R600/SI: Add initial double precision support for SI
Patch by: Niels Ole Salscheider
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@186177 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 7fad3bb..9891ad3 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -60,12 +60,18 @@
setOperationAction(ISD::STORE, MVT::v4f32, Promote);
AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
+ setOperationAction(ISD::STORE, MVT::f64, Promote);
+ AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
+
setOperationAction(ISD::LOAD, MVT::f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
+ setOperationAction(ISD::LOAD, MVT::f64, Promote);
+ AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
+
setOperationAction(ISD::MUL, MVT::i64, Expand);
setOperationAction(ISD::UDIV, MVT::i32, Expand);
diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index a314bc4..4d0fdf3 100644
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -45,6 +45,7 @@
addRegisterClass(MVT::v2i32, &AMDGPU::VReg_64RegClass);
addRegisterClass(MVT::v2f32, &AMDGPU::VReg_64RegClass);
+ addRegisterClass(MVT::f64, &AMDGPU::VReg_64RegClass);
addRegisterClass(MVT::v4i32, &AMDGPU::VReg_128RegClass);
addRegisterClass(MVT::v4f32, &AMDGPU::VReg_128RegClass);
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 5a1bf30..8436b67 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -663,7 +663,9 @@
[(set f32:$dst, (int_AMDGPU_rsq f32:$src0))]
>;
defm V_RSQ_F32 : VOP1_32 <0x0000002e, "V_RSQ_F32", []>;
-defm V_RCP_F64 : VOP1_64 <0x0000002f, "V_RCP_F64", []>;
+defm V_RCP_F64 : VOP1_64 <0x0000002f, "V_RCP_F64",
+ [(set f64:$dst, (fdiv FP_ONE, f64:$src0))]
+>;
defm V_RCP_CLAMP_F64 : VOP1_64 <0x00000030, "V_RCP_CLAMP_F64", []>;
defm V_RSQ_F64 : VOP1_64 <0x00000031, "V_RSQ_F64", []>;
defm V_RSQ_CLAMP_F64 : VOP1_64 <0x00000032, "V_RSQ_CLAMP_F64", []>;
@@ -1008,10 +1010,25 @@
>;
def V_ASHR_I64 : VOP3_64_Shift <0x00000163, "V_ASHR_I64", []>;
+let isCommutable = 1 in {
+
def V_ADD_F64 : VOP3_64 <0x00000164, "V_ADD_F64", []>;
def V_MUL_F64 : VOP3_64 <0x00000165, "V_MUL_F64", []>;
def V_MIN_F64 : VOP3_64 <0x00000166, "V_MIN_F64", []>;
def V_MAX_F64 : VOP3_64 <0x00000167, "V_MAX_F64", []>;
+
+} // isCommutable = 1
+
+def : Pat <
+ (fadd f64:$src0, f64:$src1),
+ (V_ADD_F64 $src0, $src1, (i64 0))
+>;
+
+def : Pat <
+ (fmul f64:$src0, f64:$src1),
+ (V_MUL_F64 $src0, $src1, (i64 0))
+>;
+
def V_LDEXP_F64 : VOP3_64 <0x00000168, "V_LDEXP_F64", []>;
let isCommutable = 1 in {
@@ -1434,6 +1451,10 @@
def : BitConvert <f32, i32, SReg_32>;
def : BitConvert <f32, i32, VReg_32>;
+def : BitConvert <i64, f64, VReg_64>;
+
+def : BitConvert <f64, i64, VReg_64>;
+
/********** =================== **********/
/********** Src & Dst modifiers **********/
/********** =================== **********/
@@ -1522,6 +1543,11 @@
(V_MUL_F32_e32 $src0, (V_RCP_F32_e32 $src1))
>;
+def : Pat<
+ (fdiv f64:$src0, f64:$src1),
+ (V_MUL_F64 $src0, (V_RCP_F64_e32 $src1), (i64 0))
+>;
+
def : Pat <
(fcos f32:$src0),
(V_COS_F32_e32 (V_MUL_F32_e32 $src0, (V_MOV_B32_e32 CONST.TWO_PI_INV)))
@@ -1672,6 +1698,8 @@
>;
}
+defm : MUBUFLoad_Pattern <BUFFER_LOAD_DWORDX2_ADDR64, i64,
+ global_load, constant_load>;
defm : MUBUFLoad_Pattern <BUFFER_LOAD_DWORD_ADDR64, i32,
global_load, constant_load>;
defm : MUBUFLoad_Pattern <BUFFER_LOAD_UBYTE_ADDR64, i32,