Skip to content

[AMDGPU] Prioritize allocation of low 256 VGPR classes #167978

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
+12 −1
Open

[AMDGPU] Prioritize allocation of low 256 VGPR classes #167978

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 11 additions & 0 deletions llvm/lib/Target/AMDGPU/SIRegisterInfo.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -501,6 +501,17 @@ class SIRegisterInfo final : public AMDGPUGenRegisterInfo {

SmallVector<StringLiteral>
getVRegFlagsOfReg(Register Reg, const MachineFunction &MF) const override;

float
getSpillWeightScaleFactor(const TargetRegisterClass *RC) const override {
// Prioritize VGPR_32_Lo256 over other classes which may occupy registers
// beyond v256.
return AMDGPUGenRegisterInfo::getSpillWeightScaleFactor(RC) *
((RC == &AMDGPU::VGPR_32_Lo256RegClass ||
RC == &AMDGPU::VReg_64_Lo256_Align2RegClass)
? 2.0
: 1.0);
}
Comment on lines +507 to +514
Copy link
Contributor

@arsenm arsenm Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This shouldn't require manual overriding

Copy link
Collaborator Author

@rampitec rampitec Nov 14, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At some level this is even right. Using this or that register is not more expensive. That is when it comes to spilling it matters.

};

namespace AMDGPU {
Expand Down
2 changes: 1 addition & 1 deletion llvm/lib/Target/AMDGPU/SIRegisterInfo.td
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -644,7 +644,7 @@ def VGPR_32_Lo128 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg1
// Identical to VGPR_32 except it only contains the low 256 (Lo256) registers.
def VGPR_32_Lo256 : SIRegisterClass<"AMDGPU", !listconcat(Reg32Types.types, Reg16Types.types), 32,
(add (sequence "VGPR%u", 0, 255))> {
let AllocationPriority = 0;
let AllocationPriority = !add(3, !mul(BaseClassPriority, BaseClassScaleFactor));
Copy link
Contributor

@arsenm arsenm Nov 13, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Did you try setting CostPerUse? I think that makes more sense than changing priority

Copy link
Collaborator Author

@rampitec rampitec Nov 14, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That's for registers, right? In this case it shall be per regclass. I.e., even if I increase it for low 256, it will be the same for a smaller and larger size. Here it is really a property of operand, not a register itself.

Copy link
Collaborator Author

@rampitec rampitec Nov 15, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

More or less that would make almost everything worse. Setting CostPerUse to non-zero for all low 256 VGPRs will persuade RA to allocate registers above 256 by default. At the end you will have programs using registers starting just from v256 and nothing at the range v0-v255.

let GeneratePressureSet = 0;
let Size = 32;
let Weight = 1;
Expand Down
Loading