replica
/
hanchenye-llvm-project
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[X86][BTVER2] Reduce instregex usage (PR35955) 

Most are just replaced with instrs lists, but a few regexps have been further generalized to match more instructions with a single pattern.

llvm-svn: 322734
This commit is contained in:
Simon Pilgrim 2018-01-17 19:12:48 +00:00
parent 218a0b51dd
commit 8c87a2e7bd
1 changed files with 29 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

54
llvm/lib/Target/X86/X86ScheduleBtVer2.td
View File

@ -140,24 +140,26 @@ def WriteSHLDrri : SchedWriteRes<[JALU01]> {
let ResourceCycles = [6];
let NumMicroOps = 6;
}
def: InstRW<[WriteSHLDrri], (instregex "SHLD(16|32|64)rri8")>;
def: InstRW<[WriteSHLDrri], (instregex "SHRD(16|32|64)rri8")>;
def: InstRW<[WriteSHLDrri], (instrs SHLD16rri8, SHLD32rri8, SHLD64rri8,
SHRD16rri8, SHRD32rri8, SHRD64rri8)>;
def WriteSHLDrrCL : SchedWriteRes<[JALU01]> {
let Latency = 4;
let ResourceCycles = [8];
let NumMicroOps = 7;
}
def: InstRW<[WriteSHLDrrCL], (instregex "SHLD(16|32|64)rrCL")>;
def: InstRW<[WriteSHLDrrCL], (instregex "SHRD(16|32|64)rrCL")>;
def: InstRW<[WriteSHLDrrCL], (instrs SHLD16rrCL, SHLD32rrCL, SHLD64rrCL,
SHRD16rrCL, SHRD32rrCL, SHRD64rrCL)>;
def WriteSHLDm : SchedWriteRes<[JLAGU, JALU01]> {
let Latency = 9;
let ResourceCycles = [1, 22];
let NumMicroOps = 8;
}
def: InstRW<[WriteSHLDm], (instregex "SHLD(16|32|64)mr(i8|CL)")>;
def: InstRW<[WriteSHLDm], (instregex "SHRD(16|32|64)mr(i8|CL)")>;
def: InstRW<[WriteSHLDm],(instrs SHLD16mri8, SHLD32mri8, SHLD64mri8,
SHLD16mrCL, SHLD32mrCL, SHLD64mrCL,
SHRD16mri8, SHRD32mri8, SHRD64mri8,
SHRD16mrCL, SHRD32mrCL, SHRD64mrCL)>;
////////////////////////////////////////////////////////////////////////////////
// Loads, stores, and moves, not folded with other operations.
@ -378,13 +380,13 @@ def WriteFHAddY: SchedWriteRes<[JFPU0]> {
let Latency = 3;
let ResourceCycles = [2];
}
def : InstRW<[WriteFHAddY], (instregex "VH(ADD|SUB)P(S|D)Yrr")>;
def : InstRW<[WriteFHAddY], (instrs VHADDPDYrr, VHADDPSYrr, VHSUBPDYrr, VHSUBPSYrr)>;
def WriteFHAddYLd: SchedWriteRes<[JLAGU, JFPU0]> {
let Latency = 8;
let ResourceCycles = [1, 2];
}
def : InstRW<[WriteFHAddYLd], (instregex "VH(ADD|SUB)P(S|D)Yrm")>;
def : InstRW<[WriteFHAddYLd], (instrs VHADDPDYrm, VHADDPSYrm, VHSUBPDYrm, VHSUBPSYrm)>;
////////////////////////////////////////////////////////////////////////////////
// Carry-less multiplication instructions.
@ -524,13 +526,17 @@ def WriteFAddY: SchedWriteRes<[JFPU0]> {
let Latency = 3;
let ResourceCycles = [2];
}
def : InstRW<[WriteFAddY], (instregex "VADD(SUB)?P(S|D)Yrr", "VSUBP(S|D)Yrr")>;
def : InstRW<[WriteFAddY], (instrs VADDPDYrr, VADDPSYrr,
VSUBPDYrr, VSUBPSYrr,
VADDSUBPDYrr, VADDSUBPSYrr)>;
def WriteFAddYLd: SchedWriteRes<[JLAGU, JFPU0]> {
let Latency = 8;
let ResourceCycles = [1, 2];
}
def : InstRW<[WriteFAddYLd, ReadAfterLd], (instregex "VADD(SUB)?P(S|D)Yrm", "VSUBP(S|D)Yrm")>;
def : InstRW<[WriteFAddYLd, ReadAfterLd], (instrs VADDPDYrm, VADDPSYrm,
VSUBPDYrm, VSUBPSYrm,
VADDSUBPDYrm, VADDSUBPSYrm)>;
def WriteFDivY: SchedWriteRes<[JFPU1]> {
let Latency = 38;
@ -584,17 +590,17 @@ def WriteVCVTY: SchedWriteRes<[JSTC]> {
let Latency = 3;
let ResourceCycles = [2];
}
def : InstRW<[WriteVCVTY], (instregex "VCVTDQ2P(S|D)Yrr")>;
def : InstRW<[WriteVCVTY], (instregex "VROUNDYP(S|D)r")>;
def : InstRW<[WriteVCVTY], (instrs VCVTPS2DQYrr, VCVTTPS2DQYrr)>;
def : InstRW<[WriteVCVTY], (instrs VCVTDQ2PDYrr, VCVTDQ2PSYrr,
VCVTPS2DQYrr, VCVTTPS2DQYrr,
VROUNDYPDr, VROUNDYPSr)>;
def WriteVCVTYLd: SchedWriteRes<[JLAGU, JSTC]> {
let Latency = 8;
let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instregex "VCVTDQ2P(S|D)Yrm")>;
def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instregex "VROUNDYP(S|D)m")>;
def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instrs VCVTPS2DQYrm, VCVTTPS2DQYrm)>;
def : InstRW<[WriteVCVTYLd, ReadAfterLd], (instrs VCVTDQ2PDYrm, VCVTDQ2PSYrm,
VCVTPS2DQYrm, VCVTTPS2DQYrm,
VROUNDYPDm, VROUNDYPSm)>;
def WriteVMOVNTDQSt: SchedWriteRes<[JSTC, JSAGU]> {
let Latency = 2;
@ -616,17 +622,15 @@ def WriteFCmp: SchedWriteRes<[JFPU0]> {
let Latency = 2;
}
def : InstRW<[WriteFCmp], (instregex "VMAXP(D|S)rr", "VMAXS(D|S)rr")>;
def : InstRW<[WriteFCmp], (instregex "VMINP(D|S)rr", "VMINS(D|S)rr")>;
def : InstRW<[WriteFCmp], (instregex "VCMPP(S|D)rri", "VCMPS(S|D)rr")>;
def : InstRW<[WriteFCmp], (instregex "(V)?M(AX|IN)(P|S)(D|S)rr",
"(V)?CMPP(S|D)rri", "(V)?CMPS(S|D)rr")>;
def WriteFCmpLd: SchedWriteRes<[JLAGU, JFPU0]> {
let Latency = 7;
}
def : InstRW<[WriteFCmpLd], (instregex "VMAXP(D|S)rm", "VMAXS(D|S)rm")>;
def : InstRW<[WriteFCmpLd], (instregex "VMINP(D|S)rm", "VMINS(D|S)rm")>;
def : InstRW<[WriteFCmpLd], (instregex "VCMPP(S|D)rmi", "VCMPS(S|D)rm")>;
def : InstRW<[WriteFCmpLd], (instregex "(V)?M(AX|IN)(P|S)(D|S)rm",
"(V)?CMPP(S|D)rmi", "(V)?CMPS(S|D)rm")>;
def WriteVCVTPDY: SchedWriteRes<[JSTC, JFPU01]> {
let Latency = 6;
@ -674,13 +678,13 @@ def WriteVHAddSubY: SchedWriteRes<[JFPU0]> {
let Latency = 3;
let ResourceCycles = [2];
}
def : InstRW<[WriteVHAddSubY], (instregex "VH(ADD|SUB)P(D|S)Yrr")>;
def : InstRW<[WriteVHAddSubY], (instrs VHADDPDYrr, VHADDPSYrr, VHSUBPDYrr, VHSUBPSYrr)>;
def WriteVHAddSubYLd: SchedWriteRes<[JLAGU, JFPU0]> {
let Latency = 8;
let ResourceCycles = [1, 2];
}
def : InstRW<[WriteVHAddSubYLd], (instregex "VH(ADD|SUB)P(D|S)Yrm")>;
def : InstRW<[WriteVHAddSubYLd], (instrs VHADDPDYrm, VHADDPSYrm, VHSUBPDYrm, VHSUBPSYrm)>;
def WriteVMaskMovLd: SchedWriteRes<[JLAGU,JFPU01]> {
let Latency = 6;
@ -713,7 +717,7 @@ def : InstRW<[WriteVMaskMovYSt], (instrs VMASKMOVPDYmr, VMASKMOVPSYmr)>;
def WriteVMOVMSK: SchedWriteRes<[JFPU0]> {
let Latency = 3;
}
def : InstRW<[WriteVMOVMSK], (instregex "(V)?MOVMSKP(D|S)(Y)?rr")>;
def : InstRW<[WriteVMOVMSK], (instrs VMOVMSKPDrr, VMOVMSKPDYrr, VMOVMSKPSrr, VMOVMSKPSYrr)>;
// TODO: In fact we have latency '3+i'. The +i represents an additional 1 cycle transfer
// operation which moves the floating point result to the integer unit. During this