hi, Int32Regs:$amt))]>;
+
+def FUNSHFRCLAMP : NVPTXInst<(outs Int32Regs:$dst),
+ (ins Int32Regs:$lo, Int32Regs:$hi, Int32Regs:$amt),
+ "shf.r.clamp.b32 \t$dst, $lo, $hi, $amt;",
+ [(set Int32Regs:$dst,
+ (FUN_SHFR_CLAMP Int32Regs:$lo,
+ Int32Regs:$hi, Int32Regs:$amt))]>;
//-----------------------------------
// Data Movement (Load / Store, Move)
"mov.u64 \t$dst, $a;",
[(set Int64Regs:$dst, (Wrapper tglobaladdr:$a))]>;
+// Get pointer to local stack
+def MOV_DEPOT_ADDR
+ : NVPTXInst<(outs Int32Regs:$d), (ins i32imm:$num),
+ "mov.u32 \t$d, __local_depot$num;", []>;
+def MOV_DEPOT_ADDR_64
+ : NVPTXInst<(outs Int64Regs:$d), (ins i32imm:$num),
+ "mov.u64 \t$d, __local_depot$num;", []>;
+
+
// copyPhysreg is hard-coded in NVPTXInstrInfo.cpp
let IsSimpleMove=1 in {
def IMOV1rr: NVPTXInst<(outs Int1Regs:$dst), (ins Int1Regs:$sss),
"mov.pred \t$dst, $sss;", []>;
-def IMOV8rr: NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$sss),
- "mov.u16 \t$dst, $sss;", []>;
def IMOV16rr: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$sss),
"mov.u16 \t$dst, $sss;", []>;
def IMOV32rr: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$sss),
def IMOV1ri: NVPTXInst<(outs Int1Regs:$dst), (ins i1imm:$src),
"mov.pred \t$dst, $src;",
[(set Int1Regs:$dst, imm:$src)]>;
-def IMOV8ri: NVPTXInst<(outs Int8Regs:$dst), (ins i8imm:$src),
- "mov.u16 \t$dst, $src;",
- [(set Int8Regs:$dst, imm:$src)]>;
def IMOV16ri: NVPTXInst<(outs Int16Regs:$dst), (ins i16imm:$src),
"mov.u16 \t$dst, $src;",
[(set Int16Regs:$dst, imm:$src)]>;
// Comparison and Selection
//-----------------------------------
-// Generate string block like
-// {
-// .reg .pred p;
-// setp.gt.s16 p, %a, %b;
-// selp.s16 %dst, -1, 0, p;
-// }
-// when OpcStr=setp.gt.s sz1=16 sz2=16 d=%dst a=%a b=%b
-class Set_Str<string OpcStr, string sz1, string sz2, string d, string a,
- string b> {
- string t1 = "{{\n\t.reg .pred p;\n\t";
- string t2 = !strconcat(t1 , OpcStr);
- string t3 = !strconcat(t2 , sz1);
- string t4 = !strconcat(t3 , " \tp, ");
- string t5 = !strconcat(t4 , a);
- string t6 = !strconcat(t5 , ", ");
- string t7 = !strconcat(t6 , b);
- string t8 = !strconcat(t7 , ";\n\tselp.s");
- string t9 = !strconcat(t8 , sz2);
- string t10 = !strconcat(t9, " \t");
- string t11 = !strconcat(t10, d);
- string s = !strconcat(t11, ", -1, 0, p;\n\t}}");
+multiclass ISET_FORMAT<PatFrag OpNode, PatLeaf Mode,
+ Instruction setp_16rr,
+ Instruction setp_16ri,
+ Instruction setp_16ir,
+ Instruction setp_32rr,
+ Instruction setp_32ri,
+ Instruction setp_32ir,
+ Instruction setp_64rr,
+ Instruction setp_64ri,
+ Instruction setp_64ir,
+ Instruction set_16rr,
+ Instruction set_16ri,
+ Instruction set_16ir,
+ Instruction set_32rr,
+ Instruction set_32ri,
+ Instruction set_32ir,
+ Instruction set_64rr,
+ Instruction set_64ri,
+ Instruction set_64ir> {
+ // i16 -> pred
+ def : Pat<(i1 (OpNode Int16Regs:$a, Int16Regs:$b)),
+ (setp_16rr Int16Regs:$a, Int16Regs:$b, Mode)>;
+ def : Pat<(i1 (OpNode Int16Regs:$a, imm:$b)),
+ (setp_16ri Int16Regs:$a, imm:$b, Mode)>;
+ def : Pat<(i1 (OpNode imm:$a, Int16Regs:$b)),
+ (setp_16ir imm:$a, Int16Regs:$b, Mode)>;
+ // i32 -> pred
+ def : Pat<(i1 (OpNode Int32Regs:$a, Int32Regs:$b)),
+ (setp_32rr Int32Regs:$a, Int32Regs:$b, Mode)>;
+ def : Pat<(i1 (OpNode Int32Regs:$a, imm:$b)),
+ (setp_32ri Int32Regs:$a, imm:$b, Mode)>;
+ def : Pat<(i1 (OpNode imm:$a, Int32Regs:$b)),
+ (setp_32ir imm:$a, Int32Regs:$b, Mode)>;
+ // i64 -> pred
+ def : Pat<(i1 (OpNode Int64Regs:$a, Int64Regs:$b)),
+ (setp_64rr Int64Regs:$a, Int64Regs:$b, Mode)>;
+ def : Pat<(i1 (OpNode Int64Regs:$a, imm:$b)),
+ (setp_64ri Int64Regs:$a, imm:$b, Mode)>;
+ def : Pat<(i1 (OpNode imm:$a, Int64Regs:$b)),
+ (setp_64ir imm:$a, Int64Regs:$b, Mode)>;
+
+ // i16 -> i32
+ def : Pat<(i32 (OpNode Int16Regs:$a, Int16Regs:$b)),
+ (set_16rr Int16Regs:$a, Int16Regs:$b, Mode)>;
+ def : Pat<(i32 (OpNode Int16Regs:$a, imm:$b)),
+ (set_16ri Int16Regs:$a, imm:$b, Mode)>;
+ def : Pat<(i32 (OpNode imm:$a, Int16Regs:$b)),
+ (set_16ir imm:$a, Int16Regs:$b, Mode)>;
+ // i32 -> i32
+ def : Pat<(i32 (OpNode Int32Regs:$a, Int32Regs:$b)),
+ (set_32rr Int32Regs:$a, Int32Regs:$b, Mode)>;
+ def : Pat<(i32 (OpNode Int32Regs:$a, imm:$b)),
+ (set_32ri Int32Regs:$a, imm:$b, Mode)>;
+ def : Pat<(i32 (OpNode imm:$a, Int32Regs:$b)),
+ (set_32ir imm:$a, Int32Regs:$b, Mode)>;
+ // i64 -> i32
+ def : Pat<(i32 (OpNode Int64Regs:$a, Int64Regs:$b)),
+ (set_64rr Int64Regs:$a, Int64Regs:$b, Mode)>;
+ def : Pat<(i32 (OpNode Int64Regs:$a, imm:$b)),
+ (set_64ri Int64Regs:$a, imm:$b, Mode)>;
+ def : Pat<(i32 (OpNode imm:$a, Int64Regs:$b)),
+ (set_64ir imm:$a, Int64Regs:$b, Mode)>;
}
-// Generate string block like
-// {
-// .reg .pred p;
-// .reg .s16 %temp1;
-// .reg .s16 %temp2;
-// cvt.s16.s8 %temp1, %a;
-// cvt s16.s8 %temp1, %b;
-// setp.gt.s16 p, %temp1, %temp2;
-// selp.s16 %dst, -1, 0, p;
-// }
-// when OpcStr=setp.gt.s d=%dst a=%a b=%b type=s16 cvt=cvt.s16.s8
-class Set_Stri8<string OpcStr, string d, string a, string b, string type,
- string cvt> {
- string t1 = "{{\n\t.reg .pred p;\n\t";
- string t2 = !strconcat(t1, ".reg .");
- string t3 = !strconcat(t2, type);
- string t4 = !strconcat(t3, " %temp1;\n\t");
- string t5 = !strconcat(t4, ".reg .");
- string t6 = !strconcat(t5, type);
- string t7 = !strconcat(t6, " %temp2;\n\t");
- string t8 = !strconcat(t7, cvt);
- string t9 = !strconcat(t8, " \t%temp1, ");
- string t10 = !strconcat(t9, a);
- string t11 = !strconcat(t10, ";\n\t");
- string t12 = !strconcat(t11, cvt);
- string t13 = !strconcat(t12, " \t%temp2, ");
- string t14 = !strconcat(t13, b);
- string t15 = !strconcat(t14, ";\n\t");
- string t16 = !strconcat(t15, OpcStr);
- string t17 = !strconcat(t16, "16");
- string t18 = !strconcat(t17, " \tp, %temp1, %temp2;\n\t");
- string t19 = !strconcat(t18, "selp.s16 \t");
- string t20 = !strconcat(t19, d);
- string s = !strconcat(t20, ", -1, 0, p;\n\t}}");
+multiclass ISET_FORMAT_SIGNED<PatFrag OpNode, PatLeaf Mode>
+ : ISET_FORMAT<OpNode, Mode,
+ SETP_s16rr, SETP_s16ri, SETP_s16ir,
+ SETP_s32rr, SETP_s32ri, SETP_s32ir,
+ SETP_s64rr, SETP_s64ri, SETP_s64ir,
+ SET_s16rr, SET_s16ri, SET_s16ir,
+ SET_s32rr, SET_s32ri, SET_s32ir,
+ SET_s64rr, SET_s64ri, SET_s64ir> {
+ // TableGen doesn't like empty multiclasses
+ def : PatLeaf<(i32 0)>;
}
-multiclass ISET_FORMAT<string OpcStr, string OpcStr_u32, PatFrag OpNode,
- string TypeStr, string CVTStr> {
- def i8rr_toi8: NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- Set_Stri8<OpcStr, "$dst", "$a", "$b", TypeStr, CVTStr>.s,
- []>;
- def i16rr_toi16: NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a,
- Int16Regs:$b),
- Set_Str<OpcStr, "16", "16", "$dst", "$a", "$b">.s,
- []>;
- def i32rr_toi32: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
- Int32Regs:$b),
- Set_Str<OpcStr, "32", "32", "$dst", "$a", "$b">.s,
- []>;
- def i64rr_toi64: NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$a,
- Int64Regs:$b),
- Set_Str<OpcStr, "64", "64", "$dst", "$a", "$b">.s,
- []>;
-
- def i8rr_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- Handle_i8rr<OpcStr, TypeStr, CVTStr>.s,
- [(set Int1Regs:$dst, (OpNode Int8Regs:$a, Int8Regs:$b))]>;
- def i8ri_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- Handle_i8ri<OpcStr, TypeStr, CVTStr>.s,
- [(set Int1Regs:$dst, (OpNode Int8Regs:$a, imm:$b))]>;
- def i8ir_p: NVPTXInst<(outs Int1Regs:$dst), (ins i8imm:$a, Int8Regs:$b),
- Handle_i8ir<OpcStr, TypeStr, CVTStr>.s,
- [(set Int1Regs:$dst, (OpNode imm:$a, Int8Regs:$b))]>;
- def i16rr_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Int16Regs:$a, Int16Regs:$b))]>;
- def i16ri_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Int16Regs:$a, imm:$b))]>;
- def i16ir_p: NVPTXInst<(outs Int1Regs:$dst), (ins i16imm:$a, Int16Regs:$b),
- !strconcat(OpcStr, "16 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode imm:$a, Int16Regs:$b))]>;
- def i32rr_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int32Regs:$a, Int32Regs:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Int32Regs:$a, Int32Regs:$b))]>;
- def i32ri_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
- def i32ir_p: NVPTXInst<(outs Int1Regs:$dst), (ins i32imm:$a, Int32Regs:$b),
- !strconcat(OpcStr, "32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode imm:$a, Int32Regs:$b))]>;
- def i64rr_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int64Regs:$a, Int64Regs:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Int64Regs:$a, Int64Regs:$b))]>;
- def i64ri_p: NVPTXInst<(outs Int1Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
- def i64ir_p: NVPTXInst<(outs Int1Regs:$dst), (ins i64imm:$a, Int64Regs:$b),
- !strconcat(OpcStr, "64 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode imm:$a, Int64Regs:$b))]>;
-
- def i8rr_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int8Regs:$a, Int8Regs:$b),
- Handle_i8rr<OpcStr_u32, TypeStr, CVTStr>.s,
- [(set Int32Regs:$dst, (OpNode Int8Regs:$a, Int8Regs:$b))]>;
- def i8ri_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int8Regs:$a, i8imm:$b),
- Handle_i8ri<OpcStr_u32, TypeStr, CVTStr>.s,
- [(set Int32Regs:$dst, (OpNode Int8Regs:$a, imm:$b))]>;
- def i8ir_u32: NVPTXInst<(outs Int32Regs:$dst), (ins i8imm:$a, Int8Regs:$b),
- Handle_i8ir<OpcStr_u32, TypeStr, CVTStr>.s,
- [(set Int32Regs:$dst, (OpNode imm:$a, Int8Regs:$b))]>;
- def i16rr_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a,
- Int16Regs:$b),
- !strconcat(OpcStr_u32, "16 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int16Regs:$a, Int16Regs:$b))]>;
- def i16ri_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int16Regs:$a, i16imm:$b),
- !strconcat(OpcStr_u32, "16 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int16Regs:$a, imm:$b))]>;
- def i16ir_u32: NVPTXInst<(outs Int32Regs:$dst), (ins i16imm:$a, Int16Regs:$b),
- !strconcat(OpcStr_u32, "16 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode imm:$a, Int16Regs:$b))]>;
- def i32rr_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a,
- Int32Regs:$b),
- !strconcat(OpcStr_u32, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int32Regs:$a, Int32Regs:$b))]>;
- def i32ri_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int32Regs:$a, i32imm:$b),
- !strconcat(OpcStr_u32, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int32Regs:$a, imm:$b))]>;
- def i32ir_u32: NVPTXInst<(outs Int32Regs:$dst), (ins i32imm:$a, Int32Regs:$b),
- !strconcat(OpcStr_u32, "32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode imm:$a, Int32Regs:$b))]>;
- def i64rr_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int64Regs:$a,
- Int64Regs:$b),
- !strconcat(OpcStr_u32, "64 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int64Regs:$a, Int64Regs:$b))]>;
- def i64ri_u32: NVPTXInst<(outs Int32Regs:$dst), (ins Int64Regs:$a, i64imm:$b),
- !strconcat(OpcStr_u32, "64 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Int64Regs:$a, imm:$b))]>;
- def i64ir_u32: NVPTXInst<(outs Int32Regs:$dst), (ins i64imm:$a, Int64Regs:$b),
- !strconcat(OpcStr_u32, "64 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode imm:$a, Int64Regs:$b))]>;
+multiclass ISET_FORMAT_UNSIGNED<PatFrag OpNode, PatLeaf Mode>
+ : ISET_FORMAT<OpNode, Mode,
+ SETP_u16rr, SETP_u16ri, SETP_u16ir,
+ SETP_u32rr, SETP_u32ri, SETP_u32ir,
+ SETP_u64rr, SETP_u64ri, SETP_u64ir,
+ SET_u16rr, SET_u16ri, SET_u16ir,
+ SET_u32rr, SET_u32ri, SET_u32ir,
+ SET_u64rr, SET_u64ri, SET_u64ir> {
+ // TableGen doesn't like empty multiclasses
+ def : PatLeaf<(i32 0)>;
}
-multiclass FSET_FORMAT<string OpcStr, string OpcStr_u32, PatFrag OpNode> {
- def f32rr_toi32_ftz: NVPTXInst<(outs Int32Regs:$dst), (ins Float32Regs:$a,
- Float32Regs:$b),
- Set_Str<OpcStr, "ftz.f32", "32", "$dst", "$a", "$b">.s,
- []>, Requires<[doF32FTZ]>;
- def f32rr_toi32: NVPTXInst<(outs Int32Regs:$dst), (ins Float32Regs:$a,
- Float32Regs:$b),
- Set_Str<OpcStr, "f32", "32", "$dst", "$a", "$b">.s,
- []>;
- def f64rr_toi64: NVPTXInst<(outs Int64Regs:$dst), (ins Float64Regs:$a,
- Float64Regs:$b),
- Set_Str<OpcStr, "f64", "64", "$dst", "$a", "$b">.s,
- []>;
- def f64rr_toi32: NVPTXInst<(outs Int32Regs:$dst), (ins Float64Regs:$a,
- Float64Regs:$b),
- Set_Str<OpcStr, "f64", "32", "$dst", "$a", "$b">.s,
- []>;
-
- def f32rr_p_ftz: NVPTXInst<(outs Int1Regs:$dst), (ins Float32Regs:$a
- , Float32Regs:$b),
- !strconcat(OpcStr, "ftz.f32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Float32Regs:$a, Float32Regs:$b))]>
- , Requires<[doF32FTZ]>;
- def f32rr_p: NVPTXInst<(outs Int1Regs:$dst),
- (ins Float32Regs:$a, Float32Regs:$b),
- !strconcat(OpcStr, "f32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Float32Regs:$a, Float32Regs:$b))]>;
- def f32ri_p_ftz: NVPTXInst<(outs Int1Regs:$dst),
- (ins Float32Regs:$a, f32imm:$b),
- !strconcat(OpcStr, "ftz.f32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Float32Regs:$a, fpimm:$b))]>,
- Requires<[doF32FTZ]>;
- def f32ri_p: NVPTXInst<(outs Int1Regs:$dst), (ins Float32Regs:$a, f32imm:$b),
- !strconcat(OpcStr, "f32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Float32Regs:$a, fpimm:$b))]>;
- def f32ir_p_ftz: NVPTXInst<(outs Int1Regs:$dst),
- (ins f32imm:$a, Float32Regs:$b),
- !strconcat(OpcStr, "ftz.f32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode fpimm:$a, Float32Regs:$b))]>,
- Requires<[doF32FTZ]>;
- def f32ir_p: NVPTXInst<(outs Int1Regs:$dst), (ins f32imm:$a, Float32Regs:$b),
- !strconcat(OpcStr, "f32 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode fpimm:$a, Float32Regs:$b))]>;
- def f64rr_p: NVPTXInst<(outs Int1Regs:$dst),
- (ins Float64Regs:$a, Float64Regs:$b),
- !strconcat(OpcStr, "f64 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Float64Regs:$a, Float64Regs:$b))]>;
- def f64ri_p: NVPTXInst<(outs Int1Regs:$dst), (ins Float64Regs:$a, f64imm:$b),
- !strconcat(OpcStr, "f64 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode Float64Regs:$a, fpimm:$b))]>;
- def f64ir_p: NVPTXInst<(outs Int1Regs:$dst), (ins f64imm:$a, Float64Regs:$b),
- !strconcat(OpcStr, "f64 \t$dst, $a, $b;"),
- [(set Int1Regs:$dst, (OpNode fpimm:$a, Float64Regs:$b))]>;
-
- def f32rr_u32_ftz: NVPTXInst<(outs Int32Regs:$dst),
- (ins Float32Regs:$a, Float32Regs:$b),
- !strconcat(OpcStr_u32, "ftz.f32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Float32Regs:$a, Float32Regs:$b))]>;
- def f32rr_u32: NVPTXInst<(outs Int32Regs:$dst),
- (ins Float32Regs:$a, Float32Regs:$b),
- !strconcat(OpcStr_u32, "f32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Float32Regs:$a, Float32Regs:$b))]>;
- def f32ri_u32_ftz: NVPTXInst<(outs Int32Regs:$dst),
- (ins Float32Regs:$a, f32imm:$b),
- !strconcat(OpcStr_u32, "ftz.f32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Float32Regs:$a, fpimm:$b))]>;
- def f32ri_u32: NVPTXInst<(outs Int32Regs:$dst),
- (ins Float32Regs:$a, f32imm:$b),
- !strconcat(OpcStr_u32, "f32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Float32Regs:$a, fpimm:$b))]>;
- def f32ir_u32_ftz: NVPTXInst<(outs Int32Regs:$dst),
- (ins f32imm:$a, Float32Regs:$b),
- !strconcat(OpcStr_u32, "ftz.f32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode fpimm:$a, Float32Regs:$b))]>;
- def f32ir_u32: NVPTXInst<(outs Int32Regs:$dst),
- (ins f32imm:$a, Float32Regs:$b),
- !strconcat(OpcStr_u32, "f32 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode fpimm:$a, Float32Regs:$b))]>;
- def f64rr_u32: NVPTXInst<(outs Int32Regs:$dst),
- (ins Float64Regs:$a, Float64Regs:$b),
- !strconcat(OpcStr_u32, "f64 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Float64Regs:$a, Float64Regs:$b))]>;
- def f64ri_u32: NVPTXInst<(outs Int32Regs:$dst),
- (ins Float64Regs:$a, f64imm:$b),
- !strconcat(OpcStr_u32, "f64 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode Float64Regs:$a, fpimm:$b))]>;
- def f64ir_u32: NVPTXInst<(outs Int32Regs:$dst),
- (ins f64imm:$a, Float64Regs:$b),
- !strconcat(OpcStr_u32, "f64 \t$dst, $a, $b;"),
- [(set Int32Regs:$dst, (OpNode fpimm:$a, Float64Regs:$b))]>;
+defm : ISET_FORMAT_SIGNED<setgt, CmpGT>;
+defm : ISET_FORMAT_UNSIGNED<setugt, CmpGT>;
+defm : ISET_FORMAT_SIGNED<setlt, CmpLT>;
+defm : ISET_FORMAT_UNSIGNED<setult, CmpLT>;
+defm : ISET_FORMAT_SIGNED<setge, CmpGE>;
+defm : ISET_FORMAT_UNSIGNED<setuge, CmpGE>;
+defm : ISET_FORMAT_SIGNED<setle, CmpLE>;
+defm : ISET_FORMAT_UNSIGNED<setule, CmpLE>;
+defm : ISET_FORMAT_SIGNED<seteq, CmpEQ>;
+defm : ISET_FORMAT_UNSIGNED<setueq, CmpEQ>;
+defm : ISET_FORMAT_SIGNED<setne, CmpNE>;
+defm : ISET_FORMAT_UNSIGNED<setune, CmpNE>;
+
+// i1 compares
+def : Pat<(setne Int1Regs:$a, Int1Regs:$b),
+ (XORb1rr Int1Regs:$a, Int1Regs:$b)>;
+def : Pat<(setune Int1Regs:$a, Int1Regs:$b),
+ (XORb1rr Int1Regs:$a, Int1Regs:$b)>;
+
+def : Pat<(seteq Int1Regs:$a, Int1Regs:$b),
+ (NOT1 (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
+def : Pat<(setueq Int1Regs:$a, Int1Regs:$b),
+ (NOT1 (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
+
+// i1 compare -> i32
+def : Pat<(i32 (setne Int1Regs:$a, Int1Regs:$b)),
+ (SELP_u32ii -1, 0, (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
+def : Pat<(i32 (setne Int1Regs:$a, Int1Regs:$b)),
+ (SELP_u32ii 0, -1, (XORb1rr Int1Regs:$a, Int1Regs:$b))>;
+
+
+
+multiclass FSET_FORMAT<PatFrag OpNode, PatLeaf Mode, PatLeaf ModeFTZ> {
+ // f32 -> pred
+ def : Pat<(i1 (OpNode Float32Regs:$a, Float32Regs:$b)),
+ (SETP_f32rr Float32Regs:$a, Float32Regs:$b, ModeFTZ)>,
+ Requires<[doF32FTZ]>;
+ def : Pat<(i1 (OpNode Float32Regs:$a, Float32Regs:$b)),
+ (SETP_f32rr Float32Regs:$a, Float32Regs:$b, Mode)>;
+ def : Pat<(i1 (OpNode Float32Regs:$a, fpimm:$b)),
+ (SETP_f32ri Float32Regs:$a, fpimm:$b, ModeFTZ)>,
+ Requires<[doF32FTZ]>;
+ def : Pat<(i1 (OpNode Float32Regs:$a, fpimm:$b)),
+ (SETP_f32ri Float32Regs:$a, fpimm:$b, Mode)>;
+ def : Pat<(i1 (OpNode fpimm:$a, Float32Regs:$b)),
+ (SETP_f32ir fpimm:$a, Float32Regs:$b, ModeFTZ)>,
+ Requires<[doF32FTZ]>;
+ def : Pat<(i1 (OpNode fpimm:$a, Float32Regs:$b)),
+ (SETP_f32ir fpimm:$a, Float32Regs:$b, Mode)>;
+
+ // f64 -> pred
+ def : Pat<(i1 (OpNode Float64Regs:$a, Float64Regs:$b)),
+ (SETP_f64rr Float64Regs:$a, Float64Regs:$b, Mode)>;
+ def : Pat<(i1 (OpNode Float64Regs:$a, fpimm:$b)),
+ (SETP_f64ri Float64Regs:$a, fpimm:$b, Mode)>;
+ def : Pat<(i1 (OpNode fpimm:$a, Float64Regs:$b)),
+ (SETP_f64ir fpimm:$a, Float64Regs:$b, Mode)>;
+
+ // f32 -> i32
+ def : Pat<(i32 (OpNode Float32Regs:$a, Float32Regs:$b)),
+ (SET_f32rr Float32Regs:$a, Float32Regs:$b, ModeFTZ)>,
+ Requires<[doF32FTZ]>;
+ def : Pat<(i32 (OpNode Float32Regs:$a, Float32Regs:$b)),
+ (SET_f32rr Float32Regs:$a, Float32Regs:$b, Mode)>;
+ def : Pat<(i32 (OpNode Float32Regs:$a, fpimm:$b)),
+ (SET_f32ri Float32Regs:$a, fpimm:$b, ModeFTZ)>,
+ Requires<[doF32FTZ]>;
+ def : Pat<(i32 (OpNode Float32Regs:$a, fpimm:$b)),
+ (SET_f32ri Float32Regs:$a, fpimm:$b, Mode)>;
+ def : Pat<(i32 (OpNode fpimm:$a, Float32Regs:$b)),
+ (SET_f32ir fpimm:$a, Float32Regs:$b, ModeFTZ)>,
+ Requires<[doF32FTZ]>;
+ def : Pat<(i32 (OpNode fpimm:$a, Float32Regs:$b)),
+ (SET_f32ir fpimm:$a, Float32Regs:$b, Mode)>;
+
+ // f64 -> i32
+ def : Pat<(i32 (OpNode Float64Regs:$a, Float64Regs:$b)),
+ (SET_f64rr Float64Regs:$a, Float64Regs:$b, Mode)>;
+ def : Pat<(i32 (OpNode Float64Regs:$a, fpimm:$b)),
+ (SET_f64ri Float64Regs:$a, fpimm:$b, Mode)>;
+ def : Pat<(i32 (OpNode fpimm:$a, Float64Regs:$b)),
+ (SET_f64ir fpimm:$a, Float64Regs:$b, Mode)>;
}
-defm ISetSGT
-: ISET_FORMAT<"setp.gt.s", "set.gt.u32.s", setgt, "s16", "cvt.s16.s8">;
-defm ISetUGT
-: ISET_FORMAT<"setp.gt.u", "set.gt.u32.u", setugt, "u16", "cvt.u16.u8">;
-defm ISetSLT
-: ISET_FORMAT<"setp.lt.s", "set.lt.u32.s", setlt, "s16", "cvt.s16.s8">;
-defm ISetULT
-: ISET_FORMAT<"setp.lt.u", "set.lt.u32.u", setult, "u16", "cvt.u16.u8">;
-defm ISetSGE
-: ISET_FORMAT<"setp.ge.s", "set.ge.u32.s", setge, "s16", "cvt.s16.s8">;
-defm ISetUGE
-: ISET_FORMAT<"setp.ge.u", "set.ge.u32.u", setuge, "u16", "cvt.u16.u8">;
-defm ISetSLE
-: ISET_FORMAT<"setp.le.s", "set.le.u32.s", setle, "s16", "cvt.s16.s8">;
-defm ISetULE
-: ISET_FORMAT<"setp.le.u", "set.le.u32.u", setule, "u16", "cvt.u16.u8">;
-defm ISetSEQ
-: ISET_FORMAT<"setp.eq.s", "set.eq.u32.s", seteq, "s16", "cvt.s16.s8">;
-defm ISetUEQ
-: ISET_FORMAT<"setp.eq.u", "set.eq.u32.u", setueq, "u16", "cvt.u16.u8">;
-defm ISetSNE
-: ISET_FORMAT<"setp.ne.s", "set.ne.u32.s", setne, "s16", "cvt.s16.s8">;
-defm ISetUNE
-: ISET_FORMAT<"setp.ne.u", "set.ne.u32.u", setune, "u16", "cvt.u16.u8">;
-
-def ISetSNEi1rr_p : NVPTXInst<(outs Int1Regs:$dst),
- (ins Int1Regs:$a, Int1Regs:$b),
- "xor.pred \t$dst, $a, $b;",
- [(set Int1Regs:$dst, (setne Int1Regs:$a, Int1Regs:$b))]>;
-def ISetUNEi1rr_p : NVPTXInst<(outs Int1Regs:$dst),
- (ins Int1Regs:$a, Int1Regs:$b),
- "xor.pred \t$dst, $a, $b;",
- [(set Int1Regs:$dst, (setune Int1Regs:$a, Int1Regs:$b))]>;
-def ISetSEQi1rr_p : NVPTXInst<(outs Int1Regs:$dst),
- (ins Int1Regs:$a, Int1Regs:$b),
- !strconcat("{{\n\t",
- !strconcat(".reg .pred temp;\n\t",
- !strconcat("xor.pred \ttemp, $a, $b;\n\t",
- !strconcat("not.pred \t$dst, temp;\n\t}}","")))),
- [(set Int1Regs:$dst, (seteq Int1Regs:$a, Int1Regs:$b))]>;
-def ISetUEQi1rr_p : NVPTXInst<(outs Int1Regs:$dst),
- (ins Int1Regs:$a, Int1Regs:$b),
- !strconcat("{{\n\t",
- !strconcat(".reg .pred temp;\n\t",
- !strconcat("xor.pred \ttemp, $a, $b;\n\t",
- !strconcat("not.pred \t$dst, temp;\n\t}}","")))),
- [(set Int1Regs:$dst, (setueq Int1Regs:$a, Int1Regs:$b))]>;
-
-// Compare 2 i1's and produce a u32
-def ISETSNEi1rr_u32 : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int1Regs:$a, Int1Regs:$b),
- !strconcat("{{\n\t",
- !strconcat(".reg .pred temp;\n\t",
- !strconcat("xor.pred \ttemp, $a, $b;\n\t",
- !strconcat("selp.u32 \t$dst, -1, 0, temp;", "\n\t}}")))),
- [(set Int32Regs:$dst, (setne Int1Regs:$a, Int1Regs:$b))]>;
-def ISETSEQi1rr_u32 : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int1Regs:$a, Int1Regs:$b),
- !strconcat("{{\n\t",
- !strconcat(".reg .pred temp;\n\t",
- !strconcat("xor.pred \ttemp, $a, $b;\n\t",
- !strconcat("selp.u32 \t$dst, 0, -1, temp;", "\n\t}}")))),
- [(set Int32Regs:$dst, (seteq Int1Regs:$a, Int1Regs:$b))]>;
-
-defm FSetGT : FSET_FORMAT<"setp.gt.", "set.gt.u32.", setogt>;
-defm FSetLT : FSET_FORMAT<"setp.lt.", "set.lt.u32.", setolt>;
-defm FSetGE : FSET_FORMAT<"setp.ge.", "set.ge.u32.", setoge>;
-defm FSetLE : FSET_FORMAT<"setp.le.", "set.le.u32.", setole>;
-defm FSetEQ : FSET_FORMAT<"setp.eq.", "set.eq.u32.", setoeq>;
-defm FSetNE : FSET_FORMAT<"setp.ne.", "set.ne.u32.", setone>;
-
-defm FSetUGT : FSET_FORMAT<"setp.gtu.", "set.gtu.u32.", setugt>;
-defm FSetULT : FSET_FORMAT<"setp.ltu.", "set.ltu.u32.",setult>;
-defm FSetUGE : FSET_FORMAT<"setp.geu.", "set.geu.u32.",setuge>;
-defm FSetULE : FSET_FORMAT<"setp.leu.", "set.leu.u32.",setule>;
-defm FSetUEQ : FSET_FORMAT<"setp.equ.", "set.equ.u32.",setueq>;
-defm FSetUNE : FSET_FORMAT<"setp.neu.", "set.neu.u32.",setune>;
-
-defm FSetNUM : FSET_FORMAT<"setp.num.", "set.num.u32.",seto>;
-defm FSetNAN : FSET_FORMAT<"setp.nan.", "set.nan.u32.",setuo>;
-
-def SELECTi1rr : Pat<(i1 (select Int1Regs:$p, Int1Regs:$a, Int1Regs:$b)),
- (ORb1rr (ANDb1rr Int1Regs:$p, Int1Regs:$a),
- (ANDb1rr (NOT1 Int1Regs:$p), Int1Regs:$b))>;
-def SELECTi8rr : NVPTXInst<(outs Int8Regs:$dst),
- (ins Int8Regs:$a, Int8Regs:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int8Regs:$dst, (select Int1Regs:$p, Int8Regs:$a, Int8Regs:$b))]>;
-def SELECTi8ri : NVPTXInst<(outs Int8Regs:$dst),
- (ins Int8Regs:$a, i8imm:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int8Regs:$dst, (select Int1Regs:$p, Int8Regs:$a, imm:$b))]>;
-def SELECTi8ir : NVPTXInst<(outs Int8Regs:$dst),
- (ins i8imm:$a, Int8Regs:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int8Regs:$dst, (select Int1Regs:$p, imm:$a, Int8Regs:$b))]>;
-def SELECTi8ii : NVPTXInst<(outs Int8Regs:$dst),
- (ins i8imm:$a, i8imm:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int8Regs:$dst, (select Int1Regs:$p, imm:$a, imm:$b))]>;
-
-def SELECTi16rr : NVPTXInst<(outs Int16Regs:$dst),
- (ins Int16Regs:$a, Int16Regs:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int16Regs:$dst, (select Int1Regs:$p, Int16Regs:$a, Int16Regs:$b))]>;
-def SELECTi16ri : NVPTXInst<(outs Int16Regs:$dst),
- (ins Int16Regs:$a, i16imm:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int16Regs:$dst, (select Int1Regs:$p, Int16Regs:$a, imm:$b))]>;
-def SELECTi16ir : NVPTXInst<(outs Int16Regs:$dst),
- (ins i16imm:$a, Int16Regs:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int16Regs:$dst, (select Int1Regs:$p, imm:$a, Int16Regs:$b))]>;
-def SELECTi16ii : NVPTXInst<(outs Int16Regs:$dst),
- (ins i16imm:$a, i16imm:$b, Int1Regs:$p),
- "selp.b16 \t$dst, $a, $b, $p;",
- [(set Int16Regs:$dst, (select Int1Regs:$p, imm:$a, imm:$b))]>;
-
-def SELECTi32rr : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int32Regs:$a, Int32Regs:$b, Int1Regs:$p),
- "selp.b32 \t$dst, $a, $b, $p;",
- [(set Int32Regs:$dst, (select Int1Regs:$p, Int32Regs:$a, Int32Regs:$b))]>;
-def SELECTi32ri : NVPTXInst<(outs Int32Regs:$dst),
- (ins Int32Regs:$a, i32imm:$b, Int1Regs:$p),
- "selp.b32 \t$dst, $a, $b, $p;",
- [(set Int32Regs:$dst, (select Int1Regs:$p, Int32Regs:$a, imm:$b))]>;
-def SELECTi32ir : NVPTXInst<(outs Int32Regs:$dst),
- (ins i32imm:$a, Int32Regs:$b, Int1Regs:$p),
- "selp.b32 \t$dst, $a, $b, $p;",
- [(set Int32Regs:$dst, (select Int1Regs:$p, imm:$a, Int32Regs:$b))]>;
-def SELECTi32ii : NVPTXInst<(outs Int32Regs:$dst),
- (ins i32imm:$a, i32imm:$b, Int1Regs:$p),
- "selp.b32 \t$dst, $a, $b, $p;",
- [(set Int32Regs:$dst, (select Int1Regs:$p, imm:$a, imm:$b))]>;
-
-def SELECTi64rr : NVPTXInst<(outs Int64Regs:$dst),
- (ins Int64Regs:$a, Int64Regs:$b, Int1Regs:$p),
- "selp.b64 \t$dst, $a, $b, $p;",
- [(set Int64Regs:$dst, (select Int1Regs:$p, Int64Regs:$a, Int64Regs:$b))]>;
-def SELECTi64ri : NVPTXInst<(outs Int64Regs:$dst),
- (ins Int64Regs:$a, i64imm:$b, Int1Regs:$p),
- "selp.b64 \t$dst, $a, $b, $p;",
- [(set Int64Regs:$dst, (select Int1Regs:$p, Int64Regs:$a, imm:$b))]>;
-def SELECTi64ir : NVPTXInst<(outs Int64Regs:$dst),
- (ins i64imm:$a, Int64Regs:$b, Int1Regs:$p),
- "selp.b64 \t$dst, $a, $b, $p;",
- [(set Int64Regs:$dst, (select Int1Regs:$p, imm:$a, Int64Regs:$b))]>;
-def SELECTi64ii : NVPTXInst<(outs Int64Regs:$dst),
- (ins i64imm:$a, i64imm:$b, Int1Regs:$p),
- "selp.b64 \t$dst, $a, $b, $p;",
- [(set Int64Regs:$dst, (select Int1Regs:$p, imm:$a, imm:$b))]>;
-
-def SELECTf32rr : NVPTXInst<(outs Float32Regs:$dst),
- (ins Float32Regs:$a, Float32Regs:$b, Int1Regs:$p),
- "selp.f32 \t$dst, $a, $b, $p;",
- [(set Float32Regs:$dst,
- (select Int1Regs:$p, Float32Regs:$a, Float32Regs:$b))]>;
-def SELECTf32ri : NVPTXInst<(outs Float32Regs:$dst),
- (ins Float32Regs:$a, f32imm:$b, Int1Regs:$p),
- "selp.f32 \t$dst, $a, $b, $p;",
- [(set Float32Regs:$dst, (select Int1Regs:$p, Float32Regs:$a, fpimm:$b))]>;
-def SELECTf32ir : NVPTXInst<(outs Float32Regs:$dst),
- (ins f32imm:$a, Float32Regs:$b, Int1Regs:$p),
- "selp.f32 \t$dst, $a, $b, $p;",
- [(set Float32Regs:$dst, (select Int1Regs:$p, fpimm:$a, Float32Regs:$b))]>;
-def SELECTf32ii : NVPTXInst<(outs Float32Regs:$dst),
- (ins f32imm:$a, f32imm:$b, Int1Regs:$p),
- "selp.f32 \t$dst, $a, $b, $p;",
- [(set Float32Regs:$dst, (select Int1Regs:$p, fpimm:$a, fpimm:$b))]>;
-
-def SELECTf64rr : NVPTXInst<(outs Float64Regs:$dst),
- (ins Float64Regs:$a, Float64Regs:$b, Int1Regs:$p),
- "selp.f64 \t$dst, $a, $b, $p;",
- [(set Float64Regs:$dst,
- (select Int1Regs:$p, Float64Regs:$a, Float64Regs:$b))]>;
-def SELECTf64ri : NVPTXInst<(outs Float64Regs:$dst),
- (ins Float64Regs:$a, f64imm:$b, Int1Regs:$p),
- "selp.f64 \t$dst, $a, $b, $p;",
- [(set Float64Regs:$dst, (select Int1Regs:$p, Float64Regs:$a, fpimm:$b))]>;
-def SELECTf64ir : NVPTXInst<(outs Float64Regs:$dst),
- (ins f64imm:$a, Float64Regs:$b, Int1Regs:$p),
- "selp.f64 \t$dst, $a, $b, $p;",
- [(set Float64Regs:$dst, (select Int1Regs:$p, fpimm:$a, Float64Regs:$b))]>;
-def SELECTf64ii : NVPTXInst<(outs Float64Regs:$dst),
- (ins f64imm:$a, f64imm:$b, Int1Regs:$p),
- "selp.f64 \t $dst, $a, $b, $p;",
- [(set Float64Regs:$dst, (select Int1Regs:$p, fpimm:$a, fpimm:$b))]>;
+defm FSetOGT : FSET_FORMAT<setogt, CmpGT, CmpGT_FTZ>;
+defm FSetOLT : FSET_FORMAT<setolt, CmpLT, CmpLT_FTZ>;
+defm FSetOGE : FSET_FORMAT<setoge, CmpGE, CmpGE_FTZ>;
+defm FSetOLE : FSET_FORMAT<setole, CmpLE, CmpLE_FTZ>;
+defm FSetOEQ : FSET_FORMAT<setoeq, CmpEQ, CmpEQ_FTZ>;
+defm FSetONE : FSET_FORMAT<setone, CmpNE, CmpNE_FTZ>;
+
+defm FSetUGT : FSET_FORMAT<setugt, CmpGTU, CmpGTU_FTZ>;
+defm FSetULT : FSET_FORMAT<setult, CmpLTU, CmpLTU_FTZ>;
+defm FSetUGE : FSET_FORMAT<setuge, CmpGEU, CmpGEU_FTZ>;
+defm FSetULE : FSET_FORMAT<setule, CmpLEU, CmpLEU_FTZ>;
+defm FSetUEQ : FSET_FORMAT<setueq, CmpEQU, CmpEQU_FTZ>;
+defm FSetUNE : FSET_FORMAT<setune, CmpNEU, CmpNEU_FTZ>;
+
+defm FSetGT : FSET_FORMAT<setgt, CmpGT, CmpGT_FTZ>;
+defm FSetLT : FSET_FORMAT<setlt, CmpLT, CmpLT_FTZ>;
+defm FSetGE : FSET_FORMAT<setge, CmpGE, CmpGE_FTZ>;
+defm FSetLE : FSET_FORMAT<setle, CmpLE, CmpLE_FTZ>;
+defm FSetEQ : FSET_FORMAT<seteq, CmpEQ, CmpEQ_FTZ>;
+defm FSetNE : FSET_FORMAT<setne, CmpNE, CmpNE_FTZ>;
+
+defm FSetNUM : FSET_FORMAT<seto, CmpNUM, CmpNUM_FTZ>;
+defm FSetNAN : FSET_FORMAT<setuo, CmpNAN, CmpNAN_FTZ>;
//def ld_param : SDNode<"NVPTXISD::LOAD_PARAM", SDTLoad,
// [SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
def SDTDeclareScalarParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>,
SDTCisInt<1>, SDTCisInt<2>]>;
def SDTLoadParamProfile : SDTypeProfile<1, 2, [SDTCisInt<1>, SDTCisInt<2>]>;
+def SDTLoadParamV2Profile : SDTypeProfile<2, 2, [SDTCisSameAs<0, 1>, SDTCisInt<2>, SDTCisInt<3>]>;
+def SDTLoadParamV4Profile : SDTypeProfile<4, 2, [SDTCisInt<4>, SDTCisInt<5>]>;
def SDTPrintCallProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
def SDTPrintCallUniProfile : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
def SDTStoreParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
+def SDTStoreParamV2Profile : SDTypeProfile<0, 4, [SDTCisInt<0>, SDTCisInt<1>]>;
+def SDTStoreParamV4Profile : SDTypeProfile<0, 6, [SDTCisInt<0>, SDTCisInt<1>]>;
def SDTStoreParam32Profile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>]>;
def SDTCallArgProfile : SDTypeProfile<0, 2, [SDTCisInt<0>]>;
def SDTCallArgMarkProfile : SDTypeProfile<0, 0, []>;
def SDTCallVoidProfile : SDTypeProfile<0, 1, []>;
def SDTCallValProfile : SDTypeProfile<1, 0, []>;
def SDTMoveParamProfile : SDTypeProfile<1, 1, []>;
-def SDTMoveRetvalProfile : SDTypeProfile<0, 1, []>;
def SDTStoreRetvalProfile : SDTypeProfile<0, 2, [SDTCisInt<0>]>;
+def SDTStoreRetvalV2Profile : SDTypeProfile<0, 3, [SDTCisInt<0>]>;
+def SDTStoreRetvalV4Profile : SDTypeProfile<0, 5, [SDTCisInt<0>]>;
def SDTPseudoUseParamProfile : SDTypeProfile<0, 1, []>;
def DeclareParam : SDNode<"NVPTXISD::DeclareParam", SDTDeclareParamProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def LoadParam : SDNode<"NVPTXISD::LoadParam", SDTLoadParamProfile,
[SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
+def LoadParamV2 : SDNode<"NVPTXISD::LoadParamV2", SDTLoadParamV2Profile,
+ [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
+def LoadParamV4 : SDNode<"NVPTXISD::LoadParamV4", SDTLoadParamV4Profile,
+ [SDNPHasChain, SDNPMayLoad, SDNPOutGlue, SDNPInGlue]>;
def PrintCall : SDNode<"NVPTXISD::PrintCall", SDTPrintCallProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def PrintCallUni : SDNode<"NVPTXISD::PrintCallUni", SDTPrintCallUniProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def StoreParam : SDNode<"NVPTXISD::StoreParam", SDTStoreParamProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
+def StoreParamV2 : SDNode<"NVPTXISD::StoreParamV2", SDTStoreParamV2Profile,
+ [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
+def StoreParamV4 : SDNode<"NVPTXISD::StoreParamV4", SDTStoreParamV4Profile,
+ [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def StoreParamU32 : SDNode<"NVPTXISD::StoreParamU32", SDTStoreParam32Profile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def StoreParamS32 : SDNode<"NVPTXISD::StoreParamS32", SDTStoreParam32Profile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
-def MoveToParam : SDNode<"NVPTXISD::MoveToParam", SDTStoreParamProfile,
- [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def CallArgBegin : SDNode<"NVPTXISD::CallArgBegin", SDTCallArgMarkProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def CallArg : SDNode<"NVPTXISD::CallArg", SDTCallArgProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
def MoveParam : SDNode<"NVPTXISD::MoveParam", SDTMoveParamProfile,
[]>;
-def MoveRetval : SDNode<"NVPTXISD::MoveRetval", SDTMoveRetvalProfile,
- [SDNPHasChain, SDNPSideEffect]>;
def StoreRetval : SDNode<"NVPTXISD::StoreRetval", SDTStoreRetvalProfile,
[SDNPHasChain, SDNPSideEffect]>;
-def MoveToRetval : SDNode<"NVPTXISD::MoveToRetval", SDTStoreRetvalProfile,
- [SDNPHasChain, SDNPSideEffect]>;
+def StoreRetvalV2 : SDNode<"NVPTXISD::StoreRetvalV2", SDTStoreRetvalV2Profile,
+ [SDNPHasChain, SDNPSideEffect]>;
+def StoreRetvalV4 : SDNode<"NVPTXISD::StoreRetvalV4", SDTStoreRetvalV4Profile,
+ [SDNPHasChain, SDNPSideEffect]>;
def PseudoUseParam : SDNode<"NVPTXISD::PseudoUseParam",
SDTPseudoUseParamProfile,
[SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
NVPTXInst<(outs regclass:$dst), (ins i32imm:$b),
!strconcat(!strconcat("ld.param", opstr),
"\t$dst, [retval0+$b];"),
- [(set regclass:$dst, (LoadParam (i32 1), (i32 imm:$b)))]>;
+ []>;
class LoadParamRegInst<NVPTXRegClass regclass, string opstr> :
NVPTXInst<(outs regclass:$dst), (ins i32imm:$b),
"\t$dst, retval$b;"),
[(set regclass:$dst, (LoadParam (i32 0), (i32 imm:$b)))]>;
+class LoadParamV2MemInst<NVPTXRegClass regclass, string opstr> :
+ NVPTXInst<(outs regclass:$dst, regclass:$dst2), (ins i32imm:$b),
+ !strconcat(!strconcat("ld.param.v2", opstr),
+ "\t{{$dst, $dst2}}, [retval0+$b];"), []>;
+
+class LoadParamV4MemInst<NVPTXRegClass regclass, string opstr> :
+ NVPTXInst<(outs regclass:$dst, regclass:$dst2, regclass:$dst3,
+ regclass:$dst4),
+ (ins i32imm:$b),
+ !strconcat(!strconcat("ld.param.v4", opstr),
+ "\t{{$dst, $dst2, $dst3, $dst4}}, [retval0+$b];"), []>;
+
class StoreParamInst<NVPTXRegClass regclass, string opstr> :
NVPTXInst<(outs), (ins regclass:$val, i32imm:$a, i32imm:$b),
!strconcat(!strconcat("st.param", opstr),
"\t[param$a+$b], $val;"),
- [(StoreParam (i32 imm:$a), (i32 imm:$b), regclass:$val)]>;
+ []>;
-class MoveToParamInst<NVPTXRegClass regclass, string opstr> :
- NVPTXInst<(outs), (ins regclass:$val, i32imm:$a, i32imm:$b),
- !strconcat(!strconcat("mov", opstr),
- "\tparam$a, $val;"),
- [(MoveToParam (i32 imm:$a), (i32 imm:$b), regclass:$val)]>;
+class StoreParamV2Inst<NVPTXRegClass regclass, string opstr> :
+ NVPTXInst<(outs), (ins regclass:$val, regclass:$val2,
+ i32imm:$a, i32imm:$b),
+ !strconcat(!strconcat("st.param.v2", opstr),
+ "\t[param$a+$b], {{$val, $val2}};"),
+ []>;
+
+class StoreParamV4Inst<NVPTXRegClass regclass, string opstr> :
+ NVPTXInst<(outs), (ins regclass:$val, regclass:$val1, regclass:$val2,
+ regclass:$val3, i32imm:$a, i32imm:$b),
+ !strconcat(!strconcat("st.param.v4", opstr),
+ "\t[param$a+$b], {{$val, $val2, $val3, $val4}};"),
+ []>;
class StoreRetvalInst<NVPTXRegClass regclass, string opstr> :
NVPTXInst<(outs), (ins regclass:$val, i32imm:$a),
!strconcat(!strconcat("st.param", opstr),
"\t[func_retval0+$a], $val;"),
- [(StoreRetval (i32 imm:$a), regclass:$val)]>;
+ []>;
-class MoveToRetvalInst<NVPTXRegClass regclass, string opstr> :
- NVPTXInst<(outs), (ins i32imm:$num, regclass:$val),
- !strconcat(!strconcat("mov", opstr),
- "\tfunc_retval$num, $val;"),
- [(MoveToRetval (i32 imm:$num), regclass:$val)]>;
+class StoreRetvalV2Inst<NVPTXRegClass regclass, string opstr> :
+ NVPTXInst<(outs), (ins regclass:$val, regclass:$val2, i32imm:$a),
+ !strconcat(!strconcat("st.param.v2", opstr),
+ "\t[func_retval0+$a], {{$val, $val2}};"),
+ []>;
-class MoveRetvalInst<NVPTXRegClass regclass, string opstr> :
- NVPTXInst<(outs), (ins regclass:$val),
- !strconcat(!strconcat("mov", opstr),
- "\tfunc_retval0, $val;"),
- [(MoveRetval regclass:$val)]>;
+class StoreRetvalV4Inst<NVPTXRegClass regclass, string opstr> :
+ NVPTXInst<(outs),
+ (ins regclass:$val, regclass:$val2, regclass:$val3,
+ regclass:$val4, i32imm:$a),
+ !strconcat(!strconcat("st.param.v4", opstr),
+ "\t[func_retval0+$a], {{$val, $val2, $val3, $val4}};"),
+ []>;
def PrintCallRetInst1 : NVPTXInst<(outs), (ins),
"call (retval0), ",
def LoadParamMemI64 : LoadParamMemInst<Int64Regs, ".b64">;
def LoadParamMemI32 : LoadParamMemInst<Int32Regs, ".b32">;
def LoadParamMemI16 : LoadParamMemInst<Int16Regs, ".b16">;
-def LoadParamMemI8 : LoadParamMemInst<Int8Regs, ".b8">;
-
-//def LoadParamMemI16 : NVPTXInst<(outs Int16Regs:$dst), (ins i32imm:$b),
-// !strconcat("ld.param.b32\ttemp_param_reg, [retval0+$b];\n\t",
-// "cvt.u16.u32\t$dst, temp_param_reg;"),
-// [(set Int16Regs:$dst, (LoadParam (i32 1), (i32 imm:$b)))]>;
-//def LoadParamMemI8 : NVPTXInst<(outs Int8Regs:$dst), (ins i32imm:$b),
-// !strconcat("ld.param.b32\ttemp_param_reg, [retval0+$b];\n\t",
-// "cvt.u16.u32\t$dst, temp_param_reg;"),
-// [(set Int8Regs:$dst, (LoadParam (i32 1), (i32 imm:$b)))]>;
-
+def LoadParamMemI8 : LoadParamMemInst<Int16Regs, ".b8">;
+def LoadParamMemV2I64 : LoadParamV2MemInst<Int64Regs, ".b64">;
+def LoadParamMemV2I32 : LoadParamV2MemInst<Int32Regs, ".b32">;
+def LoadParamMemV2I16 : LoadParamV2MemInst<Int16Regs, ".b16">;
+def LoadParamMemV2I8 : LoadParamV2MemInst<Int16Regs, ".b8">;
+def LoadParamMemV4I32 : LoadParamV4MemInst<Int32Regs, ".b32">;
+def LoadParamMemV4I16 : LoadParamV4MemInst<Int16Regs, ".b16">;
+def LoadParamMemV4I8 : LoadParamV4MemInst<Int16Regs, ".b8">;
def LoadParamMemF32 : LoadParamMemInst<Float32Regs, ".f32">;
def LoadParamMemF64 : LoadParamMemInst<Float64Regs, ".f64">;
-
-def LoadParamRegI64 : LoadParamRegInst<Int64Regs, ".b64">;
-def LoadParamRegI32 : LoadParamRegInst<Int32Regs, ".b32">;
-def LoadParamRegI16 : NVPTXInst<(outs Int16Regs:$dst), (ins i32imm:$b),
- "cvt.u16.u32\t$dst, retval$b;",
- [(set Int16Regs:$dst,
- (LoadParam (i32 0), (i32 imm:$b)))]>;
-def LoadParamRegI8 : NVPTXInst<(outs Int8Regs:$dst), (ins i32imm:$b),
- "cvt.u16.u32\t$dst, retval$b;",
- [(set Int8Regs:$dst,
- (LoadParam (i32 0), (i32 imm:$b)))]>;
-
-def LoadParamRegF32 : LoadParamRegInst<Float32Regs, ".f32">;
-def LoadParamRegF64 : LoadParamRegInst<Float64Regs, ".f64">;
+def LoadParamMemV2F32 : LoadParamV2MemInst<Float32Regs, ".f32">;
+def LoadParamMemV2F64 : LoadParamV2MemInst<Float64Regs, ".f64">;
+def LoadParamMemV4F32 : LoadParamV4MemInst<Float32Regs, ".f32">;
def StoreParamI64 : StoreParamInst<Int64Regs, ".b64">;
def StoreParamI32 : StoreParamInst<Int32Regs, ".b32">;
-def StoreParamI16 : NVPTXInst<(outs),
- (ins Int16Regs:$val, i32imm:$a, i32imm:$b),
- "st.param.b16\t[param$a+$b], $val;",
- [(StoreParam (i32 imm:$a), (i32 imm:$b), Int16Regs:$val)]>;
-
-def StoreParamI8 : NVPTXInst<(outs),
- (ins Int8Regs:$val, i32imm:$a, i32imm:$b),
- "st.param.b8\t[param$a+$b], $val;",
- [(StoreParam
- (i32 imm:$a), (i32 imm:$b), Int8Regs:$val)]>;
-
-def StoreParamS32I16 : NVPTXInst<(outs),
- (ins Int16Regs:$val, i32imm:$a, i32imm:$b),
- !strconcat("cvt.s32.s16\ttemp_param_reg, $val;\n\t",
- "st.param.b32\t[param$a+$b], temp_param_reg;"),
- [(StoreParamS32 (i32 imm:$a), (i32 imm:$b), Int16Regs:$val)]>;
-def StoreParamU32I16 : NVPTXInst<(outs),
- (ins Int16Regs:$val, i32imm:$a, i32imm:$b),
- !strconcat("cvt.u32.u16\ttemp_param_reg, $val;\n\t",
- "st.param.b32\t[param$a+$b], temp_param_reg;"),
- [(StoreParamU32 (i32 imm:$a), (i32 imm:$b), Int16Regs:$val)]>;
-
-def StoreParamU32I8 : NVPTXInst<(outs),
- (ins Int8Regs:$val, i32imm:$a, i32imm:$b),
- !strconcat("cvt.u32.u8\ttemp_param_reg, $val;\n\t",
- "st.param.b32\t[param$a+$b], temp_param_reg;"),
- [(StoreParamU32 (i32 imm:$a), (i32 imm:$b), Int8Regs:$val)]>;
-def StoreParamS32I8 : NVPTXInst<(outs),
- (ins Int8Regs:$val, i32imm:$a, i32imm:$b),
- !strconcat("cvt.s32.s8\ttemp_param_reg, $val;\n\t",
- "st.param.b32\t[param$a+$b], temp_param_reg;"),
- [(StoreParamS32 (i32 imm:$a), (i32 imm:$b), Int8Regs:$val)]>;
+def StoreParamI16 : StoreParamInst<Int16Regs, ".b16">;
+def StoreParamI8 : StoreParamInst<Int16Regs, ".b8">;
+def StoreParamV2I64 : StoreParamV2Inst<Int64Regs, ".b64">;
+def StoreParamV2I32 : StoreParamV2Inst<Int32Regs, ".b32">;
+def StoreParamV2I16 : StoreParamV2Inst<Int16Regs, ".b16">;
+def StoreParamV2I8 : StoreParamV2Inst<Int16Regs, ".b8">;
+
+// FIXME: StoreParamV4Inst crashes llvm-tblgen :(
+//def StoreParamV4I32 : StoreParamV4Inst<Int32Regs, ".b32">;
+def StoreParamV4I32 : NVPTXInst<(outs), (ins Int32Regs:$val, Int32Regs:$val2,
+ Int32Regs:$val3, Int32Regs:$val4,
+ i32imm:$a, i32imm:$b),
+ "st.param.v4.b32\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
+ []>;
+
+def StoreParamV4I16 : NVPTXInst<(outs), (ins Int16Regs:$val, Int16Regs:$val2,
+ Int16Regs:$val3, Int16Regs:$val4,
+ i32imm:$a, i32imm:$b),
+ "st.param.v4.b16\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
+ []>;
+
+def StoreParamV4I8 : NVPTXInst<(outs), (ins Int16Regs:$val, Int16Regs:$val2,
+ Int16Regs:$val3, Int16Regs:$val4,
+ i32imm:$a, i32imm:$b),
+ "st.param.v4.b8\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
+ []>;
def StoreParamF32 : StoreParamInst<Float32Regs, ".f32">;
def StoreParamF64 : StoreParamInst<Float64Regs, ".f64">;
+def StoreParamV2F32 : StoreParamV2Inst<Float32Regs, ".f32">;
+def StoreParamV2F64 : StoreParamV2Inst<Float64Regs, ".f64">;
+// FIXME: StoreParamV4Inst crashes llvm-tblgen :(
+//def StoreParamV4F32 : StoreParamV4Inst<Float32Regs, ".f32">;
+def StoreParamV4F32 : NVPTXInst<(outs),
+ (ins Float32Regs:$val, Float32Regs:$val2,
+ Float32Regs:$val3, Float32Regs:$val4,
+ i32imm:$a, i32imm:$b),
+ "st.param.v4.f32\t[param$a+$b], {{$val, $val2, $val3, $val4}};",
+ []>;
-def MoveToParamI64 : MoveToParamInst<Int64Regs, ".b64">;
-def MoveToParamI32 : MoveToParamInst<Int32Regs, ".b32">;
-def MoveToParamF64 : MoveToParamInst<Float64Regs, ".f64">;
-def MoveToParamF32 : MoveToParamInst<Float32Regs, ".f32">;
-def MoveToParamI16 : NVPTXInst<(outs),
- (ins Int16Regs:$val, i32imm:$a, i32imm:$b),
- !strconcat("cvt.u32.u16\ttemp_param_reg, $val;\n\t",
- "mov.b32\tparam$a, temp_param_reg;"),
- [(MoveToParam (i32 imm:$a), (i32 imm:$b), Int16Regs:$val)]>;
-def MoveToParamI8 : NVPTXInst<(outs),
- (ins Int8Regs:$val, i32imm:$a, i32imm:$b),
- !strconcat("cvt.u32.u16\ttemp_param_reg, $val;\n\t",
- "mov.b32\tparam$a, temp_param_reg;"),
- [(MoveToParam (i32 imm:$a), (i32 imm:$b), Int8Regs:$val)]>;
def StoreRetvalI64 : StoreRetvalInst<Int64Regs, ".b64">;
def StoreRetvalI32 : StoreRetvalInst<Int32Regs, ".b32">;
def StoreRetvalI16 : StoreRetvalInst<Int16Regs, ".b16">;
-def StoreRetvalI8 : StoreRetvalInst<Int8Regs, ".b8">;
-
-//def StoreRetvalI16 : NVPTXInst<(outs), (ins Int16Regs:$val, i32imm:$a),
-// !strconcat("\{\n\t",
-// !strconcat(".reg .b32 temp_retval_reg;\n\t",
-// !strconcat("cvt.u32.u16\ttemp_retval_reg, $val;\n\t",
-// "st.param.b32\t[func_retval0+$a], temp_retval_reg;\n\t\}"))),
-// [(StoreRetval (i32 imm:$a), Int16Regs:$val)]>;
-//def StoreRetvalI8 : NVPTXInst<(outs), (ins Int8Regs:$val, i32imm:$a),
-// !strconcat("\{\n\t",
-// !strconcat(".reg .b32 temp_retval_reg;\n\t",
-// !strconcat("cvt.u32.u16\ttemp_retval_reg, $val;\n\t",
-// "st.param.b32\t[func_retval0+$a], temp_retval_reg;\n\t\}"))),
-// [(StoreRetval (i32 imm:$a), Int8Regs:$val)]>;
+def StoreRetvalI8 : StoreRetvalInst<Int16Regs, ".b8">;
+def StoreRetvalV2I64 : StoreRetvalV2Inst<Int64Regs, ".b64">;
+def StoreRetvalV2I32 : StoreRetvalV2Inst<Int32Regs, ".b32">;
+def StoreRetvalV2I16 : StoreRetvalV2Inst<Int16Regs, ".b16">;
+def StoreRetvalV2I8 : StoreRetvalV2Inst<Int16Regs, ".b8">;
+def StoreRetvalV4I32 : StoreRetvalV4Inst<Int32Regs, ".b32">;
+def StoreRetvalV4I16 : StoreRetvalV4Inst<Int16Regs, ".b16">;
+def StoreRetvalV4I8 : StoreRetvalV4Inst<Int16Regs, ".b8">;
def StoreRetvalF64 : StoreRetvalInst<Float64Regs, ".f64">;
def StoreRetvalF32 : StoreRetvalInst<Float32Regs, ".f32">;
-
-def MoveRetvalI64 : MoveRetvalInst<Int64Regs, ".b64">;
-def MoveRetvalI32 : MoveRetvalInst<Int32Regs, ".b32">;
-def MoveRetvalI16 : MoveRetvalInst<Int16Regs, ".b16">;
-def MoveRetvalI8 : MoveRetvalInst<Int8Regs, ".b8">;
-def MoveRetvalF64 : MoveRetvalInst<Float64Regs, ".f64">;
-def MoveRetvalF32 : MoveRetvalInst<Float32Regs, ".f32">;
-
-def MoveToRetvalI64 : MoveToRetvalInst<Int64Regs, ".b64">;
-def MoveToRetvalI32 : MoveToRetvalInst<Int32Regs, ".b32">;
-def MoveToRetvalF64 : MoveToRetvalInst<Float64Regs, ".f64">;
-def MoveToRetvalF32 : MoveToRetvalInst<Float32Regs, ".f32">;
-def MoveToRetvalI16 : NVPTXInst<(outs), (ins i32imm:$num, Int16Regs:$val),
- "cvt.u32.u16\tfunc_retval$num, $val;",
- [(MoveToRetval (i32 imm:$num), Int16Regs:$val)]>;
-def MoveToRetvalI8 : NVPTXInst<(outs), (ins i32imm:$num, Int8Regs:$val),
- "cvt.u32.u16\tfunc_retval$num, $val;",
- [(MoveToRetval (i32 imm:$num), Int8Regs:$val)]>;
+def StoreRetvalV2F64 : StoreRetvalV2Inst<Float64Regs, ".f64">;
+def StoreRetvalV2F32 : StoreRetvalV2Inst<Float32Regs, ".f32">;
+def StoreRetvalV4F32 : StoreRetvalV4Inst<Float32Regs, ".f32">;
def CallArgBeginInst : NVPTXInst<(outs), (ins), "(", [(CallArgBegin)]>;
def CallArgEndInst1 : NVPTXInst<(outs), (ins), ");", [(CallArgEnd (i32 1))]>;
def CallArgI64 : CallArgInst<Int64Regs>;
def CallArgI32 : CallArgInst<Int32Regs>;
def CallArgI16 : CallArgInst<Int16Regs>;
-def CallArgI8 : CallArgInst<Int8Regs>;
def CallArgF64 : CallArgInst<Float64Regs>;
def CallArgF32 : CallArgInst<Float32Regs>;
def LastCallArgI64 : LastCallArgInst<Int64Regs>;
def LastCallArgI32 : LastCallArgInst<Int32Regs>;
def LastCallArgI16 : LastCallArgInst<Int16Regs>;
-def LastCallArgI8 : LastCallArgInst<Int8Regs>;
def LastCallArgF64 : LastCallArgInst<Float64Regs>;
def LastCallArgF32 : LastCallArgInst<Float32Regs>;
def MoveParamI16 : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$src),
"cvt.u16.u32\t$dst, $src;",
[(set Int16Regs:$dst, (MoveParam Int16Regs:$src))]>;
-def MoveParamI8 : NVPTXInst<(outs Int8Regs:$dst), (ins Int8Regs:$src),
- "cvt.u16.u32\t$dst, $src;",
- [(set Int8Regs:$dst, (MoveParam Int8Regs:$src))]>;
def MoveParamF64 : MoveParamInst<Float64Regs, ".f64">;
def MoveParamF32 : MoveParamInst<Float32Regs, ".f32">;
def PseudoUseParamI64 : PseudoUseParamInst<Int64Regs>;
def PseudoUseParamI32 : PseudoUseParamInst<Int32Regs>;
def PseudoUseParamI16 : PseudoUseParamInst<Int16Regs>;
-def PseudoUseParamI8 : PseudoUseParamInst<Int8Regs>;
def PseudoUseParamF64 : PseudoUseParamInst<Float64Regs>;
def PseudoUseParamF32 : PseudoUseParamInst<Float32Regs>;
i32imm:$fromWidth, Int32Regs:$addr),
!strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t$dst, [$addr];"), []>;
+ def _areg_64 : NVPTXInst<(outs regclass:$dst),
+ (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr),
+ !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$fromWidth",
+ " \t$dst, [$addr];"), []>;
def _ari : NVPTXInst<(outs regclass:$dst),
(ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
!strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t$dst, [$addr+$offset];"), []>;
+ def _ari_64 : NVPTXInst<(outs regclass:$dst),
+ (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
+ !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$fromWidth",
+ " \t$dst, [$addr+$offset];"), []>;
def _asi : NVPTXInst<(outs regclass:$dst),
(ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
i32imm:$fromWidth, imem:$addr, i32imm:$offset),
"$fromWidth \t$dst, [$addr+$offset];"), []>;
}
-let mayLoad=1, neverHasSideEffects=1 in {
-defm LD_i8 : LD<Int8Regs>;
+let mayLoad=1, hasSideEffects=0 in {
+defm LD_i8 : LD<Int16Regs>;
defm LD_i16 : LD<Int16Regs>;
defm LD_i32 : LD<Int32Regs>;
defm LD_i64 : LD<Int64Regs>;
defm LD_f64 : LD<Float64Regs>;
}
-let VecInstType=isVecLD.Value, mayLoad=1, neverHasSideEffects=1 in {
-defm LD_v2i8 : LD<V2I8Regs>;
-defm LD_v4i8 : LD<V4I8Regs>;
-defm LD_v2i16 : LD<V2I16Regs>;
-defm LD_v4i16 : LD<V4I16Regs>;
-defm LD_v2i32 : LD<V2I32Regs>;
-defm LD_v4i32 : LD<V4I32Regs>;
-defm LD_v2f32 : LD<V2F32Regs>;
-defm LD_v4f32 : LD<V4F32Regs>;
-defm LD_v2i64 : LD<V2I64Regs>;
-defm LD_v2f64 : LD<V2F64Regs>;
-}
-
multiclass ST<NVPTXRegClass regclass> {
def _avar : NVPTXInst<(outs),
(ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
LdStCode:$Sign, i32imm:$toWidth, Int32Regs:$addr),
!strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
" \t[$addr], $src;"), []>;
+ def _areg_64 : NVPTXInst<(outs),
+ (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
+ LdStCode:$Sign, i32imm:$toWidth, Int64Regs:$addr),
+ !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth ",
+ "\t[$addr], $src;"), []>;
def _ari : NVPTXInst<(outs),
(ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
LdStCode:$Sign, i32imm:$toWidth, Int32Regs:$addr, i32imm:$offset),
!strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth",
" \t[$addr+$offset], $src;"), []>;
+ def _ari_64 : NVPTXInst<(outs),
+ (ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
+ LdStCode:$Sign, i32imm:$toWidth, Int64Regs:$addr, i32imm:$offset),
+ !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}$toWidth ",
+ "\t[$addr+$offset], $src;"), []>;
def _asi : NVPTXInst<(outs),
(ins regclass:$src, LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec,
LdStCode:$Sign, i32imm:$toWidth, imem:$addr, i32imm:$offset),
" \t[$addr+$offset], $src;"), []>;
}
-let mayStore=1, neverHasSideEffects=1 in {
-defm ST_i8 : ST<Int8Regs>;
+let mayStore=1, hasSideEffects=0 in {
+defm ST_i8 : ST<Int16Regs>;
defm ST_i16 : ST<Int16Regs>;
defm ST_i32 : ST<Int32Regs>;
defm ST_i64 : ST<Int64Regs>;
defm ST_f64 : ST<Float64Regs>;
}
-let VecInstType=isVecST.Value, mayStore=1, neverHasSideEffects=1 in {
-defm ST_v2i8 : ST<V2I8Regs>;
-defm ST_v4i8 : ST<V4I8Regs>;
-defm ST_v2i16 : ST<V2I16Regs>;
-defm ST_v4i16 : ST<V4I16Regs>;
-defm ST_v2i32 : ST<V2I32Regs>;
-defm ST_v4i32 : ST<V4I32Regs>;
-defm ST_v2f32 : ST<V2F32Regs>;
-defm ST_v4f32 : ST<V4F32Regs>;
-defm ST_v2i64 : ST<V2I64Regs>;
-defm ST_v2f64 : ST<V2F64Regs>;
-}
-
// The following is used only in and after vector elementizations.
// Vector elementization happens at the machine instruction level, so the
// following instruction
i32imm:$fromWidth, Int32Regs:$addr),
!strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
+ def _v2_areg_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
+ (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr),
+ !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t{{$dst1, $dst2}}, [$addr];"), []>;
def _v2_ari : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
i32imm:$fromWidth, Int32Regs:$addr, i32imm:$offset),
!strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
+ def _v2_ari_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
+ (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
+ !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t{{$dst1, $dst2}}, [$addr+$offset];"), []>;
def _v2_asi : NVPTXInst<(outs regclass:$dst1, regclass:$dst2),
(ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
i32imm:$fromWidth, imem:$addr, i32imm:$offset),
i32imm:$fromWidth, Int32Regs:$addr),
!strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
+ def _v4_areg_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
+ regclass:$dst3, regclass:$dst4),
+ (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr),
+ !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr];"), []>;
def _v4_ari : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4),
(ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
!strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
[]>;
+ def _v4_ari_64 : NVPTXInst<(outs regclass:$dst1, regclass:$dst2,
+ regclass:$dst3, regclass:$dst4),
+ (ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
+ !strconcat("ld${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
+ []>;
def _v4_asi : NVPTXInst<(outs regclass:$dst1, regclass:$dst2, regclass:$dst3,
regclass:$dst4),
(ins LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
"$fromWidth \t{{$dst1, $dst2, $dst3, $dst4}}, [$addr+$offset];"),
[]>;
}
-let mayLoad=1, neverHasSideEffects=1 in {
-defm LDV_i8 : LD_VEC<Int8Regs>;
+let mayLoad=1, hasSideEffects=0 in {
+defm LDV_i8 : LD_VEC<Int16Regs>;
defm LDV_i16 : LD_VEC<Int16Regs>;
defm LDV_i32 : LD_VEC<Int32Regs>;
defm LDV_i64 : LD_VEC<Int64Regs>;
LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int32Regs:$addr),
!strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
+ def _v2_areg_64 : NVPTXInst<(outs),
+ (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
+ LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int64Regs:$addr),
+ !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t[$addr], {{$src1, $src2}};"), []>;
def _v2_ari : NVPTXInst<(outs),
(ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int32Regs:$addr,
i32imm:$offset),
!strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
+ def _v2_ari_64 : NVPTXInst<(outs),
+ (ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
+ LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, Int64Regs:$addr,
+ i32imm:$offset),
+ !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t[$addr+$offset], {{$src1, $src2}};"), []>;
def _v2_asi : NVPTXInst<(outs),
(ins regclass:$src1, regclass:$src2, LdStCode:$isVol, LdStCode:$addsp,
LdStCode:$Vec, LdStCode:$Sign, i32imm:$fromWidth, imem:$addr,
i32imm:$fromWidth, Int32Regs:$addr),
!strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
+ def _v4_areg_64 : NVPTXInst<(outs),
+ (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
+ LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr),
+ !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t[$addr], {{$src1, $src2, $src3, $src4}};"), []>;
def _v4_ari : NVPTXInst<(outs),
(ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
!strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
"$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
[]>;
+ def _v4_ari_64 : NVPTXInst<(outs),
+ (ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
+ LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
+ i32imm:$fromWidth, Int64Regs:$addr, i32imm:$offset),
+ !strconcat("st${isVol:volatile}${addsp:addsp}${Vec:vec}.${Sign:sign}",
+ "$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
+ []>;
def _v4_asi : NVPTXInst<(outs),
(ins regclass:$src1, regclass:$src2, regclass:$src3, regclass:$src4,
LdStCode:$isVol, LdStCode:$addsp, LdStCode:$Vec, LdStCode:$Sign,
"$fromWidth \t[$addr+$offset], {{$src1, $src2, $src3, $src4}};"),
[]>;
}
-let mayStore=1, neverHasSideEffects=1 in {
-defm STV_i8 : ST_VEC<Int8Regs>;
+let mayStore=1, hasSideEffects=0 in {
+defm STV_i8 : ST_VEC<Int16Regs>;
defm STV_i16 : ST_VEC<Int16Regs>;
defm STV_i32 : ST_VEC<Int32Regs>;
defm STV_i64 : ST_VEC<Int64Regs>;
//---- Conversion ----
-multiclass CVT_INT_TO_FP <string OpStr, SDNode OpNode> {
-// FIXME: need to add f16 support
-// def CVTf16i8 :
-// NVPTXInst<(outs Float16Regs:$d), (ins Int8Regs:$a),
-// !strconcat(!strconcat("cvt.rn.f16.", OpStr), "8 \t$d, $a;"),
-// [(set Float16Regs:$d, (OpNode Int8Regs:$a))]>;
-// def CVTf16i16 :
-// NVPTXInst<(outs Float16Regs:$d), (ins Int16Regs:$a),
-// !strconcat(!strconcat("cvt.rn.f16.", OpStr), "16 \t$d, $a;"),
-// [(set Float16Regs:$d, (OpNode Int16Regs:$a))]>;
-// def CVTf16i32 :
-// NVPTXInst<(outs Float16Regs:$d), (ins Int32Regs:$a),
-// !strconcat(!strconcat("cvt.rn.f16.", OpStr), "32 \t$d, $a;"),
-// [(set Float16Regs:$d, (OpNode Int32Regs:$a))]>;
-// def CVTf16i64:
-// NVPTXInst<(outs Float16Regs:$d), (ins Int64Regs:$a),
-// !strconcat(!strconcat("cvt.rn.f32.", OpStr), "64 \t$d, $a;"),
-// [(set Float32Regs:$d, (OpNode Int64Regs:$a))]>;
-
- def CVTf32i1 :
- NVPTXInst<(outs Float32Regs:$d), (ins Int1Regs:$a),
- "selp.f32 \t$d, 1.0, 0.0, $a;",
- [(set Float32Regs:$d, (OpNode Int1Regs:$a))]>;
- def CVTf32i8 :
- NVPTXInst<(outs Float32Regs:$d), (ins Int8Regs:$a),
- !strconcat(!strconcat("cvt.rn.f32.", OpStr), "8 \t$d, $a;"),
- [(set Float32Regs:$d, (OpNode Int8Regs:$a))]>;
- def CVTf32i16 :
- NVPTXInst<(outs Float32Regs:$d), (ins Int16Regs:$a),
- !strconcat(!strconcat("cvt.rn.f32.", OpStr), "16 \t$d, $a;"),
- [(set Float32Regs:$d, (OpNode Int16Regs:$a))]>;
- def CVTf32i32 :
- NVPTXInst<(outs Float32Regs:$d), (ins Int32Regs:$a),
- !strconcat(!strconcat("cvt.rn.f32.", OpStr), "32 \t$d, $a;"),
- [(set Float32Regs:$d, (OpNode Int32Regs:$a))]>;
- def CVTf32i64:
- NVPTXInst<(outs Float32Regs:$d), (ins Int64Regs:$a),
- !strconcat(!strconcat("cvt.rn.f32.", OpStr), "64 \t$d, $a;"),
- [(set Float32Regs:$d, (OpNode Int64Regs:$a))]>;
-
- def CVTf64i1 :
- NVPTXInst<(outs Float64Regs:$d), (ins Int1Regs:$a),
- "selp.f64 \t$d, 1.0, 0.0, $a;",
- [(set Float64Regs:$d, (OpNode Int1Regs:$a))]>;
- def CVTf64i8 :
- NVPTXInst<(outs Float64Regs:$d), (ins Int8Regs:$a),
- !strconcat(!strconcat("cvt.rn.f64.", OpStr), "8 \t$d, $a;"),
- [(set Float64Regs:$d, (OpNode Int8Regs:$a))]>;
- def CVTf64i16 :
- NVPTXInst<(outs Float64Regs:$d), (ins Int16Regs:$a),
- !strconcat(!strconcat("cvt.rn.f64.", OpStr), "16 \t$d, $a;"),
- [(set Float64Regs:$d, (OpNode Int16Regs:$a))]>;
- def CVTf64i32 :
- NVPTXInst<(outs Float64Regs:$d), (ins Int32Regs:$a),
- !strconcat(!strconcat("cvt.rn.f64.", OpStr), "32 \t$d, $a;"),
- [(set Float64Regs:$d, (OpNode Int32Regs:$a))]>;
- def CVTf64i64:
- NVPTXInst<(outs Float64Regs:$d), (ins Int64Regs:$a),
- !strconcat(!strconcat("cvt.rn.f64.", OpStr), "64 \t$d, $a;"),
- [(set Float64Regs:$d, (OpNode Int64Regs:$a))]>;
-}
-
-defm Sint_to_fp : CVT_INT_TO_FP <"s", sint_to_fp>;
-defm Uint_to_fp : CVT_INT_TO_FP <"u", uint_to_fp>;
-
-multiclass CVT_FP_TO_INT <string OpStr, SDNode OpNode> {
-// FIXME: need to add f16 support
-// def CVTi8f16:
-// NVPTXInst<(outs Int8Regs:$d), (ins Float16Regs:$a),
-// !strconcat(!strconcat("cvt.rzi.", OpStr), "8.f16 $d, $a;"),
-// [(set Int8Regs:$d, (OpNode Float16Regs:$a))]>;
- def CVTi8f32_ftz:
- NVPTXInst<(outs Int8Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.ftz.", OpStr), "16.f32 \t$d, $a;"),
- [(set Int8Regs:$d, (OpNode Float32Regs:$a))]>, Requires<[doF32FTZ]>;
- def CVTi8f32:
- NVPTXInst<(outs Int8Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "16.f32 \t$d, $a;"),
- [(set Int8Regs:$d, (OpNode Float32Regs:$a))]>;
- def CVTi8f64:
- NVPTXInst<(outs Int8Regs:$d), (ins Float64Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "16.f64 \t$d, $a;"),
- [(set Int8Regs:$d, (OpNode Float64Regs:$a))]>;
-
-// FIXME: need to add f16 support
-// def CVTi16f16:
-// NVPTXInst<(outs Int16Regs:$d), (ins Float16Regs:$a),
-// !strconcat(!strconcat("cvt.rzi.", OpStr), "16.f16 \t$d, $a;"),
-// [(set Int16Regs:$d, (OpNode Float16Regs:$a))]>;
- def CVTi16f32_ftz:
- NVPTXInst<(outs Int16Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.ftz.", OpStr), "16.f32 \t$d, $a;"),
- [(set Int16Regs:$d, (OpNode Float32Regs:$a))]>, Requires<[doF32FTZ]>;
- def CVTi16f32:
- NVPTXInst<(outs Int16Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "16.f32 \t$d, $a;"),
- [(set Int16Regs:$d, (OpNode Float32Regs:$a))]>;
- def CVTi16f64:
- NVPTXInst<(outs Int16Regs:$d), (ins Float64Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "16.f64 \t$d, $a;"),
- [(set Int16Regs:$d, (OpNode Float64Regs:$a))]>;
-
-// FIXME: need to add f16 support
-// def CVTi32f16: def CVTi32f16:
-// NVPTXInst<(outs Int32Regs:$d), (ins Float16Regs:$a),
-// !strconcat(!strconcat("cvt.rzi.", OpStr), "32.f16 \t$d, $a;"),
-// [(set Int32Regs:$d, (OpNode Float16Regs:$a))]>;
- def CVTi32f32_ftz:
- NVPTXInst<(outs Int32Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.ftz.", OpStr), "32.f32 \t$d, $a;"),
- [(set Int32Regs:$d, (OpNode Float32Regs:$a))]>, Requires<[doF32FTZ]>;
- def CVTi32f32:
- NVPTXInst<(outs Int32Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "32.f32 \t$d, $a;"),
- [(set Int32Regs:$d, (OpNode Float32Regs:$a))]>;
- def CVTi32f64:
- NVPTXInst<(outs Int32Regs:$d), (ins Float64Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "32.f64 \t$d, $a;"),
- [(set Int32Regs:$d, (OpNode Float64Regs:$a))]>;
-
-// FIXME: need to add f16 support
-// def CVTi64f16:
-// NVPTXInst<(outs Int64Regs:$d), (ins Float16Regs:$a),
-// !strconcat(!strconcat("cvt.rzi.", OpStr), "64.f16 \t$d, $a;"),
-// [(set Int64Regs:$d, (OpNode Float16Regs:$a))]>;
- def CVTi64f32_ftz:
- NVPTXInst<(outs Int64Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.ftz.", OpStr), "64.f32 \t$d, $a;"),
- [(set Int64Regs:$d, (OpNode Float32Regs:$a))]>, Requires<[doF32FTZ]>;
- def CVTi64f32:
- NVPTXInst<(outs Int64Regs:$d), (ins Float32Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "64.f32 \t$d, $a;"),
- [(set Int64Regs:$d, (OpNode Float32Regs:$a))]>;
- def CVTi64f64:
- NVPTXInst<(outs Int64Regs:$d), (ins Float64Regs:$a),
- !strconcat(!strconcat("cvt.rzi.", OpStr), "64.f64 \t$d, $a;"),
- [(set Int64Regs:$d, (OpNode Float64Regs:$a))]>;
-}
-
-defm Fp_to_sint : CVT_FP_TO_INT <"s", fp_to_sint>;
-defm Fp_to_uint : CVT_FP_TO_INT <"u", fp_to_uint>;
-
-multiclass INT_EXTEND_UNSIGNED_1 <SDNode OpNode> {
- def ext1to8:
- NVPTXInst<(outs Int8Regs:$d), (ins Int1Regs:$a),
- "selp.u16 \t$d, 1, 0, $a;",
- [(set Int8Regs:$d, (OpNode Int1Regs:$a))]>;
- def ext1to16:
- NVPTXInst<(outs Int16Regs:$d), (ins Int1Regs:$a),
- "selp.u16 \t$d, 1, 0, $a;",
- [(set Int16Regs:$d, (OpNode Int1Regs:$a))]>;
- def ext1to32:
- NVPTXInst<(outs Int32Regs:$d), (ins Int1Regs:$a),
- "selp.u32 \t$d, 1, 0, $a;",
- [(set Int32Regs:$d, (OpNode Int1Regs:$a))]>;
- def ext1to64:
- NVPTXInst<(outs Int64Regs:$d), (ins Int1Regs:$a),
- "selp.u64 \t$d, 1, 0, $a;",
- [(set Int64Regs:$d, (OpNode Int1Regs:$a))]>;
-}
-
-multiclass INT_EXTEND_SIGNED_1 <SDNode OpNode> {
- def ext1to8:
- NVPTXInst<(outs Int8Regs:$d), (ins Int1Regs:$a),
- "selp.s16 \t$d, -1, 0, $a;",
- [(set Int8Regs:$d, (OpNode Int1Regs:$a))]>;
- def ext1to16:
- NVPTXInst<(outs Int16Regs:$d), (ins Int1Regs:$a),
- "selp.s16 \t$d, -1, 0, $a;",
- [(set Int16Regs:$d, (OpNode Int1Regs:$a))]>;
- def ext1to32:
- NVPTXInst<(outs Int32Regs:$d), (ins Int1Regs:$a),
- "selp.s32 \t$d, -1, 0, $a;",
- [(set Int32Regs:$d, (OpNode Int1Regs:$a))]>;
- def ext1to64:
- NVPTXInst<(outs Int64Regs:$d), (ins Int1Regs:$a),
- "selp.s64 \t$d, -1, 0, $a;",
- [(set Int64Regs:$d, (OpNode Int1Regs:$a))]>;
-}
-
-multiclass INT_EXTEND <string OpStr, SDNode OpNode> {
- // All Int8Regs are emiited as 16bit registers in ptx.
- // And there is no selp.u8 in ptx.
- def ext8to16:
- NVPTXInst<(outs Int16Regs:$d), (ins Int8Regs:$a),
- !strconcat("cvt.", !strconcat(OpStr, !strconcat("16.",
- !strconcat(OpStr, "8 \t$d, $a;")))),
- [(set Int16Regs:$d, (OpNode Int8Regs:$a))]>;
- def ext8to32:
- NVPTXInst<(outs Int32Regs:$d), (ins Int8Regs:$a),
- !strconcat("cvt.", !strconcat(OpStr, !strconcat("32.",
- !strconcat(OpStr, "8 \t$d, $a;")))),
- [(set Int32Regs:$d, (OpNode Int8Regs:$a))]>;
- def ext8to64:
- NVPTXInst<(outs Int64Regs:$d), (ins Int8Regs:$a),
- !strconcat("cvt.", !strconcat(OpStr, !strconcat("64.",
- !strconcat(OpStr, "8 \t$d, $a;")))),
- [(set Int64Regs:$d, (OpNode Int8Regs:$a))]>;
- def ext16to32:
- NVPTXInst<(outs Int32Regs:$d), (ins Int16Regs:$a),
- !strconcat("cvt.", !strconcat(OpStr, !strconcat("32.",
- !strconcat(OpStr, "16 \t$d, $a;")))),
- [(set Int32Regs:$d, (OpNode Int16Regs:$a))]>;
- def ext16to64:
- NVPTXInst<(outs Int64Regs:$d), (ins Int16Regs:$a),
- !strconcat("cvt.", !strconcat(OpStr, !strconcat("64.",
- !strconcat(OpStr, "16 \t$d, $a;")))),
- [(set Int64Regs:$d, (OpNode Int16Regs:$a))]>;
- def ext32to64:
- NVPTXInst<(outs Int64Regs:$d), (ins Int32Regs:$a),
- !strconcat("cvt.", !strconcat(OpStr, !strconcat("64.",
- !strconcat(OpStr, "32 \t$d, $a;")))),
- [(set Int64Regs:$d, (OpNode Int32Regs:$a))]>;
-}
-
-defm Sint_extend_1 : INT_EXTEND_SIGNED_1<sext>;
-defm Zint_extend_1 : INT_EXTEND_UNSIGNED_1<zext>;
-defm Aint_extend_1 : INT_EXTEND_UNSIGNED_1<anyext>;
-
-defm Sint_extend : INT_EXTEND <"s", sext>;
-defm Zint_extend : INT_EXTEND <"u", zext>;
-defm Aint_extend : INT_EXTEND <"u", anyext>;
-
-class TRUNC_to1_asm<string sz> {
- string s = !strconcat("{{\n\t",
- !strconcat(".reg ",
- !strconcat(sz,
- !strconcat(" temp;\n\t",
- !strconcat("and",
- !strconcat(sz,
- !strconcat("\t temp, $a, 1;\n\t",
- !strconcat("setp",
- !strconcat(sz, ".eq \t $d, temp, 1;\n\t}}")))))))));
-}
-
-def TRUNC_64to32 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
- "cvt.u32.u64 \t$d, $a;",
- [(set Int32Regs:$d, (trunc Int64Regs:$a))]>;
-def TRUNC_64to16 : NVPTXInst<(outs Int16Regs:$d), (ins Int64Regs:$a),
- "cvt.u16.u64 \t$d, $a;",
- [(set Int16Regs:$d, (trunc Int64Regs:$a))]>;
-def TRUNC_64to8 : NVPTXInst<(outs Int8Regs:$d), (ins Int64Regs:$a),
- "cvt.u8.u64 \t$d, $a;",
- [(set Int8Regs:$d, (trunc Int64Regs:$a))]>;
-def TRUNC_32to16 : NVPTXInst<(outs Int16Regs:$d), (ins Int32Regs:$a),
- "cvt.u16.u32 \t$d, $a;",
- [(set Int16Regs:$d, (trunc Int32Regs:$a))]>;
-def TRUNC_32to8 : NVPTXInst<(outs Int8Regs:$d), (ins Int32Regs:$a),
- "cvt.u8.u32 \t$d, $a;",
- [(set Int8Regs:$d, (trunc Int32Regs:$a))]>;
-def TRUNC_16to8 : NVPTXInst<(outs Int8Regs:$d), (ins Int16Regs:$a),
- "cvt.u8.u16 \t$d, $a;",
- [(set Int8Regs:$d, (trunc Int16Regs:$a))]>;
-def TRUNC_64to1 : NVPTXInst<(outs Int1Regs:$d), (ins Int64Regs:$a),
- TRUNC_to1_asm<".b64">.s,
- [(set Int1Regs:$d, (trunc Int64Regs:$a))]>;
-def TRUNC_32to1 : NVPTXInst<(outs Int1Regs:$d), (ins Int32Regs:$a),
- TRUNC_to1_asm<".b32">.s,
- [(set Int1Regs:$d, (trunc Int32Regs:$a))]>;
-def TRUNC_16to1 : NVPTXInst<(outs Int1Regs:$d), (ins Int16Regs:$a),
- TRUNC_to1_asm<".b16">.s,
- [(set Int1Regs:$d, (trunc Int16Regs:$a))]>;
-def TRUNC_8to1 : NVPTXInst<(outs Int1Regs:$d), (ins Int8Regs:$a),
- TRUNC_to1_asm<".b16">.s,
- [(set Int1Regs:$d, (trunc Int8Regs:$a))]>;
-
-// Select instructions
-def : Pat<(select Int32Regs:$pred, Int8Regs:$a, Int8Regs:$b),
- (SELECTi8rr Int8Regs:$a, Int8Regs:$b, (TRUNC_32to1 Int32Regs:$pred))>;
-def : Pat<(select Int32Regs:$pred, Int16Regs:$a, Int16Regs:$b),
- (SELECTi16rr Int16Regs:$a, Int16Regs:$b,
- (TRUNC_32to1 Int32Regs:$pred))>;
-def : Pat<(select Int32Regs:$pred, Int32Regs:$a, Int32Regs:$b),
- (SELECTi32rr Int32Regs:$a, Int32Regs:$b,
- (TRUNC_32to1 Int32Regs:$pred))>;
-def : Pat<(select Int32Regs:$pred, Int64Regs:$a, Int64Regs:$b),
- (SELECTi64rr Int64Regs:$a, Int64Regs:$b,
- (TRUNC_32to1 Int32Regs:$pred))>;
-def : Pat<(select Int32Regs:$pred, Float32Regs:$a, Float32Regs:$b),
- (SELECTf32rr Float32Regs:$a, Float32Regs:$b,
- (TRUNC_32to1 Int32Regs:$pred))>;
-def : Pat<(select Int32Regs:$pred, Float64Regs:$a, Float64Regs:$b),
- (SELECTf64rr Float64Regs:$a, Float64Regs:$b,
- (TRUNC_32to1 Int32Regs:$pred))>;
-
class F_BITCONVERT<string SzStr, NVPTXRegClass regclassIn,
NVPTXRegClass regclassOut> :
NVPTXInst<(outs regclassOut:$d), (ins regclassIn:$a),
def BITCONVERT_64_I2F : F_BITCONVERT<"64", Int64Regs, Float64Regs>;
def BITCONVERT_64_F2I : F_BITCONVERT<"64", Float64Regs, Int64Regs>;
+// NOTE: pred->fp are currently sub-optimal due to an issue in TableGen where
+// we cannot specify floating-point literals in isel patterns. Therefore, we
+// use an integer selp to select either 1 or 0 and then cvt to floating-point.
+
+// sint -> f32
+def : Pat<(f32 (sint_to_fp Int1Regs:$a)),
+ (CVT_f32_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+def : Pat<(f32 (sint_to_fp Int16Regs:$a)),
+ (CVT_f32_s16 Int16Regs:$a, CvtRN)>;
+def : Pat<(f32 (sint_to_fp Int32Regs:$a)),
+ (CVT_f32_s32 Int32Regs:$a, CvtRN)>;
+def : Pat<(f32 (sint_to_fp Int64Regs:$a)),
+ (CVT_f32_s64 Int64Regs:$a, CvtRN)>;
+
+// uint -> f32
+def : Pat<(f32 (uint_to_fp Int1Regs:$a)),
+ (CVT_f32_u32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+def : Pat<(f32 (uint_to_fp Int16Regs:$a)),
+ (CVT_f32_u16 Int16Regs:$a, CvtRN)>;
+def : Pat<(f32 (uint_to_fp Int32Regs:$a)),
+ (CVT_f32_u32 Int32Regs:$a, CvtRN)>;
+def : Pat<(f32 (uint_to_fp Int64Regs:$a)),
+ (CVT_f32_u64 Int64Regs:$a, CvtRN)>;
+
+// sint -> f64
+def : Pat<(f64 (sint_to_fp Int1Regs:$a)),
+ (CVT_f64_s32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+def : Pat<(f64 (sint_to_fp Int16Regs:$a)),
+ (CVT_f64_s16 Int16Regs:$a, CvtRN)>;
+def : Pat<(f64 (sint_to_fp Int32Regs:$a)),
+ (CVT_f64_s32 Int32Regs:$a, CvtRN)>;
+def : Pat<(f64 (sint_to_fp Int64Regs:$a)),
+ (CVT_f64_s64 Int64Regs:$a, CvtRN)>;
+
+// uint -> f64
+def : Pat<(f64 (uint_to_fp Int1Regs:$a)),
+ (CVT_f64_u32 (SELP_u32ii 1, 0, Int1Regs:$a), CvtRN)>;
+def : Pat<(f64 (uint_to_fp Int16Regs:$a)),
+ (CVT_f64_u16 Int16Regs:$a, CvtRN)>;
+def : Pat<(f64 (uint_to_fp Int32Regs:$a)),
+ (CVT_f64_u32 Int32Regs:$a, CvtRN)>;
+def : Pat<(f64 (uint_to_fp Int64Regs:$a)),
+ (CVT_f64_u64 Int64Regs:$a, CvtRN)>;
+
+
+// f32 -> sint
+def : Pat<(i1 (fp_to_sint Float32Regs:$a)),
+ (SETP_b32ri (BITCONVERT_32_F2I Float32Regs:$a), 0, CmpEQ)>;
+def : Pat<(i16 (fp_to_sint Float32Regs:$a)),
+ (CVT_s16_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(i16 (fp_to_sint Float32Regs:$a)),
+ (CVT_s16_f32 Float32Regs:$a, CvtRZI)>;
+def : Pat<(i32 (fp_to_sint Float32Regs:$a)),
+ (CVT_s32_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(i32 (fp_to_sint Float32Regs:$a)),
+ (CVT_s32_f32 Float32Regs:$a, CvtRZI)>;
+def : Pat<(i64 (fp_to_sint Float32Regs:$a)),
+ (CVT_s64_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(i64 (fp_to_sint Float32Regs:$a)),
+ (CVT_s64_f32 Float32Regs:$a, CvtRZI)>;
+
+// f32 -> uint
+def : Pat<(i1 (fp_to_uint Float32Regs:$a)),
+ (SETP_b32ri (BITCONVERT_32_F2I Float32Regs:$a), 0, CmpEQ)>;
+def : Pat<(i16 (fp_to_uint Float32Regs:$a)),
+ (CVT_u16_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(i16 (fp_to_uint Float32Regs:$a)),
+ (CVT_u16_f32 Float32Regs:$a, CvtRZI)>;
+def : Pat<(i32 (fp_to_uint Float32Regs:$a)),
+ (CVT_u32_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(i32 (fp_to_uint Float32Regs:$a)),
+ (CVT_u32_f32 Float32Regs:$a, CvtRZI)>;
+def : Pat<(i64 (fp_to_uint Float32Regs:$a)),
+ (CVT_u64_f32 Float32Regs:$a, CvtRZI_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(i64 (fp_to_uint Float32Regs:$a)),
+ (CVT_u64_f32 Float32Regs:$a, CvtRZI)>;
+
+// f64 -> sint
+def : Pat<(i1 (fp_to_sint Float64Regs:$a)),
+ (SETP_b64ri (BITCONVERT_64_F2I Float64Regs:$a), 0, CmpEQ)>;
+def : Pat<(i16 (fp_to_sint Float64Regs:$a)),
+ (CVT_s16_f64 Float64Regs:$a, CvtRZI)>;
+def : Pat<(i32 (fp_to_sint Float64Regs:$a)),
+ (CVT_s32_f64 Float64Regs:$a, CvtRZI)>;
+def : Pat<(i64 (fp_to_sint Float64Regs:$a)),
+ (CVT_s64_f64 Float64Regs:$a, CvtRZI)>;
+
+// f64 -> uint
+def : Pat<(i1 (fp_to_uint Float64Regs:$a)),
+ (SETP_b64ri (BITCONVERT_64_F2I Float64Regs:$a), 0, CmpEQ)>;
+def : Pat<(i16 (fp_to_uint Float64Regs:$a)),
+ (CVT_u16_f64 Float64Regs:$a, CvtRZI)>;
+def : Pat<(i32 (fp_to_uint Float64Regs:$a)),
+ (CVT_u32_f64 Float64Regs:$a, CvtRZI)>;
+def : Pat<(i64 (fp_to_uint Float64Regs:$a)),
+ (CVT_u64_f64 Float64Regs:$a, CvtRZI)>;
+
+// sext i1
+def : Pat<(i16 (sext Int1Regs:$a)),
+ (SELP_s16ii -1, 0, Int1Regs:$a)>;
+def : Pat<(i32 (sext Int1Regs:$a)),
+ (SELP_s32ii -1, 0, Int1Regs:$a)>;
+def : Pat<(i64 (sext Int1Regs:$a)),
+ (SELP_s64ii -1, 0, Int1Regs:$a)>;
+
+// zext i1
+def : Pat<(i16 (zext Int1Regs:$a)),
+ (SELP_u16ii 1, 0, Int1Regs:$a)>;
+def : Pat<(i32 (zext Int1Regs:$a)),
+ (SELP_u32ii 1, 0, Int1Regs:$a)>;
+def : Pat<(i64 (zext Int1Regs:$a)),
+ (SELP_u64ii 1, 0, Int1Regs:$a)>;
+
+// anyext i1
+def : Pat<(i16 (anyext Int1Regs:$a)),
+ (SELP_u16ii -1, 0, Int1Regs:$a)>;
+def : Pat<(i32 (anyext Int1Regs:$a)),
+ (SELP_u32ii -1, 0, Int1Regs:$a)>;
+def : Pat<(i64 (anyext Int1Regs:$a)),
+ (SELP_u64ii -1, 0, Int1Regs:$a)>;
+
+// sext i16
+def : Pat<(i32 (sext Int16Regs:$a)),
+ (CVT_s32_s16 Int16Regs:$a, CvtNONE)>;
+def : Pat<(i64 (sext Int16Regs:$a)),
+ (CVT_s64_s16 Int16Regs:$a, CvtNONE)>;
+
+// zext i16
+def : Pat<(i32 (zext Int16Regs:$a)),
+ (CVT_u32_u16 Int16Regs:$a, CvtNONE)>;
+def : Pat<(i64 (zext Int16Regs:$a)),
+ (CVT_u64_u16 Int16Regs:$a, CvtNONE)>;
+
+// anyext i16
+def : Pat<(i32 (anyext Int16Regs:$a)),
+ (CVT_u32_u16 Int16Regs:$a, CvtNONE)>;
+def : Pat<(i64 (anyext Int16Regs:$a)),
+ (CVT_u64_u16 Int16Regs:$a, CvtNONE)>;
+
+// sext i32
+def : Pat<(i64 (sext Int32Regs:$a)),
+ (CVT_s64_s32 Int32Regs:$a, CvtNONE)>;
+
+// zext i32
+def : Pat<(i64 (zext Int32Regs:$a)),
+ (CVT_u64_u32 Int32Regs:$a, CvtNONE)>;
+
+// anyext i32
+def : Pat<(i64 (anyext Int32Regs:$a)),
+ (CVT_u64_u32 Int32Regs:$a, CvtNONE)>;
+
+
+// truncate i64
+def : Pat<(i32 (trunc Int64Regs:$a)),
+ (CVT_u32_u64 Int64Regs:$a, CvtNONE)>;
+def : Pat<(i16 (trunc Int64Regs:$a)),
+ (CVT_u16_u64 Int64Regs:$a, CvtNONE)>;
+def : Pat<(i1 (trunc Int64Regs:$a)),
+ (SETP_b64ri (ANDb64ri Int64Regs:$a, 1), 1, CmpEQ)>;
+
+// truncate i32
+def : Pat<(i16 (trunc Int32Regs:$a)),
+ (CVT_u16_u32 Int32Regs:$a, CvtNONE)>;
+def : Pat<(i1 (trunc Int32Regs:$a)),
+ (SETP_b32ri (ANDb32ri Int32Regs:$a, 1), 1, CmpEQ)>;
+
+// truncate i16
+def : Pat<(i1 (trunc Int16Regs:$a)),
+ (SETP_b16ri (ANDb16ri Int16Regs:$a, 1), 1, CmpEQ)>;
+
+// sext_inreg
+def : Pat<(sext_inreg Int16Regs:$a, i8), (CVT_INREG_s16_s8 Int16Regs:$a)>;
+def : Pat<(sext_inreg Int32Regs:$a, i8), (CVT_INREG_s32_s8 Int32Regs:$a)>;
+def : Pat<(sext_inreg Int32Regs:$a, i16), (CVT_INREG_s32_s16 Int32Regs:$a)>;
+def : Pat<(sext_inreg Int64Regs:$a, i8), (CVT_INREG_s64_s8 Int64Regs:$a)>;
+def : Pat<(sext_inreg Int64Regs:$a, i16), (CVT_INREG_s64_s16 Int64Regs:$a)>;
+def : Pat<(sext_inreg Int64Regs:$a, i32), (CVT_INREG_s64_s32 Int64Regs:$a)>;
+
+
+// Select instructions with 32-bit predicates
+def : Pat<(select Int32Regs:$pred, Int16Regs:$a, Int16Regs:$b),
+ (SELP_b16rr Int16Regs:$a, Int16Regs:$b,
+ (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
+def : Pat<(select Int32Regs:$pred, Int32Regs:$a, Int32Regs:$b),
+ (SELP_b32rr Int32Regs:$a, Int32Regs:$b,
+ (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
+def : Pat<(select Int32Regs:$pred, Int64Regs:$a, Int64Regs:$b),
+ (SELP_b64rr Int64Regs:$a, Int64Regs:$b,
+ (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
+def : Pat<(select Int32Regs:$pred, Float32Regs:$a, Float32Regs:$b),
+ (SELP_f32rr Float32Regs:$a, Float32Regs:$b,
+ (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
+def : Pat<(select Int32Regs:$pred, Float64Regs:$a, Float64Regs:$b),
+ (SELP_f64rr Float64Regs:$a, Float64Regs:$b,
+ (SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
+
+
// pack a set of smaller int registers to a larger int register
-def V4I8toI32 : NVPTXInst<(outs Int32Regs:$d),
- (ins Int8Regs:$s1, Int8Regs:$s2,
- Int8Regs:$s3, Int8Regs:$s4),
- !strconcat("{{\n\t.reg .b8\t%t<4>;",
- !strconcat("\n\tcvt.u8.u8\t%t0, $s1;",
- !strconcat("\n\tcvt.u8.u8\t%t1, $s2;",
- !strconcat("\n\tcvt.u8.u8\t%t2, $s3;",
- !strconcat("\n\tcvt.u8.u8\t%t3, $s4;",
- "\n\tmov.b32\t$d, {%t0, %t1, %t2, %t3};\n\t}}"))))),
- []>;
def V4I16toI64 : NVPTXInst<(outs Int64Regs:$d),
(ins Int16Regs:$s1, Int16Regs:$s2,
Int16Regs:$s3, Int16Regs:$s4),
"mov.b64\t$d, {{$s1, $s2, $s3, $s4}};",
[]>;
-def V2I8toI16 : NVPTXInst<(outs Int16Regs:$d),
- (ins Int8Regs:$s1, Int8Regs:$s2),
- !strconcat("{{\n\t.reg .b8\t%t<2>;",
- !strconcat("\n\tcvt.u8.u8\t%t0, $s1;",
- !strconcat("\n\tcvt.u8.u8\t%t1, $s2;",
- "\n\tmov.b16\t$d, {%t0, %t1};\n\t}}"))),
- []>;
def V2I16toI32 : NVPTXInst<(outs Int32Regs:$d),
(ins Int16Regs:$s1, Int16Regs:$s2),
"mov.b32\t$d, {{$s1, $s2}};",
[]>;
// unpack a larger int register to a set of smaller int registers
-def I32toV4I8 : NVPTXInst<(outs Int8Regs:$d1, Int8Regs:$d2,
- Int8Regs:$d3, Int8Regs:$d4),
- (ins Int32Regs:$s),
- !strconcat("{{\n\t.reg .b8\t%t<4>;",
- !strconcat("\n\tmov.b32\t{%t0, %t1, %t2, %t3}, $s;",
- !strconcat("\n\tcvt.u8.u8\t$d1, %t0;",
- !strconcat("\n\tcvt.u8.u8\t$d2, %t1;",
- !strconcat("\n\tcvt.u8.u8\t$d3, %t2;",
- "\n\tcvt.u8.u8\t$d4, %t3;\n\t}}"))))),
- []>;
def I64toV4I16 : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2,
Int16Regs:$d3, Int16Regs:$d4),
(ins Int64Regs:$s),
"mov.b64\t{{$d1, $d2, $d3, $d4}}, $s;",
[]>;
-def I16toV2I8 : NVPTXInst<(outs Int8Regs:$d1, Int8Regs:$d2),
- (ins Int16Regs:$s),
- !strconcat("{{\n\t.reg .b8\t%t<2>;",
- !strconcat("\n\tmov.b16\t{%t0, %t1}, $s;",
- !strconcat("\n\tcvt.u8.u8\t$d1, %t0;",
- "\n\tcvt.u8.u8\t$d2, %t1;\n\t}}"))),
- []>;
def I32toV2I16 : NVPTXInst<(outs Int16Regs:$d1, Int16Regs:$d2),
(ins Int32Regs:$s),
"mov.b32\t{{$d1, $d2}}, $s;",
"mov.b64\t{{$d1, $d2}}, $s;",
[]>;
-def FPRound_ftz : NVPTXInst<(outs Float32Regs:$d), (ins Float64Regs:$a),
- "cvt.rn.ftz.f32.f64 \t$d, $a;",
- [(set Float32Regs:$d, (fround Float64Regs:$a))]>, Requires<[doF32FTZ]>;
-
-def FPRound : NVPTXInst<(outs Float32Regs:$d), (ins Float64Regs:$a),
- "cvt.rn.f32.f64 \t$d, $a;",
- [(set Float32Regs:$d, (fround Float64Regs:$a))]>;
-
-def FPExtend_ftz : NVPTXInst<(outs Float64Regs:$d), (ins Float32Regs:$a),
- "cvt.ftz.f64.f32 \t$d, $a;",
- [(set Float64Regs:$d, (fextend Float32Regs:$a))]>, Requires<[doF32FTZ]>;
-
-def FPExtend : NVPTXInst<(outs Float64Regs:$d), (ins Float32Regs:$a),
- "cvt.f64.f32 \t$d, $a;",
- [(set Float64Regs:$d, (fextend Float32Regs:$a))]>;
+// Count leading zeros
+def CLZr32 : NVPTXInst<(outs Int32Regs:$d), (ins Int32Regs:$a),
+ "clz.b32\t$d, $a;",
+ []>;
+def CLZr64 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
+ "clz.b64\t$d, $a;",
+ []>;
+
+// 32-bit has a direct PTX instruction
+def : Pat<(ctlz Int32Regs:$a),
+ (CLZr32 Int32Regs:$a)>;
+def : Pat<(ctlz_zero_undef Int32Regs:$a),
+ (CLZr32 Int32Regs:$a)>;
+
+// For 64-bit, the result in PTX is actually 32-bit so we zero-extend
+// to 64-bit to match the LLVM semantics
+def : Pat<(ctlz Int64Regs:$a),
+ (CVT_u64_u32 (CLZr64 Int64Regs:$a), CvtNONE)>;
+def : Pat<(ctlz_zero_undef Int64Regs:$a),
+ (CVT_u64_u32 (CLZr64 Int64Regs:$a), CvtNONE)>;
+
+// For 16-bit, we zero-extend to 32-bit, then trunc the result back
+// to 16-bits (ctlz of a 16-bit value is guaranteed to require less
+// than 16 bits to store). We also need to subtract 16 because the
+// high-order 16 zeros were counted.
+def : Pat<(ctlz Int16Regs:$a),
+ (SUBi16ri (CVT_u16_u32 (CLZr32
+ (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
+ CvtNONE), 16)>;
+def : Pat<(ctlz_zero_undef Int16Regs:$a),
+ (SUBi16ri (CVT_u16_u32 (CLZr32
+ (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
+ CvtNONE), 16)>;
+
+// Population count
+def POPCr32 : NVPTXInst<(outs Int32Regs:$d), (ins Int32Regs:$a),
+ "popc.b32\t$d, $a;",
+ []>;
+def POPCr64 : NVPTXInst<(outs Int32Regs:$d), (ins Int64Regs:$a),
+ "popc.b64\t$d, $a;",
+ []>;
+
+// 32-bit has a direct PTX instruction
+def : Pat<(ctpop Int32Regs:$a),
+ (POPCr32 Int32Regs:$a)>;
+
+// For 64-bit, the result in PTX is actually 32-bit so we zero-extend
+// to 64-bit to match the LLVM semantics
+def : Pat<(ctpop Int64Regs:$a),
+ (CVT_u64_u32 (POPCr64 Int64Regs:$a), CvtNONE)>;
+
+// For 16-bit, we zero-extend to 32-bit, then trunc the result back
+// to 16-bits (ctpop of a 16-bit value is guaranteed to require less
+// than 16 bits to store)
+def : Pat<(ctpop Int16Regs:$a),
+ (CVT_u16_u32 (POPCr32 (CVT_u32_u16 Int16Regs:$a, CvtNONE)),
+ CvtNONE)>;
+
+// fround f64 -> f32
+def : Pat<(f32 (fround Float64Regs:$a)),
+ (CVT_f32_f64 Float64Regs:$a, CvtRN_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(f32 (fround Float64Regs:$a)),
+ (CVT_f32_f64 Float64Regs:$a, CvtRN)>;
+
+// fextend f32 -> f64
+def : Pat<(f64 (fextend Float32Regs:$a)),
+ (CVT_f64_f32 Float32Regs:$a, CvtNONE_FTZ)>, Requires<[doF32FTZ]>;
+def : Pat<(f64 (fextend Float32Regs:$a)),
+ (CVT_f64_f32 Float32Regs:$a, CvtNONE)>;
def retflag : SDNode<"NVPTXISD::RET_FLAG", SDTNone,
[SDNPHasChain, SDNPOptInGlue]>;
[(br bb:$target)]>;
}
-def : Pat<(brcond Int32Regs:$a, bb:$target), (CBranch
- (ISetUNEi32ri_p Int32Regs:$a, 0), bb:$target)>;
+def : Pat<(brcond Int32Regs:$a, bb:$target),
+ (CBranch (SETP_u32ri Int32Regs:$a, 0, CmpNE), bb:$target)>;
// SelectionDAGBuilder::visitSWitchCase() will invert the condition of a
// conditional branch if
"trap;",
[(trap)]>;
-include "NVPTXVector.td"
+// Call prototype wrapper
+def SDTCallPrototype : SDTypeProfile<0, 1, [SDTCisInt<0>]>;
+def CallPrototype
+ : SDNode<"NVPTXISD::CallPrototype", SDTCallPrototype,
+ [SDNPHasChain, SDNPOutGlue, SDNPInGlue, SDNPSideEffect]>;
+def ProtoIdent : Operand<i32> {
+ let PrintMethod = "printProtoIdent";
+}
+def CALL_PROTOTYPE
+ : NVPTXInst<(outs), (ins ProtoIdent:$ident),
+ "$ident", [(CallPrototype (i32 texternalsym:$ident))]>;
+
+
include "NVPTXIntrinsics.td"