[llvm] r206547 - R600/SI: Match sign_extend_inreg to s_sext_i32_i8 and s_sext_i32_i16
Matt Arsenault
Matthew.Arsenault at amd.com
Thu Apr 17 18:53:19 PDT 2014
Author: arsenm
Date: Thu Apr 17 20:53:18 2014
New Revision: 206547
URL: http://llvm.org/viewvc/llvm-project?rev=206547&view=rev
Log:
R600/SI: Match sign_extend_inreg to s_sext_i32_i8 and s_sext_i32_i16
Modified:
llvm/trunk/lib/Target/R600/SIISelLowering.cpp
llvm/trunk/lib/Target/R600/SIInstrInfo.cpp
llvm/trunk/lib/Target/R600/SIInstrInfo.h
llvm/trunk/lib/Target/R600/SIInstructions.td
llvm/trunk/test/CodeGen/R600/sext-in-reg.ll
Modified: llvm/trunk/lib/Target/R600/SIISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIISelLowering.cpp?rev=206547&r1=206546&r2=206547&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/R600/SIISelLowering.cpp Thu Apr 17 20:53:18 2014
@@ -165,6 +165,9 @@ SITargetLowering::SITargetLowering(Targe
setOperationAction(ISD::LOAD, MVT::i1, Custom);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Legal);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
+
setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
setOperationAction(ISD::FrameIndex, MVT::i32, Custom);
Modified: llvm/trunk/lib/Target/R600/SIInstrInfo.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstrInfo.cpp?rev=206547&r1=206546&r2=206547&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstrInfo.cpp (original)
+++ llvm/trunk/lib/Target/R600/SIInstrInfo.cpp Thu Apr 17 20:53:18 2014
@@ -537,6 +537,8 @@ unsigned SIInstrInfo::getVALUOp(const Ma
case AMDGPU::S_LSHL_B64: return AMDGPU::V_LSHL_B64;
case AMDGPU::S_LSHR_B32: return AMDGPU::V_LSHR_B32_e32;
case AMDGPU::S_LSHR_B64: return AMDGPU::V_LSHR_B64;
+ case AMDGPU::S_SEXT_I32_I8: return AMDGPU::V_BFE_I32;
+ case AMDGPU::S_SEXT_I32_I16: return AMDGPU::V_BFE_I32;
case AMDGPU::S_NOT_B32: return AMDGPU::V_NOT_B32_e32;
case AMDGPU::S_CMP_EQ_I32: return AMDGPU::V_CMP_EQ_I32_e32;
case AMDGPU::S_CMP_LG_I32: return AMDGPU::V_CMP_NE_I32_e32;
@@ -915,8 +917,10 @@ void SIInstrInfo::moveToVALU(MachineInst
MachineBasicBlock *MBB = Inst->getParent();
MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
+ unsigned Opcode = Inst->getOpcode();
+
// Handle some special cases
- switch(Inst->getOpcode()) {
+ switch (Opcode) {
case AMDGPU::S_MOV_B64: {
DebugLoc DL = Inst->getDebugLoc();
@@ -988,26 +992,28 @@ void SIInstrInfo::moveToVALU(MachineInst
Inst->RemoveOperand(i);
}
- // Add the implict and explicit register definitions.
- if (NewDesc.ImplicitUses) {
- for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
- unsigned Reg = NewDesc.ImplicitUses[i];
- Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
- }
+ if (Opcode == AMDGPU::S_SEXT_I32_I8 || Opcode == AMDGPU::S_SEXT_I32_I16) {
+ // We are converting these to a BFE, so we need to add the missing
+ // operands for the size and offset.
+ unsigned Size = (Opcode == AMDGPU::S_SEXT_I32_I8) ? 8 : 16;
+ Inst->addOperand(MachineOperand::CreateImm(0));
+ Inst->addOperand(MachineOperand::CreateImm(Size));
+
+ // XXX - Other pointless operands. There are 4, but it seems you only need
+ // 3 to not hit an assertion later in MCInstLower.
+ Inst->addOperand(MachineOperand::CreateImm(0));
+ Inst->addOperand(MachineOperand::CreateImm(0));
+ Inst->addOperand(MachineOperand::CreateImm(0));
+ Inst->addOperand(MachineOperand::CreateImm(0));
}
- if (NewDesc.ImplicitDefs) {
- for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
- unsigned Reg = NewDesc.ImplicitDefs[i];
- Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
- }
- }
+ addDescImplicitUseDef(NewDesc, Inst);
// Update the destination register class.
const TargetRegisterClass *NewDstRC = getOpRegClass(*Inst, 0);
- switch (Inst->getOpcode()) {
+ switch (Opcode) {
// For target instructions, getOpRegClass just returns the virtual
// register class associated with the operand, so we need to find an
// equivalent VGPR register class in order to move the instruction to the
@@ -1120,6 +1126,24 @@ void SIInstrInfo::splitScalar64BitOp(Sma
Worklist.push_back(HiHalf);
}
+void SIInstrInfo::addDescImplicitUseDef(const MCInstrDesc &NewDesc,
+ MachineInstr *Inst) const {
+ // Add the implict and explicit register definitions.
+ if (NewDesc.ImplicitUses) {
+ for (unsigned i = 0; NewDesc.ImplicitUses[i]; ++i) {
+ unsigned Reg = NewDesc.ImplicitUses[i];
+ Inst->addOperand(MachineOperand::CreateReg(Reg, false, true));
+ }
+ }
+
+ if (NewDesc.ImplicitDefs) {
+ for (unsigned i = 0; NewDesc.ImplicitDefs[i]; ++i) {
+ unsigned Reg = NewDesc.ImplicitDefs[i];
+ Inst->addOperand(MachineOperand::CreateReg(Reg, true, true));
+ }
+ }
+}
+
MachineInstrBuilder SIInstrInfo::buildIndirectWrite(
MachineBasicBlock *MBB,
MachineBasicBlock::iterator I,
Modified: llvm/trunk/lib/Target/R600/SIInstrInfo.h
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstrInfo.h?rev=206547&r1=206546&r2=206547&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstrInfo.h (original)
+++ llvm/trunk/lib/Target/R600/SIInstrInfo.h Thu Apr 17 20:53:18 2014
@@ -47,6 +47,7 @@ private:
void splitScalar64BitOp(SmallVectorImpl<MachineInstr *> & Worklist,
MachineInstr *Inst, unsigned Opcode) const;
+ void addDescImplicitUseDef(const MCInstrDesc &Desc, MachineInstr *MI) const;
public:
explicit SIInstrInfo(AMDGPUTargetMachine &tm);
Modified: llvm/trunk/lib/Target/R600/SIInstructions.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/R600/SIInstructions.td?rev=206547&r1=206546&r2=206547&view=diff
==============================================================================
--- llvm/trunk/lib/Target/R600/SIInstructions.td (original)
+++ llvm/trunk/lib/Target/R600/SIInstructions.td Thu Apr 17 20:53:18 2014
@@ -68,8 +68,12 @@ def S_BREV_B64 : SOP1_64 <0x0000000c, "S
//def S_FLBIT_I32_B64 : SOP1_32 <0x00000016, "S_FLBIT_I32_B64", []>;
def S_FLBIT_I32 : SOP1_32 <0x00000017, "S_FLBIT_I32", []>;
//def S_FLBIT_I32_I64 : SOP1_32 <0x00000018, "S_FLBIT_I32_I64", []>;
-//def S_SEXT_I32_I8 : SOP1_32 <0x00000019, "S_SEXT_I32_I8", []>;
-//def S_SEXT_I32_I16 : SOP1_32 <0x0000001a, "S_SEXT_I32_I16", []>;
+def S_SEXT_I32_I8 : SOP1_32 <0x00000019, "S_SEXT_I32_I8",
+ [(set i32:$dst, (sext_inreg i32:$src0, i8))]
+>;
+def S_SEXT_I32_I16 : SOP1_32 <0x0000001a, "S_SEXT_I32_I16",
+ [(set i32:$dst, (sext_inreg i32:$src0, i16))]
+>;
////def S_BITSET0_B32 : SOP1_BITSET0 <0x0000001b, "S_BITSET0_B32", []>;
////def S_BITSET0_B64 : SOP1_BITSET0 <0x0000001c, "S_BITSET0_B64", []>;
////def S_BITSET1_B32 : SOP1_BITSET1 <0x0000001d, "S_BITSET1_B32", []>;
@@ -124,6 +128,21 @@ def S_CMPK_EQ_I32 : SOPK <
>;
*/
+// Handle sext_inreg in i64
+def : Pat <
+ (i64 (sext_inreg i64:$src, i8)),
+ (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
+ (S_SEXT_I32_I8 (EXTRACT_SUBREG i64:$src, sub0)), sub0),
+ (S_MOV_B32 -1), sub1)
+>;
+
+def : Pat <
+ (i64 (sext_inreg i64:$src, i16)),
+ (INSERT_SUBREG (INSERT_SUBREG (i64 (IMPLICIT_DEF)),
+ (S_SEXT_I32_I16 (EXTRACT_SUBREG i64:$src, sub0)), sub0),
+ (S_MOV_B32 -1), sub1)
+>;
+
let isCompare = 1 in {
def S_CMPK_LG_I32 : SOPK_32 <0x00000004, "S_CMPK_LG_I32", []>;
def S_CMPK_GT_I32 : SOPK_32 <0x00000005, "S_CMPK_GT_I32", []>;
Modified: llvm/trunk/test/CodeGen/R600/sext-in-reg.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/R600/sext-in-reg.ll?rev=206547&r1=206546&r2=206547&view=diff
==============================================================================
--- llvm/trunk/test/CodeGen/R600/sext-in-reg.ll (original)
+++ llvm/trunk/test/CodeGen/R600/sext-in-reg.ll Thu Apr 17 20:53:18 2014
@@ -21,8 +21,9 @@ define void @sext_in_reg_i1_i32(i32 addr
; FUNC-LABEL: @sext_in_reg_i8_to_i32
; SI: S_ADD_I32 [[VAL:s[0-9]+]],
-; SI: V_BFE_I32 [[EXTRACT:v[0-9]+]], [[VAL]], 0, 8
-; SI: BUFFER_STORE_DWORD [[EXTRACT]],
+; SI: S_SEXT_I32_I8 [[EXTRACT:s[0-9]+]], [[VAL]]
+; SI: V_MOV_B32_e32 [[VEXTRACT:v[0-9]+]], [[EXTRACT]]
+; SI: BUFFER_STORE_DWORD [[VEXTRACT]],
; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+\.[XYZW]]], [[ADDR:T[0-9]+.[XYZW]]]
; EG: ADD_INT
@@ -38,8 +39,9 @@ define void @sext_in_reg_i8_to_i32(i32 a
; FUNC-LABEL: @sext_in_reg_i16_to_i32
; SI: S_ADD_I32 [[VAL:s[0-9]+]],
-; SI: V_BFE_I32 [[EXTRACT:v[0-9]+]], [[VAL]], 0, 16
-; SI: BUFFER_STORE_DWORD [[EXTRACT]],
+; SI: S_SEXT_I32_I16 [[EXTRACT:s[0-9]+]], [[VAL]]
+; SI: V_MOV_B32_e32 [[VEXTRACT:v[0-9]+]], [[EXTRACT]]
+; SI: BUFFER_STORE_DWORD [[VEXTRACT]],
; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+\.[XYZW]]], [[ADDR:T[0-9]+.[XYZW]]]
; EG: ADD_INT
@@ -55,8 +57,9 @@ define void @sext_in_reg_i16_to_i32(i32
; FUNC-LABEL: @sext_in_reg_i8_to_v1i32
; SI: S_ADD_I32 [[VAL:s[0-9]+]],
-; SI: V_BFE_I32 [[EXTRACT:v[0-9]+]], [[VAL]], 0, 8
-; SI: BUFFER_STORE_DWORD [[EXTRACT]],
+; SI: S_SEXT_I32_I8 [[EXTRACT:s[0-9]+]], [[VAL]]
+; SI: V_MOV_B32_e32 [[VEXTRACT:v[0-9]+]], [[EXTRACT]]
+; SI: BUFFER_STORE_DWORD [[VEXTRACT]],
; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+\.[XYZW]]], [[ADDR:T[0-9]+.[XYZW]]]
; EG: ADD_INT
@@ -71,9 +74,10 @@ define void @sext_in_reg_i8_to_v1i32(<1
}
; FUNC-LABEL: @sext_in_reg_i8_to_i64
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
-; SI: V_ASHRREV_I32_e32 {{v[0-9]+}}, 31,
-; SI: BUFFER_STORE_DWORD
+; SI: S_ADD_I32 [[VAL:s[0-9]+]],
+; SI: S_SEXT_I32_I8 [[EXTRACT:s[0-9]+]], [[VAL]]
+; SI: S_MOV_B32 {{s[0-9]+}}, -1
+; SI: BUFFER_STORE_DWORDX2
; EG: MEM_{{.*}} STORE_{{.*}} [[RES_LO:T[0-9]+\.[XYZW]]], [[ADDR_LO:T[0-9]+.[XYZW]]]
; EG: MEM_{{.*}} STORE_{{.*}} [[RES_HI:T[0-9]+\.[XYZW]]], [[ADDR_HI:T[0-9]+.[XYZW]]]
@@ -94,9 +98,10 @@ define void @sext_in_reg_i8_to_i64(i64 a
}
; FUNC-LABEL: @sext_in_reg_i16_to_i64
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 16
-; SI: V_ASHRREV_I32_e32 {{v[0-9]+}}, 31,
-; SI: BUFFER_STORE_DWORD
+; SI: S_ADD_I32 [[VAL:s[0-9]+]],
+; SI: S_SEXT_I32_I16 [[EXTRACT:s[0-9]+]], [[VAL]]
+; SI: S_MOV_B32 {{s[0-9]+}}, -1
+; SI: BUFFER_STORE_DWORDX2
; EG: MEM_{{.*}} STORE_{{.*}} [[RES_LO:T[0-9]+\.[XYZW]]], [[ADDR_LO:T[0-9]+.[XYZW]]]
; EG: MEM_{{.*}} STORE_{{.*}} [[RES_HI:T[0-9]+\.[XYZW]]], [[ADDR_HI:T[0-9]+.[XYZW]]]
@@ -237,8 +242,8 @@ define void @sext_in_reg_v4i1_to_v4i32(<
}
; FUNC-LABEL: @sext_in_reg_v2i8_to_v2i32
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
+; SI: S_SEXT_I32_I8 {{s[0-9]+}}, {{s[0-9]+}}
+; SI: S_SEXT_I32_I8 {{s[0-9]+}}, {{s[0-9]+}}
; SI: BUFFER_STORE_DWORDX2
; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+]]{{\.[XYZW][XYZW]}}, [[ADDR:T[0-9]+.[XYZW]]]
@@ -254,10 +259,10 @@ define void @sext_in_reg_v2i8_to_v2i32(<
}
; FUNC-LABEL: @sext_in_reg_v4i8_to_v4i32
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
+; SI: S_SEXT_I32_I8 {{s[0-9]+}}, {{s[0-9]+}}
+; SI: S_SEXT_I32_I8 {{s[0-9]+}}, {{s[0-9]+}}
+; SI: S_SEXT_I32_I8 {{s[0-9]+}}, {{s[0-9]+}}
+; SI: S_SEXT_I32_I8 {{s[0-9]+}}, {{s[0-9]+}}
; SI: BUFFER_STORE_DWORDX4
; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+]]{{\.[XYZW][XYZW][XYZW][XYZW]}}, [[ADDR:T[0-9]+.[XYZW]]]
@@ -275,8 +280,8 @@ define void @sext_in_reg_v4i8_to_v4i32(<
}
; FUNC-LABEL: @sext_in_reg_v2i16_to_v2i32
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
-; SI: V_BFE_I32 {{v[0-9]+}}, {{s[0-9]+}}, 0, 8
+; SI: S_SEXT_I32_I16 {{s[0-9]+}}, {{s[0-9]+}}
+; SI: S_SEXT_I32_I16 {{s[0-9]+}}, {{s[0-9]+}}
; SI: BUFFER_STORE_DWORDX2
; EG: MEM_{{.*}} STORE_{{.*}} [[RES:T[0-9]+]]{{\.[XYZW][XYZW]}}, [[ADDR:T[0-9]+.[XYZW]]]
@@ -285,8 +290,8 @@ define void @sext_in_reg_v4i8_to_v4i32(<
; EG: LSHR {{\*?}} [[ADDR]]
define void @sext_in_reg_v2i16_to_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> %a, <2 x i32> %b) nounwind {
%c = add <2 x i32> %a, %b ; add to prevent folding into extload
- %shl = shl <2 x i32> %c, <i32 24, i32 24>
- %ashr = ashr <2 x i32> %shl, <i32 24, i32 24>
+ %shl = shl <2 x i32> %c, <i32 16, i32 16>
+ %ashr = ashr <2 x i32> %shl, <i32 16, i32 16>
store <2 x i32> %ashr, <2 x i32> addrspace(1)* %out, align 8
ret void
}
More information about the llvm-commits
mailing list