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

[AArch64][Falkor] Refine modeling of multiply accumulate forwarding. 

llvm-svn: 302933
This commit is contained in:
Geoff Berry 2017-05-12 18:57:10 +00:00
parent d58e2d951e
commit ddbbf6416c
2 changed files with 61 additions and 44 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
llvm/lib/Target/AArch64/AArch64SchedFalkorDetails.td
View File

@ -42,11 +42,11 @@ def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^FCVT(N|M|P|Z|A)(S|U)(v1i32|v1
def : InstRW<[FalkorWr_1VXVY_4cyc], (instrs FCVTXNv1i64)>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^FCVTZ(S|U)(v2i32|v4i16)(_shift)?$")>;
def : InstRW<[FalkorWr_1VXVY_5cyc], (instregex "^(FMUL|FMULX)(v2f32|v4f16|(v1i16_indexed|v4i16_indexed|v1i32_indexed|v2i32_indexed))$")>;
def : InstRW<[FalkorWr_1VXVY_5cyc], (instrs FMULX16, FMULX32)>;
def : InstRW<[FalkorWr_FMUL32_1VXVY_5cyc], (instregex "^(FMUL|FMULX)(v2f32|v4f16|(v1i16_indexed|v4i16_indexed|v1i32_indexed|v2i32_indexed))$")>;
def : InstRW<[FalkorWr_FMUL32_1VXVY_5cyc], (instrs FMULX16, FMULX32)>;
def : InstRW<[FalkorWr_1VXVY_6cyc], (instregex "^(FMUL|FMULX)v1i64_indexed$")>;
def : InstRW<[FalkorWr_1VXVY_6cyc], (instrs FMULX64)>;
def : InstRW<[FalkorWr_FMUL64_1VXVY_6cyc], (instregex "^(FMUL|FMULX)v1i64_indexed$")>;
def : InstRW<[FalkorWr_FMUL64_1VXVY_6cyc], (instrs FMULX64)>;
def : InstRW<[FalkorWr_2VXVY_1cyc], (instregex "^(FABS|FNEG)(v2f64|v4f32|v8f16)$")>;
@ -62,9 +62,9 @@ def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^FCVT(N|M|P|Z|A)(S|U)(v2f64|v4
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^(FCVTL|FCVTL2)(v2i32|v4i16|v4i32|v8i16)$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^FCVTZ(S|U)(v2i64|v4i32|v8i16)(_shift)?$")>;
def : InstRW<[FalkorWr_2VXVY_5cyc], (instregex "^(FMUL|FMULX)(v2f64|v4f32|v8f16|v8i16_indexed|v4i32_indexed)$")>;
def : InstRW<[FalkorWr_FMUL32_2VXVY_5cyc], (instregex "^(FMUL|FMULX)(v2f64|v4f32|v8f16|v8i16_indexed|v4i32_indexed)$")>;
def : InstRW<[FalkorWr_2VXVY_6cyc], (instregex "^(FMUL|FMULX)v2i64_indexed$")>;
def : InstRW<[FalkorWr_FMUL64_2VXVY_6cyc], (instregex "^(FMUL|FMULX)v2i64_indexed$")>;
def : InstRW<[FalkorWr_3VXVY_4cyc], (instregex "^(FCVTX?N|FCVTX?N2)(v1i32|v1i64|v1f16|v2f32|v4f16)$")>;
@ -72,13 +72,14 @@ def : InstRW<[FalkorWr_3VXVY_5cyc], (instregex "^(FCVTX?N|FCVTX?N2)(v2i32|v4i1
def : InstRW<[FalkorWr_2VX_2VY_2cyc], (instregex "^(FDIV|FSQRT)(v2f64|v4f32|v8f16)$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc, FalkorReadVMA],(instregex "^ML(A|S)(v8i8|v4i16|v2i32)(_indexed)?$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc, FalkorReadVMA],(instregex "^ML(A|S)(v16i8|v8i16|v4i32|v2i64)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_1VXVY_4cyc, FalkorReadVMA], (instregex "^ML(A|S)(v8i8|v4i16|v2i32)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc, FalkorReadVMA], (instregex "^ML(A|S)(v16i8|v8i16|v4i32|v2i64)(_indexed)?$")>;
def : InstRW<[FalkorWr_FMUL32_1VXVY_5cyc, FalkorReadFMA32], (instregex "^FML(A|S)(v2f32|v4f16|(v1i16_indexed|v4i16_indexed|v1i32_indexed|v2i32_indexed))$")>;
def : InstRW<[FalkorWr_FMUL64_1VXVY_6cyc, FalkorReadFMA64], (instregex "^FML(A|S)v1i64_indexed$")>;
def : InstRW<[FalkorWr_FMUL32_2VXVY_5cyc, FalkorReadFMA32], (instregex "^FML(A|S)(v4f32|v8f16|v8i16_indexed|v4i32_indexed)$")>;
def : InstRW<[FalkorWr_FMUL64_2VXVY_6cyc, FalkorReadFMA64], (instregex "^FML(A|S)(v2f64|v2i64_indexed)$")>;
def : InstRW<[FalkorWr_1VXVY_5cyc, FalkorReadFMA],(instregex "^FML(A|S)(v2f32|v4f16|(v1i16_indexed|v4i16_indexed|v1i32_indexed|v2i32_indexed))$")>;
def : InstRW<[FalkorWr_1VXVY_6cyc, FalkorReadFMA],(instregex "^FML(A|S)v1i64_indexed$")>;
def : InstRW<[FalkorWr_2VXVY_5cyc, FalkorReadFMA],(instregex "^FML(A|S)(v2f64|v4f32|v8f16|v8i16_indexed|v4i32_indexed)$")>;
def : InstRW<[FalkorWr_2VXVY_6cyc, FalkorReadFMA],(instregex "^FML(A|S)v2i64_indexed$")>;
// SIMD Integer Instructions
// -----------------------------------------------------------------------------
def : InstRW<[FalkorWr_1VXVY_1cyc], (instregex "^ADD(v1i64|v2i32|v4i16|v8i8)$")>;
@ -119,10 +120,10 @@ def : InstRW<[FalkorWr_1VXVY_3cyc], (instregex "^SQNEG(v1i8|v1i16|v1i32|v1i64)
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^(S|U)ADDLVv8i8v$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^(S|U)?(MAX|MIN)V(v8i8v|v8i16v)$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instrs ADDVv8i8v)>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^MUL(v2i32|v4i16|v8i8)(_indexed)?$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^SQR?DMULH(v8i8|v4i16|v1i32|v2i32|v1i16)(_indexed)?$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^SQDMULL(i16|i32)$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instregex "^SQRDML(A|S)?H(v8i8|v4i16|v1i32|v2i32|v1i16)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_1VXVY_4cyc], (instregex "^MUL(v2i32|v4i16|v8i8)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_1VXVY_4cyc], (instregex "^SQR?DMULH(v8i8|v4i16|v1i32|v2i32|v1i16)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_1VXVY_4cyc], (instregex "^SQDMULL(i16|i32)$")>;
def : InstRW<[FalkorWr_VMUL32_1VXVY_4cyc, FalkorReadVMA], (instregex "^SQRDML(A|S)H(i16|i32|v8i8|v4i16|v1i32|v2i32|v1i16)(_indexed)?$")>;
def : InstRW<[FalkorWr_1VXVY_5cyc], (instregex "^(S|U)?(MAX|MIN)Vv16i8v$")>;
@ -169,9 +170,9 @@ def : InstRW<[FalkorWr_2VXVY_3cyc], (instregex "^PMULL2?(v1i64|v2i64)$")>;
def : InstRW<[FalkorWr_2VXVY_3cyc], (instregex "^S(L|R)I(v16i8|v8i16|v4i32|v2i64)_shift$")>;
def : InstRW<[FalkorWr_2VXVY_3cyc], (instregex "^SQ(ABS|NEG)(v16i8|v8i16|v4i32|v2i64)$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^(MUL|SQR?DMULH)(v16i8|v8i16|v4i32)(_indexed)?$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^SQDMULLv.*$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^SQRDML(A|S)H(v16i8|v8i16|v4i32)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc], (instregex "^(MUL|SQR?DMULH)(v16i8|v8i16|v4i32)(_indexed)?$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc], (instregex "^SQDMULLv.*$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc, FalkorReadVMA], (instregex "^SQRDML(A|S)H(v16i8|v8i16|v4i32)(_indexed)?$")>;
def : InstRW<[FalkorWr_3VXVY_3cyc], (instregex "^(S|U)ADDLVv4i32v$")>;
@ -185,8 +186,9 @@ def : InstRW<[FalkorWr_4VXVY_3cyc], (instregex "^(S|U)ABALv.*$")>;
def : InstRW<[FalkorWr_4VXVY_4cyc], (instregex "^(S|U)ABA(v16i8|v8i16|v4i32)$")>;
def : InstRW<[FalkorWr_1VXVY_4cyc, FalkorReadVMA],(instregex "^SQD(MLAL|MLSL)(i16|i32)$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc, FalkorReadVMA],(instregex "^SQD(MLAL|MLSL)v.*$")>;
def : InstRW<[FalkorWr_VMUL32_1VXVY_4cyc, FalkorReadVMA], (instregex "^SQD(MLAL|MLSL)(i16|i32|v1i32_indexed|v1i64_indexed)$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc, FalkorReadVMA], (instregex "^SQD(MLAL|MLSL)v[248].*$")>;
// SIMD Load Instructions
// -----------------------------------------------------------------------------
def : InstRW<[WriteVLD], (instregex "^LD1(i64|Onev(8b|4h|2s|1d|16b|8h|4s|2d))$")>;
@ -294,9 +296,9 @@ def : InstRW<[FalkorWr_1VXVY_3cyc], (instrs FRECPEv1i32, FRECPEv1i64, FRSQRTEv
def : InstRW<[FalkorWr_1VXVY_3cyc], (instrs FRECPXv1i32, FRECPXv1i64)>;
def : InstRW<[FalkorWr_1VXVY_3cyc], (instrs URECPEv2i32, URSQRTEv2i32)>;
def : InstRW<[FalkorWr_1VXVY_5cyc], (instrs FRECPS32, FRSQRTS32, FRECPSv2f32, FRSQRTSv2f32)>;
def : InstRW<[FalkorWr_FMUL32_1VXVY_5cyc], (instrs FRECPS32, FRSQRTS32, FRECPSv2f32, FRSQRTSv2f32)>;
def : InstRW<[FalkorWr_1VXVY_6cyc], (instrs FRECPS64, FRSQRTS64)>;
def : InstRW<[FalkorWr_FMUL64_1VXVY_6cyc], (instrs FRECPS64, FRSQRTS64)>;
def : InstRW<[FalkorWr_1GTOV_1VXVY_2cyc],(instregex "^INSv(i32|i64)(gpr|lane)$")>;
def : InstRW<[FalkorWr_2GTOV_1cyc], (instregex "^DUP(v4i32|v2i64)(gpr|lane)$")>;
@ -311,9 +313,9 @@ def : InstRW<[FalkorWr_2VXVY_3cyc], (instrs URECPEv4i32, URSQRTEv4i32)>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instrs TBLv8i8Two)>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^TBX(v8|v16)i8One$")>;
def : InstRW<[FalkorWr_2VXVY_5cyc], (instrs FRECPSv4f32, FRSQRTSv4f32)>;
def : InstRW<[FalkorWr_FMUL32_2VXVY_5cyc], (instrs FRECPSv4f32, FRSQRTSv4f32)>;
def : InstRW<[FalkorWr_2VXVY_6cyc], (instrs FRECPSv2f64, FRSQRTSv2f64)>;
def : InstRW<[FalkorWr_FMUL64_2VXVY_6cyc], (instrs FRECPSv2f64, FRSQRTSv2f64)>;
def : InstRW<[FalkorWr_3VXVY_5cyc], (instregex "^TBL(v8i8Three|v16i8Two)$")>;
def : InstRW<[FalkorWr_3VXVY_5cyc], (instregex "^TBX(v8i8Two|v16i8Two)$")>;
@ -416,15 +418,15 @@ def : InstRW<[FalkorWr_1VXVY_3cyc], (instrs FCVTSHr, FCVTDHr)>;
def : InstRW<[FalkorWr_1VXVY_4cyc], (instrs FCVTSDr, FCVTDSr)>;
def : InstRW<[FalkorWr_1VXVY_5cyc], (instregex "^F(N)?MUL(H|S)rr$")>;
def : InstRW<[FalkorWr_FMUL32_1VXVY_5cyc], (instregex "^F(N)?MUL(H|S)rr$")>;
def : InstRW<[FalkorWr_1VXVY_6cyc], (instregex "^F(N)?MULDrr$")>;
def : InstRW<[FalkorWr_FMUL64_1VXVY_6cyc], (instregex "^F(N)?MULDrr$")>;
def : InstRW<[FalkorWr_1VX_1VY_10cyc],(instregex "^FDIV(H|S|D)rr$")>;
def : InstRW<[FalkorWr_1VX_1VY_2cyc], (instregex "^FSQRT(H|S|D)r$")>;
def : InstRW<[FalkorWr_1VXVY_5cyc, FalkorReadFMA],(instregex "^F(N)?M(ADD|SUB)(H|S)rrr$")>;
def : InstRW<[FalkorWr_1VXVY_6cyc, FalkorReadFMA],(instregex "^F(N)?M(ADD|SUB)Drrr$")>;
def : InstRW<[FalkorWr_FMUL32_1VXVY_5cyc, ReadDefault, ReadDefault, FalkorReadFMA32], (instregex "^F(N)?M(ADD|SUB)(H|S)rrr$")>;
def : InstRW<[FalkorWr_FMUL64_1VXVY_6cyc, ReadDefault, ReadDefault, FalkorReadFMA64], (instregex "^F(N)?M(ADD|SUB)Drrr$")>;
// FP Miscellaneous Instructions
// -----------------------------------------------------------------------------
@ -475,16 +477,17 @@ def : InstRW<[FalkorWr_2XYZ_2cyc], (instregex "^EXTR(W|X)rri$")>;
// Divide and Multiply Instructions
// -----------------------------------------------------------------------------
def : InstRW<[FalkorWr_1X_4cyc], (instregex "^(S|U)M(ADD|SUB)Lrrr$")>;
def : InstRW<[FalkorWr_1X_4cyc], (instregex "^M(ADD|SUB)Wrrr$")>;
def : InstRW<[FalkorWr_IMUL64_1X_4cyc, ReadDefault, ReadDefault, FalkorReadIMA64], (instregex "^(S|U)M(ADD|SUB)Lrrr$")>;
def : InstRW<[FalkorWr_IMUL32_1X_2cyc, ReadDefault, ReadDefault, FalkorReadIMA32], (instregex "^M(ADD|SUB)Wrrr$")>;
def : InstRW<[FalkorWr_1X_5cyc], (instregex "^(S|U)MULHrr$")>;
def : InstRW<[FalkorWr_1X_5cyc], (instregex "^M(ADD|SUB)Xrrr$")>;
def : InstRW<[FalkorWr_IMUL64_1X_5cyc], (instregex "^(S|U)MULHrr$")>;
def : InstRW<[FalkorWr_IMUL64_1X_5cyc, ReadDefault, ReadDefault, FalkorReadIMA64], (instregex "^M(ADD|SUB)Xrrr$")>;
def : InstRW<[FalkorWr_1X_1Z_8cyc], (instregex "^(S|U)DIVWr$")>;
def : InstRW<[FalkorWr_1X_1Z_16cyc], (instregex "^(S|U)DIVXr$")>;
def : InstRW<[FalkorWr_2VXVY_4cyc], (instregex "^(S|U)(MLAL|MLSL|MULL)v.*$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc], (instregex "^(S|U)MULLv.*$")>;
def : InstRW<[FalkorWr_VMUL32_2VXVY_4cyc, FalkorReadVMA], (instregex "^(S|U)(MLAL|MLSL)v.*$")>;
// Move and Shift Instructions
// -----------------------------------------------------------------------------

34
llvm/lib/Target/AArch64/AArch64SchedFalkorWriteRes.td
View File

@ -29,8 +29,9 @@
// Define 1 micro-op types
def FalkorWr_1X_2cyc : SchedWriteRes<[FalkorUnitX]> { let Latency = 2; }
def FalkorWr_1X_4cyc : SchedWriteRes<[FalkorUnitX]> { let Latency = 4; }
def FalkorWr_1X_5cyc : SchedWriteRes<[FalkorUnitX]> { let Latency = 5; }
def FalkorWr_IMUL32_1X_2cyc : SchedWriteRes<[FalkorUnitX]> { let Latency = 4; }
def FalkorWr_IMUL64_1X_4cyc : SchedWriteRes<[FalkorUnitX]> { let Latency = 4; }
def FalkorWr_IMUL64_1X_5cyc : SchedWriteRes<[FalkorUnitX]> { let Latency = 5; }
def FalkorWr_1Z_0cyc : SchedWriteRes<[FalkorUnitZ]> { let Latency = 0; }
def FalkorWr_1ZB_0cyc : SchedWriteRes<[FalkorUnitZB]> { let Latency = 0; }
def FalkorWr_1LD_3cyc : SchedWriteRes<[FalkorUnitLD]> { let Latency = 3; }
@ -45,8 +46,10 @@ def FalkorWr_1VXVY_1cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 1; }
def FalkorWr_1VXVY_2cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 2; }
def FalkorWr_1VXVY_3cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 3; }
def FalkorWr_1VXVY_4cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 4; }
def FalkorWr_VMUL32_1VXVY_4cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 4; }
def FalkorWr_1VXVY_5cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 5; }
def FalkorWr_1VXVY_6cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 6; }
def FalkorWr_FMUL32_1VXVY_5cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 5; }
def FalkorWr_FMUL64_1VXVY_6cyc : SchedWriteRes<[FalkorUnitVXVY]>{ let Latency = 6; }
def FalkorWr_1LD_0cyc : SchedWriteRes<[FalkorUnitLD]> { let Latency = 0; }
def FalkorWr_1ST_0cyc : SchedWriteRes<[FalkorUnitST]> { let Latency = 0; }
@ -75,14 +78,26 @@ def FalkorWr_2VXVY_4cyc : SchedWriteRes<[FalkorUnitVXVY, FalkorUnitVXVY]> {
let Latency = 4;
let NumMicroOps = 2;
}
def FalkorWr_VMUL32_2VXVY_4cyc : SchedWriteRes<[FalkorUnitVXVY, FalkorUnitVXVY]> {
let Latency = 4;
let NumMicroOps = 2;
}
def FalkorWr_2VXVY_5cyc : SchedWriteRes<[FalkorUnitVXVY, FalkorUnitVXVY]> {
let Latency = 5;
let NumMicroOps = 2;
}
def FalkorWr_FMUL32_2VXVY_5cyc : SchedWriteRes<[FalkorUnitVXVY, FalkorUnitVXVY]> {
let Latency = 5;
let NumMicroOps = 2;
}
def FalkorWr_2VXVY_6cyc : SchedWriteRes<[FalkorUnitVXVY, FalkorUnitVXVY]> {
let Latency = 6;
let NumMicroOps = 2;
}
def FalkorWr_FMUL64_2VXVY_6cyc : SchedWriteRes<[FalkorUnitVXVY, FalkorUnitVXVY]> {
let Latency = 6;
let NumMicroOps = 2;
}
def FalkorWr_1LD_1VXVY_4cyc : SchedWriteRes<[FalkorUnitLD, FalkorUnitVXVY]> {
let Latency = 4;
@ -350,14 +365,13 @@ def FalkorWr_2LD_2VXVY_1XYZ_2LD_2VXVY_4cyc:SchedWriteRes<[FalkorUnitLD,
let NumMicroOps = 9;
}
// Forwarding logic is modeled for vector multiply and accumulate
// Forwarding logic is modeled for multiply add/accumulate.
// -----------------------------------------------------------------------------
def FalkorReadVMA : SchedReadAdvance<2, [FalkorWr_1VXVY_4cyc,
FalkorWr_2VXVY_4cyc]>;
def FalkorReadFMA : SchedReadAdvance<3, [FalkorWr_1VXVY_5cyc,
FalkorWr_1VXVY_6cyc,
FalkorWr_2VXVY_5cyc,
FalkorWr_2VXVY_6cyc]>;
def FalkorReadIMA32 : SchedReadAdvance<3, [FalkorWr_IMUL32_1X_2cyc]>;
def FalkorReadIMA64 : SchedReadAdvance<4, [FalkorWr_IMUL64_1X_4cyc, FalkorWr_IMUL64_1X_5cyc]>;
def FalkorReadVMA : SchedReadAdvance<3, [FalkorWr_VMUL32_1VXVY_4cyc, FalkorWr_VMUL32_2VXVY_4cyc]>;
def FalkorReadFMA32 : SchedReadAdvance<1, [FalkorWr_FMUL32_1VXVY_5cyc, FalkorWr_FMUL32_2VXVY_5cyc]>;
def FalkorReadFMA64 : SchedReadAdvance<2, [FalkorWr_FMUL64_1VXVY_6cyc, FalkorWr_FMUL64_2VXVY_6cyc]>;
// SchedPredicates and WriteVariants for Immediate Zero and LSLFast
// -----------------------------------------------------------------------------