Skip to content

Commit 8413a7e

Browse files
committed
[AMDGPU] Allocate AVRegClass last
Change-Id: Iace3462f27ea276b22716793ebfa13b5026b0e58
1 parent d737fe2 commit 8413a7e

26 files changed

+1677
-1647
lines changed

llvm/lib/Target/AMDGPU/SIRegisterInfo.td

Lines changed: 35 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -109,6 +109,23 @@ class SIRegisterClass <string n, list<ValueType> rTypes, int Align, dag rList>
109109
let TSFlags{2} = HasVGPR;
110110
let TSFlags{3} = HasAGPR;
111111
let TSFlags{4} = HasSGPR;
112+
113+
// RA will use RegisterClass AllocationPriority amongst other info (e.g. ordering in the basic block)
114+
// to decide which registers to try to assign first. Usually, this RegisterClass priority is given
115+
// very high priority, if not the highest priority, when considering which VirtReg to allocate next.
116+
//
117+
// We have 5 bits to assign AllocationPriorities to RegisterClasses. Generally, it is beneficial to
118+
// assign more constrained RegisterClasses first. As a result, we prioritize register classes with
119+
// more 32 bit tuples (e.g. VReg_512) over registers with fewer tuples (e.g. VGPR_32).
120+
//
121+
// The interesting case is the vector register case on architectures which have ARegs, VRegs, AVRegs.
122+
// In this case, we would like to assign ARegs and VRegs before AVRegs, as AVRegs are less constrained
123+
// and can be assigned to both AGPRs and VGPRs. We use the 5th bit to encode this into the
124+
// RegisterClass AllocationPriority. BaseClassPriority is used to turn the bit on, and BaseClassScaleFactor
125+
// is used for scaling of the bit (i.e. 1 << 4).
126+
field int BaseClassPriority = 1;
127+
field int BaseClassScaleFactor = 16;
128+
112129
}
113130

114131
multiclass SIRegLoHi16 <string n, bits<8> regIdx, bit ArtificialHigh = 1,
@@ -575,7 +592,7 @@ let HasVGPR = 1 in {
575592
def VGPR_16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
576593
(add (interleave (sequence "VGPR%u_LO16", 0, 255),
577594
(sequence "VGPR%u_HI16", 0, 255)))> {
578-
let AllocationPriority = 2;
595+
let AllocationPriority = !add(2, !mul(BaseClassPriority, BaseClassScaleFactor));
579596
let Size = 16;
580597
let GeneratePressureSet = 0;
581598

@@ -601,7 +618,7 @@ def VGPR_16_Lo128 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
601618
// i16/f16 only on VI+
602619
def VGPR_32 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
603620
(add (sequence "VGPR%u", 0, 255))> {
604-
let AllocationPriority = 0;
621+
let AllocationPriority = !add(0, !mul(BaseClassPriority, BaseClassScaleFactor));
605622
let Size = 32;
606623
let Weight = 1;
607624
let BaseClassOrder = 32;
@@ -610,7 +627,7 @@ def VGPR_32 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types
610627
// Identical to VGPR_32 except it only contains the low 128 (Lo128) registers.
611628
def VGPR_32_Lo128 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
612629
(add (sequence "VGPR%u", 0, 127))> {
613-
let AllocationPriority = 0;
630+
let AllocationPriority = !add(0, !mul(BaseClassPriority, BaseClassScaleFactor));
614631
let GeneratePressureSet = 0;
615632
let Size = 32;
616633
let Weight = 1;
@@ -668,7 +685,7 @@ def AGPR_LO16 : SIRegisterClass<"AMDGPU", Reg16Types.types, 16,
668685
// AccVGPR 32-bit registers
669686
def AGPR_32 : SIRegisterClass<"AMDGPU", [i32, f32, i16, f16, bf16, v2i16, v2f16, v2bf16], 32,
670687
(add (sequence "AGPR%u", 0, 255))> {
671-
let AllocationPriority = 0;
688+
let AllocationPriority = !add(0, !mul(BaseClassPriority, BaseClassScaleFactor));
672689
let Size = 32;
673690
let Weight = 1;
674691
let BaseClassOrder = 32;
@@ -940,14 +957,23 @@ class VRegClassBase<int numRegs, list<ValueType> regTypes, dag regList> :
940957

941958
// Requires n v_mov_b32 to copy
942959
let CopyCost = numRegs;
943-
let AllocationPriority = !sub(numRegs, 1);
960+
961+
// Since we only have 5 bits for the RegisterClass Allocation Priorty, and since we use the
962+
// 5th bit for BaseClassPriority, we need to encode the SizePriority into 4 bits. As a result
963+
// of this encoding, for registers with numRegs 15 or 16, we give SizePriority of 14, and for
964+
// regsters with numRegs 17+ we give SizePriority of 15. In practice, there is only one
965+
// RegClass per Vector Register type in each of these groups (i.e. numRegs = 15,16 : {VReg_512},
966+
// and numRegs = 17+ : {VReg_1024}). Therefore, we have not lost any info by compressing.
967+
defvar SizePrioriity = !if(!le(numRegs, 14), !sub(numRegs, 1), !if(!le(numRegs, 16), 14, 15));
968+
969+
let AllocationPriority = !add(SizePrioriity, !mul(BaseClassPriority, BaseClassScaleFactor));
944970
let Weight = numRegs;
945971
}
946972

947973
// Define a register tuple class, along with one requiring an even
948974
// aligned base register.
949975
multiclass VRegClass<int numRegs, list<ValueType> regTypes, dag regList> {
950-
let HasVGPR = 1 in {
976+
let HasVGPR = 1, BaseClassPriority = 1 in {
951977
// Define the regular class.
952978
def "" : VRegClassBase<numRegs, regTypes, regList> {
953979
let BaseClassOrder = !mul(numRegs, 32);
@@ -981,7 +1007,7 @@ defm VReg_1024 : VRegClass<32, Reg1024Types.types, (add VGPR_1024)>;
9811007
}
9821008

9831009
multiclass ARegClass<int numRegs, list<ValueType> regTypes, dag regList> {
984-
let CopyCost = !add(numRegs, numRegs, 1), HasAGPR = 1 in {
1010+
let CopyCost = !add(numRegs, numRegs, 1), HasAGPR = 1, BaseClassPriority = 1 in {
9851011
// Define the regular class.
9861012
def "" : VRegClassBase<numRegs, regTypes, regList> {
9871013
let BaseClassOrder = !mul(numRegs, 32);
@@ -1066,6 +1092,7 @@ def VS_64 : SIRegisterClass<"AMDGPU", VReg_64.RegTypes, 32, (add VReg_64, SReg_6
10661092
def AV_32 : SIRegisterClass<"AMDGPU", VGPR_32.RegTypes, 32, (add VGPR_32, AGPR_32)> {
10671093
let HasVGPR = 1;
10681094
let HasAGPR = 1;
1095+
let BaseClassPriority = 0;
10691096
let Size = 32;
10701097
}
10711098
} // End GeneratePressureSet = 0
@@ -1074,7 +1101,7 @@ def AV_32 : SIRegisterClass<"AMDGPU", VGPR_32.RegTypes, 32, (add VGPR_32, AGPR_3
10741101
// aligned base register.
10751102
multiclass AVRegClass<int numRegs, list<ValueType> regTypes,
10761103
dag vregList, dag aregList> {
1077-
let HasVGPR = 1, HasAGPR = 1 in {
1104+
let HasVGPR = 1, HasAGPR = 1, BaseClassPriority = 0 in {
10781105
// Define the regular class.
10791106
def "" : VRegClassBase<numRegs, regTypes, (add vregList, aregList)>;
10801107

0 commit comments

Comments
 (0)