def hasLDU : Predicate<"Subtarget.hasLDU()">;
def hasGenericLdSt : Predicate<"Subtarget.hasGenericLdSt()">;
-def doF32FTZ : Predicate<"UseF32FTZ==1">;
-def doNoF32FTZ : Predicate<"UseF32FTZ==0">;
+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()">;
// 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.
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, $src;", []>;
+}
+
//-----------------------------------
// Integer Arithmetic
//-----------------------------------
// If we reverse the order of the following two lines, then rrr2 rule will be
// generated for FMA32, but not for rrr.
// Therefore, we manually write the rrr2 rule in FPCONTRACT32.
-defm FMAD32_ftz : FPCONTRACT32<"mad.ftz.f32", doFMADF32_ftz>;
-defm FMAD32 : FPCONTRACT32<"mad.f32", doFMADF32>;
defm FMA32_ftz : FPCONTRACT32<"fma.rn.ftz.f32", doFMAF32_ftz>;
defm FMA32 : FPCONTRACT32<"fma.rn.f32", doFMAF32>;
defm FMA64 : FPCONTRACT64<"fma.rn.f64", doFMAF64>;
defm FMAF32ext_ftz : FPCONTRACT32_SUB_PAT<FMA32_ftzrrr, doFMAF32AGG_ftz>;
defm FMAF32ext : FPCONTRACT32_SUB_PAT<FMA32rrr, doFMAF32AGG>;
-defm FMADF32ext_ftz : FPCONTRACT32_SUB_PAT_MAD<FMAD32_ftzrrr, doFMADF32_ftz>;
-defm FMADF32ext : FPCONTRACT32_SUB_PAT_MAD<FMAD32rrr, doFMADF32>;
defm FMAF64ext : FPCONTRACT64_SUB_PAT<FMA64rrr, doFMAF64AGG>;
def SINF: NVPTXInst<(outs Float32Regs:$dst), (ins Float32Regs:$src),
"mov.u64 \t$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:$sss),
//---- Conversion ----
+class F_BITCONVERT<string SzStr, NVPTXRegClass regclassIn,
+ NVPTXRegClass regclassOut> :
+ NVPTXInst<(outs regclassOut:$d), (ins regclassIn:$a),
+ !strconcat("mov.b", !strconcat(SzStr, " \t $d, $a;")),
+ [(set regclassOut:$d, (bitconvert regclassIn:$a))]>;
+
+def BITCONVERT_32_I2F : F_BITCONVERT<"32", Int32Regs, Float32Regs>;
+def BITCONVERT_32_F2I : F_BITCONVERT<"32", Float32Regs, Int32Regs>;
+def BITCONVERT_64_I2F : F_BITCONVERT<"64", Int64Regs, Float64Regs>;
+def BITCONVERT_64_F2I : F_BITCONVERT<"64", Float64Regs, Int64Regs>;
+
// NOTE: pred->fp are currently sub-optimal due to an issue in TableGen where
// we cannot specify floating-point literals in isel patterns. Therefore, we
// use an integer selp to select either 1 or 0 and then cvt to floating-point.
// f32 -> sint
+def : Pat<(i1 (fp_to_sint Float32Regs:$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_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_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_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)),
// anyext i1
def : Pat<(i16 (anyext Int1Regs:$a)),
- (SELP_u16ii 1, 0, Int1Regs:$a)>;
+ (SELP_u16ii -1, 0, Int1Regs:$a)>;
def : Pat<(i32 (anyext Int1Regs:$a)),
- (SELP_u32ii 1, 0, Int1Regs:$a)>;
+ (SELP_u32ii -1, 0, Int1Regs:$a)>;
def : Pat<(i64 (anyext Int1Regs:$a)),
- (SELP_u64ii 1, 0, Int1Regs:$a)>;
+ (SELP_u64ii -1, 0, Int1Regs:$a)>;
// sext i16
def : Pat<(i32 (sext Int16Regs:$a)),
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:$b),
(SETP_b32ri (ANDb32ri Int32Regs:$pred, 1), 1, CmpEQ))>;
-class F_BITCONVERT<string SzStr, NVPTXRegClass regclassIn,
- NVPTXRegClass regclassOut> :
- NVPTXInst<(outs regclassOut:$d), (ins regclassIn:$a),
- !strconcat("mov.b", !strconcat(SzStr, " \t $d, $a;")),
- [(set regclassOut:$d, (bitconvert regclassIn:$a))]>;
-
-def BITCONVERT_32_I2F : F_BITCONVERT<"32", Int32Regs, Float32Regs>;
-def BITCONVERT_32_F2I : F_BITCONVERT<"32", Float32Regs, Int32Regs>;
-def BITCONVERT_64_I2F : F_BITCONVERT<"64", Int64Regs, Float64Regs>;
-def BITCONVERT_64_F2I : F_BITCONVERT<"64", Float64Regs, Int64Regs>;
-
// pack a set of smaller int registers to a larger int register
def V4I16toI64 : NVPTXInst<(outs Int64Regs:$d),
(ins Int16Regs:$s1, Int16Regs:$s2,
projects / oota-llvm.git / blobdiff