diff options
Diffstat (limited to 'contrib/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td')
-rw-r--r-- | contrib/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td | 2114 |
1 files changed, 934 insertions, 1180 deletions
diff --git a/contrib/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td b/contrib/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td index da6dd39..b23f1e4 100644 --- a/contrib/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td +++ b/contrib/llvm/lib/Target/NVPTX/NVPTXInstrInfo.td @@ -32,6 +32,86 @@ def isVecOther : VecInstTypeEnum<15>; 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 //===----------------------------------------------------------------------===// @@ -56,127 +136,31 @@ def hasLDG : Predicate<"Subtarget.hasLDG()">; def hasLDU : Predicate<"Subtarget.hasLDU()">; def hasGenericLdSt : Predicate<"Subtarget.hasGenericLdSt()">; -def doF32FTZ : Predicate<"UseF32FTZ">; +def doF32FTZ : Predicate<"useF32FTZ()">; +def doNoF32FTZ : Predicate<"!useF32FTZ()">; def doFMAF32 : Predicate<"doFMAF32">; -def doFMAF32_ftz : Predicate<"(doFMAF32 && UseF32FTZ)">; +def doFMAF32_ftz : Predicate<"(doFMAF32 && useF32FTZ())">; def doFMAF32AGG : Predicate<"doFMAF32AGG">; -def doFMAF32AGG_ftz : Predicate<"(doFMAF32AGG && UseF32FTZ)">; +def doFMAF32AGG_ftz : Predicate<"(doFMAF32AGG && useF32FTZ())">; def doFMAF64 : Predicate<"doFMAF64">; def doFMAF64AGG : Predicate<"doFMAF64AGG">; -def doFMADF32 : Predicate<"doFMADF32">; -def doFMADF32_ftz : Predicate<"(doFMADF32 && UseF32FTZ)">; def doMulWide : Predicate<"doMulWide">; def allowFMA : Predicate<"allowFMA">; -def allowFMA_ftz : Predicate<"(allowFMA && UseF32FTZ)">; +def allowFMA_ftz : Predicate<"(allowFMA && useF32FTZ())">; -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 do_SQRTF32_APPROX : Predicate<"do_SQRTF32_PREC==0">; -def do_SQRTF32_RN : Predicate<"do_SQRTF32_PREC==1">; +def do_SQRTF32_APPROX : Predicate<"!usePrecSqrtF32()">; +def do_SQRTF32_RN : Predicate<"usePrecSqrtF32()">; 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. -// -// -//===----------------------------------------------------------------------===// - -// 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}}")))))))))))); -} - -// 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}}"))))))))))))); -} - -// 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}}"))))))))))))); -} - //===----------------------------------------------------------------------===// // Some Common Instruction Class Templates @@ -204,66 +188,6 @@ multiclass I3<string OpcStr, SDNode OpNode> { 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> { @@ -369,6 +293,90 @@ multiclass F2<string OpcStr, SDNode OpNode> { //===----------------------------------------------------------------------===// //----------------------------------- +// General Type Conversion +//----------------------------------- + +let neverHasSideEffects = 1 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, $src;", []>; +} + +//----------------------------------- // Integer Arithmetic //----------------------------------- @@ -522,81 +530,17 @@ def : Pat<(mul (zext Int16Regs:$a), (i32 UInt16Const:$b)), defm MULT : I3<"mul.lo.s", mul>; -defm MULTHS : I3_noi8<"mul.hi.s", mulhs>; -defm MULTHU : I3_noi8<"mul.hi.u", mulhu>; -def MULTHSi8rr : NVPTXInst<(outs Int8Regs:$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:$c))]>; - def MAD16rrr : NVPTXInst<(outs Int16Regs:$dst), (ins Int16Regs:$a, Int16Regs:$b, Int16Regs:$c), "mad.lo.s16 \t$dst, $a, $b, $c;", @@ -661,10 +605,6 @@ def MAD64rii : NVPTXInst<(outs Int64Regs:$dst), (mul 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:$src))]>; @@ -842,6 +782,16 @@ def FDIV32ri_prec : NVPTXInst<(outs Float32Regs:$dst), (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), @@ -912,8 +862,6 @@ multiclass FPCONTRACT64<string OpcStr, Predicate Pred> { // 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>; @@ -952,8 +900,6 @@ multiclass FPCONTRACT64_SUB_PAT<NVPTXInst Inst, Predicate 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>; def SINF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src), @@ -963,6 +909,41 @@ def COSF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$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:$y))>; + //----------------------------------- // Logical Arithmetic //----------------------------------- @@ -974,12 +955,6 @@ multiclass LOG_FORMAT<string OpcStr, SDNode OpNode> { def b1ri: NVPTXInst<(outs Int1Regs:$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:$a, @@ -1010,9 +985,6 @@ defm XOR : LOG_FORMAT<"xor", xor>; def NOT1: NVPTXInst<(outs Int1Regs:$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:$src))]>; @@ -1056,21 +1028,13 @@ multiclass LSHIFT_FORMAT<string OpcStr, SDNode OpNode> { !strconcat(OpcStr, "16 \t$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, $b;"), @@ -1102,20 +1066,10 @@ multiclass RSHIFT_FORMAT<string OpcStr, SDNode OpNode, string CVTStr> { !strconcat(OpcStr, "16 \t$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>; // 32bit def ROT32imm_sw : NVPTXInst<(outs Int32Regs:$dst), @@ -1213,6 +1167,120 @@ def ROTR64reg_sw : NVPTXInst<(outs Int64Regs:$dst), (ins Int64Regs:$src, //----------------------------------- +// General Comparison +//----------------------------------- + +// General setp instructions +multiclass SETP<string TypeStr, RegisterClass RC, Operand ImmCls> { + def rr : NVPTXInst<(outs Int1Regs:$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:$b))]>; +} + +defm SELP_b16 : SELP_PATTERN<"b16", Int16Regs, i16imm, imm>; +defm SELP_s16 : SELP<"s16", Int16Regs, i16imm>; +defm SELP_u16 : SELP<"u16", Int16Regs, i16imm>; +defm SELP_b32 : SELP_PATTERN<"b32", Int32Regs, i32imm, imm>; +defm SELP_s32 : SELP<"s32", Int32Regs, i32imm>; +defm SELP_u32 : SELP<"u32", Int32Regs, i32imm>; +defm SELP_b64 : SELP_PATTERN<"b64", Int64Regs, i64imm, imm>; +defm SELP_s64 : SELP<"s64", Int64Regs, i64imm>; +defm SELP_u64 : SELP<"u64", Int64Regs, i64imm>; +defm SELP_f32 : SELP_PATTERN<"f32", Float32Regs, f32imm, fpimm>; +defm SELP_f64 : SELP_PATTERN<"f64", Float64Regs, f64imm, fpimm>; + +// Special select for predicate operands +def : Pat<(i1 (select Int1Regs:$p, Int1Regs:$a, Int1Regs:$b)), + (ORb1rr (ANDb1rr Int1Regs:$p, Int1Regs:$a), + (ANDb1rr (NOT1 Int1Regs:$p), Int1Regs:$b))>; + +//----------------------------------- // Data Movement (Load / Store, Move) //----------------------------------- @@ -1253,12 +1321,19 @@ def MOV_ADDR64 : NVPTXInst<(outs Int64Regs:$dst), (ins imem:$a), "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:$sss), @@ -1274,9 +1349,6 @@ def FMOV64rr: NVPTXInst<(outs Float64Regs:$dst), (ins Float64Regs:$src), def IMOV1ri: NVPTXInst<(outs Int1Regs:$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:$src)]>; @@ -1308,440 +1380,194 @@ def LEA_ADDRi64 : NVPTXInst<(outs Int64Regs:$dst), (ins MEMri64:$addr), // Comparison and Selection //----------------------------------- -// Generate string block like -// { -// .reg .pred p; -// setp.gt.s16 p, %a, %b; -// selp.s16 %dst, -1, 0, p; -// } -// when OpcStr=setp.gt.s sz1=16 sz2=16 d=%dst a=%a b=%b -class Set_Str<string OpcStr, string sz1, string sz2, string d, string a, - string b> { - string t1 = "{{\n\t.reg .pred p;\n\t"; - string t2 = !strconcat(t1 , OpcStr); - string t3 = !strconcat(t2 , sz1); - string t4 = !strconcat(t3 , " \tp, "); - string t5 = !strconcat(t4 , a); - string t6 = !strconcat(t5 , ", "); - string t7 = !strconcat(t6 , b); - string t8 = !strconcat(t7 , ";\n\tselp.s"); - string t9 = !strconcat(t8 , sz2); - string t10 = !strconcat(t9, " \t"); - string t11 = !strconcat(t10, d); - string s = !strconcat(t11, ", -1, 0, p;\n\t}}"); +multiclass ISET_FORMAT<PatFrag OpNode, PatLeaf Mode, + Instruction setp_16rr, + Instruction setp_16ri, + Instruction setp_16ir, + Instruction setp_32rr, + Instruction setp_32ri, + Instruction setp_32ir, + Instruction setp_64rr, + Instruction setp_64ri, + Instruction setp_64ir, + Instruction set_16rr, + Instruction set_16ri, + Instruction set_16ir, + Instruction set_32rr, + Instruction set_32ri, + Instruction set_32ir, + Instruction set_64rr, + Instruction set_64ri, + Instruction set_64ir> { + // i16 -> pred + def : Pat<(i1 (OpNode Int16Regs:$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 FSetGT : FSET_FORMAT<setogt, CmpGT, CmpGT_FTZ>; +defm FSetLT : FSET_FORMAT<setolt, CmpLT, CmpLT_FTZ>; +defm FSetGE : FSET_FORMAT<setoge, CmpGE, CmpGE_FTZ>; +defm FSetLE : FSET_FORMAT<setole, CmpLE, CmpLE_FTZ>; +defm FSetEQ : FSET_FORMAT<setoeq, CmpEQ, CmpEQ_FTZ>; +defm FSetNE : 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 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]>; @@ -1751,17 +1577,22 @@ def SDTDeclareParamProfile : SDTypeProfile<0, 3, [SDTCisInt<0>, SDTCisInt<1>, 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, @@ -1776,18 +1607,24 @@ def DeclareRet : SDNode<"NVPTXISD::DeclareRet", SDTDeclareScalarParamProfile, [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, @@ -1804,12 +1641,12 @@ def CallVal : SDNode<"NVPTXISD::CallVal", SDTCallValProfile, [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]>; @@ -1820,7 +1657,7 @@ class LoadParamMemInst<NVPTXRegClass regclass, string opstr> : 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:$b), @@ -1828,35 +1665,57 @@ class LoadParamRegInst<NVPTXRegClass regclass, string opstr> : "\t$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, $val4}};"), + []>; def PrintCallRetInst1 : NVPTXInst<(outs), (ins), "call (retval0), ", @@ -1919,126 +1778,81 @@ def PrintCallUniNoRetInst : NVPTXInst<(outs), (ins), "call.uni ", def LoadParamMemI64 : LoadParamMemInst<Int64Regs, ".b64">; def LoadParamMemI32 : LoadParamMemInst<Int32Regs, ".b32">; def LoadParamMemI16 : LoadParamMemInst<Int16Regs, ".b16">; -def LoadParamMemI8 : LoadParamMemInst<Int8Regs, ".b8">; - -//def LoadParamMemI16 : NVPTXInst<(outs Int16Regs:$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.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:$val)]>; +def StoreRetvalV2F64 : StoreRetvalV2Inst<Float64Regs, ".f64">; +def StoreRetvalV2F32 : StoreRetvalV2Inst<Float32Regs, ".f32">; +def StoreRetvalV4F32 : StoreRetvalV4Inst<Float32Regs, ".f32">; def CallArgBeginInst : NVPTXInst<(outs), (ins), "(", [(CallArgBegin)]>; def CallArgEndInst1 : NVPTXInst<(outs), (ins), ");", [(CallArgEnd (i32 1))]>; @@ -2056,7 +1870,6 @@ class LastCallArgInst<NVPTXRegClass regclass> : def CallArgI64 : CallArgInst<Int64Regs>; def CallArgI32 : CallArgInst<Int32Regs>; def CallArgI16 : CallArgInst<Int16Regs>; -def CallArgI8 : CallArgInst<Int8Regs>; def CallArgF64 : CallArgInst<Float64Regs>; def CallArgF32 : CallArgInst<Float32Regs>; @@ -2064,7 +1877,6 @@ def CallArgF32 : CallArgInst<Float32Regs>; def LastCallArgI64 : LastCallArgInst<Int64Regs>; def LastCallArgI32 : LastCallArgInst<Int32Regs>; def LastCallArgI16 : LastCallArgInst<Int16Regs>; -def LastCallArgI8 : LastCallArgInst<Int8Regs>; def LastCallArgF64 : LastCallArgInst<Float64Regs>; def LastCallArgF32 : LastCallArgInst<Float32Regs>; @@ -2124,9 +1936,6 @@ def MoveParamI32 : MoveParamInst<Int32Regs, ".b32">; def MoveParamI16 : NVPTXInst<(outs Int16Regs:$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:$src))]>; def MoveParamF64 : MoveParamInst<Float64Regs, ".f64">; def MoveParamF32 : MoveParamInst<Float32Regs, ".f32">; @@ -2138,7 +1947,6 @@ class PseudoUseParamInst<NVPTXRegClass regclass> : def PseudoUseParamI64 : PseudoUseParamInst<Int64Regs>; def PseudoUseParamI32 : PseudoUseParamInst<Int32Regs>; def PseudoUseParamI16 : PseudoUseParamInst<Int16Regs>; -def PseudoUseParamI8 : PseudoUseParamInst<Int8Regs>; def PseudoUseParamF64 : PseudoUseParamInst<Float64Regs>; def PseudoUseParamF32 : PseudoUseParamInst<Float32Regs>; @@ -2180,7 +1988,7 @@ multiclass LD<NVPTXRegClass regclass> { } let mayLoad=1, neverHasSideEffects=1 in { -defm LD_i8 : LD<Int8Regs>; +defm LD_i8 : LD<Int16Regs>; defm LD_i16 : LD<Int16Regs>; defm LD_i32 : LD<Int32Regs>; defm LD_i64 : LD<Int64Regs>; @@ -2222,7 +2030,7 @@ multiclass ST<NVPTXRegClass regclass> { } let mayStore=1, neverHasSideEffects=1 in { -defm ST_i8 : ST<Int8Regs>; +defm ST_i8 : ST<Int16Regs>; defm ST_i16 : ST<Int16Regs>; defm ST_i32 : ST<Int32Regs>; defm ST_i64 : ST<Int64Regs>; @@ -2306,7 +2114,7 @@ multiclass LD_VEC<NVPTXRegClass regclass> { []>; } let mayLoad=1, neverHasSideEffects=1 in { -defm LDV_i8 : LD_VEC<Int8Regs>; +defm LDV_i8 : LD_VEC<Int16Regs>; defm LDV_i16 : LD_VEC<Int16Regs>; defm LDV_i32 : LD_VEC<Int32Regs>; defm LDV_i64 : LD_VEC<Int64Regs>; @@ -2389,7 +2197,7 @@ multiclass ST_VEC<NVPTXRegClass regclass> { []>; } let mayStore=1, neverHasSideEffects=1 in { -defm STV_i8 : ST_VEC<Int8Regs>; +defm STV_i8 : ST_VEC<Int16Regs>; defm STV_i16 : ST_VEC<Int16Regs>; defm STV_i32 : ST_VEC<Int32Regs>; defm STV_i64 : ST_VEC<Int64Regs>; @@ -2400,291 +2208,6 @@ defm STV_f64 : ST_VEC<Float64Regs>; //---- Conversion ---- -multiclass CVT_INT_TO_FP <string OpStr, SDNode OpNode> { -// FIXME: need to add f16 support -// def CVTf16i8 : -// NVPTXInst<(outs Float16Regs:$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:$a), @@ -2696,29 +2219,209 @@ def BITCONVERT_32_F2I : F_BITCONVERT<"32", Float32Regs, Int32Regs>; def BITCONVERT_64_I2F : F_BITCONVERT<"64", Int64Regs, Float64Regs>; def BITCONVERT_64_F2I : F_BITCONVERT<"64", Float64Regs, Int64Regs>; +// NOTE: pred->fp are currently sub-optimal due to an issue in TableGen where +// we cannot specify floating-point literals in isel patterns. Therefore, we +// use an integer selp to select either 1 or 0 and then cvt to floating-point. + +// sint -> f32 +def : Pat<(f32 (sint_to_fp Int1Regs:$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, $s2}};", @@ -2733,28 +2436,11 @@ def V2F32toF64 : NVPTXInst<(outs Float64Regs:$d), []>; // 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}}, $s;", @@ -2768,21 +2454,75 @@ def F64toV2F32 : NVPTXInst<(outs Float32Regs:$d1, Float32Regs:$d2), "mov.b64\t{{$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:$a, CvtNONE)>; def retflag : SDNode<"NVPTXISD::RET_FLAG", SDTNone, [SDNPHasChain, SDNPOptInGlue]>; @@ -2810,8 +2550,8 @@ let isTerminator=1 in { [(br bb:$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 @@ -2867,6 +2607,20 @@ def trapinst : NVPTXInst<(outs), (ins), "trap;", [(trap)]>; +// 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" |