def Use64BitAddresses : Predicate<"getSubtarget().is64Bit()">;
// Shader Model Support
-def SupportsSM13 : Predicate<"getSubtarget().supportsSM13()">;
-def DoesNotSupportSM13 : Predicate<"!getSubtarget().supportsSM13()">;
-def SupportsSM20 : Predicate<"getSubtarget().supportsSM20()">;
-def DoesNotSupportSM20 : Predicate<"!getSubtarget().supportsSM20()">;
+def FDivNeedsRoundingMode : Predicate<"getSubtarget().fdivNeedsRoundingMode()">;
+def FDivNoRoundingMode : Predicate<"!getSubtarget().fdivNeedsRoundingMode()">;
+def FMadNeedsRoundingMode : Predicate<"getSubtarget().fmadNeedsRoundingMode()">;
+def FMadNoRoundingMode : Predicate<"!getSubtarget().fmadNeedsRoundingMode()">;
// PTX Version Support
def SupportsPTX21 : Predicate<"getSubtarget().supportsPTX21()">;
let PrintMethod = "printParamOperand";
let MIOperandInfo = (ops i32imm);
}
+def MEMret : Operand<i32> {
+ let PrintMethod = "printReturnOperand";
+ let MIOperandInfo = (ops i32imm);
+}
// Branch & call targets have OtherVT type.
def brtarget : Operand<OtherVT>;
def PTXcopyaddress
: SDNode<"PTXISD::COPY_ADDRESS", SDTypeProfile<1, 1, []>, []>;
+// Load/store .param space
+def PTXloadparam
+ : SDNode<"PTXISD::LOAD_PARAM", SDTypeProfile<1, 1, [SDTCisVT<1, i32>]>,
+ [SDNPHasChain, SDNPOutGlue, SDNPOptInGlue]>;
def PTXstoreparam
: SDNode<"PTXISD::STORE_PARAM", SDTypeProfile<0, 2, [SDTCisVT<0, i32>]>,
[SDNPHasChain, SDNPOutGlue, SDNPOptInGlue]>;
(ins RegF32:$a, RegF32:$b),
"div.rn.f32\t$d, $a, $b",
[(set RegF32:$d, (fdiv RegF32:$a, RegF32:$b))]>,
- Requires<[SupportsSM13]>;
+ Requires<[FDivNeedsRoundingMode]>;
def FDIVri32SM13 : InstPTX<(outs RegF32:$d),
(ins RegF32:$a, f32imm:$b),
"div.rn.f32\t$d, $a, $b",
[(set RegF32:$d, (fdiv RegF32:$a, fpimm:$b))]>,
- Requires<[SupportsSM13]>;
+ Requires<[FDivNeedsRoundingMode]>;
def FDIVrr32SM10 : InstPTX<(outs RegF32:$d),
(ins RegF32:$a, RegF32:$b),
"div.f32\t$d, $a, $b",
[(set RegF32:$d, (fdiv RegF32:$a, RegF32:$b))]>,
- Requires<[DoesNotSupportSM13]>;
+ Requires<[FDivNoRoundingMode]>;
def FDIVri32SM10 : InstPTX<(outs RegF32:$d),
(ins RegF32:$a, f32imm:$b),
"div.f32\t$d, $a, $b",
[(set RegF32:$d, (fdiv RegF32:$a, fpimm:$b))]>,
- Requires<[DoesNotSupportSM13]>;
+ Requires<[FDivNoRoundingMode]>;
def FDIVrr64SM13 : InstPTX<(outs RegF64:$d),
(ins RegF64:$a, RegF64:$b),
"div.rn.f64\t$d, $a, $b",
[(set RegF64:$d, (fdiv RegF64:$a, RegF64:$b))]>,
- Requires<[SupportsSM13]>;
+ Requires<[FDivNeedsRoundingMode]>;
def FDIVri64SM13 : InstPTX<(outs RegF64:$d),
(ins RegF64:$a, f64imm:$b),
"div.rn.f64\t$d, $a, $b",
[(set RegF64:$d, (fdiv RegF64:$a, fpimm:$b))]>,
- Requires<[SupportsSM13]>;
+ Requires<[FDivNeedsRoundingMode]>;
def FDIVrr64SM10 : InstPTX<(outs RegF64:$d),
(ins RegF64:$a, RegF64:$b),
"div.f64\t$d, $a, $b",
[(set RegF64:$d, (fdiv RegF64:$a, RegF64:$b))]>,
- Requires<[DoesNotSupportSM13]>;
+ Requires<[FDivNoRoundingMode]>;
def FDIVri64SM10 : InstPTX<(outs RegF64:$d),
(ins RegF64:$a, f64imm:$b),
"div.f64\t$d, $a, $b",
[(set RegF64:$d, (fdiv RegF64:$a, fpimm:$b))]>,
- Requires<[DoesNotSupportSM13]>;
+ Requires<[FDivNoRoundingMode]>;
// In the short term, mad is supported on all PTX versions and we use a
// default rounding mode no matter what shader model or PTX version.
// TODO: Allow the rounding mode to be selectable through llc.
-defm FMADSM13 : PTX_FLOAT_4OP<"mad.rn", fmul, fadd>, Requires<[SupportsSM13, SupportsFMA]>;
-defm FMAD : PTX_FLOAT_4OP<"mad", fmul, fadd>, Requires<[DoesNotSupportSM13, SupportsFMA]>;
+defm FMADSM13 : PTX_FLOAT_4OP<"mad.rn", fmul, fadd>,
+ Requires<[FMadNeedsRoundingMode, SupportsFMA]>;
+defm FMAD : PTX_FLOAT_4OP<"mad", fmul, fadd>,
+ Requires<[FMadNoRoundingMode, SupportsFMA]>;
///===- Floating-Point Intrinsic Instructions -----------------------------===//
defm LDl : PTX_LD_ALL<"ld.local", load_local>;
defm LDs : PTX_LD_ALL<"ld.shared", load_shared>;
-// This is a special instruction that is manually inserted for parameters
-def LDpiU16 : InstPTX<(outs RegI16:$d), (ins MEMpi:$a),
- "ld.param.u16\t$d, [$a]", []>;
-def LDpiU32 : InstPTX<(outs RegI32:$d), (ins MEMpi:$a),
- "ld.param.u32\t$d, [$a]", []>;
-def LDpiU64 : InstPTX<(outs RegI64:$d), (ins MEMpi:$a),
- "ld.param.u64\t$d, [$a]", []>;
-def LDpiF32 : InstPTX<(outs RegF32:$d), (ins MEMpi:$a),
- "ld.param.f32\t$d, [$a]", []>;
-def LDpiF64 : InstPTX<(outs RegF64:$d), (ins MEMpi:$a),
- "ld.param.f64\t$d, [$a]", []>;
-
-// def STpiPred : InstPTX<(outs), (ins i1imm:$d, RegPred:$a),
-// "st.param.pred\t[$d], $a",
-// [(PTXstoreparam imm:$d, RegPred:$a)]>;
-// def STpiU16 : InstPTX<(outs), (ins i16imm:$d, RegI16:$a),
-// "st.param.u16\t[$d], $a",
-// [(PTXstoreparam imm:$d, RegI16:$a)]>;
-def STpiU32 : InstPTX<(outs), (ins i32imm:$d, RegI32:$a),
- "st.param.u32\t[$d], $a",
- [(PTXstoreparam timm:$d, RegI32:$a)]>;
-// def STpiU64 : InstPTX<(outs), (ins i64imm:$d, RegI64:$a),
-// "st.param.u64\t[$d], $a",
-// [(PTXstoreparam imm:$d, RegI64:$a)]>;
-// def STpiF32 : InstPTX<(outs), (ins MEMpi:$d, RegF32:$a),
-// "st.param.f32\t[$d], $a", []>;
-// def STpiF64 : InstPTX<(outs), (ins MEMpi:$d, RegF64:$a),
-// "st.param.f64\t[$d], $a", []>;
+// These instructions are used to load/store from the .param space for
+// device and kernel parameters
+
+let hasSideEffects = 1 in {
+ def LDpiPred : InstPTX<(outs RegPred:$d), (ins MEMpi:$a),
+ "ld.param.pred\t$d, [$a]",
+ [(set RegPred:$d, (PTXloadparam timm:$a))]>;
+ def LDpiU16 : InstPTX<(outs RegI16:$d), (ins MEMpi:$a),
+ "ld.param.u16\t$d, [$a]",
+ [(set RegI16:$d, (PTXloadparam timm:$a))]>;
+ def LDpiU32 : InstPTX<(outs RegI32:$d), (ins MEMpi:$a),
+ "ld.param.u32\t$d, [$a]",
+ [(set RegI32:$d, (PTXloadparam timm:$a))]>;
+ def LDpiU64 : InstPTX<(outs RegI64:$d), (ins MEMpi:$a),
+ "ld.param.u64\t$d, [$a]",
+ [(set RegI64:$d, (PTXloadparam timm:$a))]>;
+ def LDpiF32 : InstPTX<(outs RegF32:$d), (ins MEMpi:$a),
+ "ld.param.f32\t$d, [$a]",
+ [(set RegF32:$d, (PTXloadparam timm:$a))]>;
+ def LDpiF64 : InstPTX<(outs RegF64:$d), (ins MEMpi:$a),
+ "ld.param.f64\t$d, [$a]",
+ [(set RegF64:$d, (PTXloadparam timm:$a))]>;
+
+ def STpiPred : InstPTX<(outs), (ins MEMret:$d, RegPred:$a),
+ "st.param.pred\t[$d], $a",
+ [(PTXstoreparam timm:$d, RegPred:$a)]>;
+ def STpiU16 : InstPTX<(outs), (ins MEMret:$d, RegI16:$a),
+ "st.param.u16\t[$d], $a",
+ [(PTXstoreparam timm:$d, RegI16:$a)]>;
+ def STpiU32 : InstPTX<(outs), (ins MEMret:$d, RegI32:$a),
+ "st.param.u32\t[$d], $a",
+ [(PTXstoreparam timm:$d, RegI32:$a)]>;
+ def STpiU64 : InstPTX<(outs), (ins MEMret:$d, RegI64:$a),
+ "st.param.u64\t[$d], $a",
+ [(PTXstoreparam timm:$d, RegI64:$a)]>;
+ def STpiF32 : InstPTX<(outs), (ins MEMret:$d, RegF32:$a),
+ "st.param.f32\t[$d], $a",
+ [(PTXstoreparam timm:$d, RegF32:$a)]>;
+ def STpiF64 : InstPTX<(outs), (ins MEMret:$d, RegF64:$a),
+ "st.param.f64\t[$d], $a",
+ [(PTXstoreparam timm:$d, RegF64:$a)]>;
+}
// Stores
defm STg : PTX_ST_ALL<"st.global", store_global>;
projects / oota-llvm.git / blobdiff