Skip to content

AMDGPU: Update live intervals in convertToThreeAddress #104610

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

Merged
merged 8 commits into from
Sep 6, 2024
Merged

AMDGPU: Update live intervals in convertToThreeAddress #104610

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
c71c5dd
AMDGPU: Update live intervals in convertToThreeAddress
arsenm Aug 16, 2024
15573e4
Comments
arsenm Aug 19, 2024
a332148
Rename variable
arsenm Aug 19, 2024
270f780
Rename variable
arsenm Sep 4, 2024
1fc765d
Remove less
arsenm Sep 4, 2024
7d7a5e6
Add tracksRegLiveness to every function
arsenm Sep 4, 2024
61b61f2
Hackily use shrinkToUses
arsenm Sep 4, 2024
7a56477
Remove unused index argument
arsenm Sep 6, 2024
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
40 changes: 31 additions & 9 deletions llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4059,17 +4059,37 @@ MachineInstr *SIInstrInfo::convertToThreeAddress(MachineInstr &MI,
!RI.isSGPRReg(MBB.getParent()->getRegInfo(), Src0->getReg()))) {
MachineInstr *DefMI;
const auto killDef = [&]() -> void {
const MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
MachineRegisterInfo &MRI = MBB.getParent()->getRegInfo();
// The only user is the instruction which will be killed.
Register DefReg = DefMI->getOperand(0).getReg();
if (!MRI.hasOneNonDBGUse(DefReg))
return;
// We cannot just remove the DefMI here, calling pass will crash.
DefMI->setDesc(get(AMDGPU::IMPLICIT_DEF));
for (unsigned I = DefMI->getNumOperands() - 1; I != 0; --I)
DefMI->removeOperand(I);
if (LV)
LV->getVarInfo(DefReg).AliveBlocks.clear();

if (MRI.hasOneNonDBGUse(DefReg)) {
// We cannot just remove the DefMI here, calling pass will crash.
DefMI->setDesc(get(AMDGPU::IMPLICIT_DEF));
DefMI->getOperand(0).setIsDead(true);
for (unsigned I = DefMI->getNumOperands() - 1; I != 0; --I)
DefMI->removeOperand(I);
if (LV)
LV->getVarInfo(DefReg).AliveBlocks.clear();
}

if (LIS) {
LiveInterval &DefLI = LIS->getInterval(DefReg);

// We cannot delete the original instruction here, so hack out the use
Copy link
Contributor

Choose a reason for hiding this comment

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

Can you add a test that demonstrates this?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

This is already covered by several tests (the problematic cases were any of the ones with other uses)

// in the original instruction with a dummy register so we can use
// shrinkToUses to deal with any multi-use edge cases. Other targets do
// not have the complexity of deleting a use to consider here.
Register DummyReg = MRI.cloneVirtualRegister(DefReg);
for (MachineOperand &MIOp : MI.uses()) {
if (MIOp.isReg() && MIOp.getReg() == DefReg) {
MIOp.setIsUndef(true);
MIOp.setReg(DummyReg);
}
}

LIS->shrinkToUses(&DefLI);
}
};

int64_t Imm;
Expand Down Expand Up @@ -4107,6 +4127,7 @@ MachineInstr *SIInstrInfo::convertToThreeAddress(MachineInstr &MI,
.add(*Src2)
.setMIFlags(MI.getFlags());
updateLiveVariables(LV, MI, *MIB);

if (LIS)
LIS->ReplaceMachineInstrInMaps(MI, *MIB);
killDef();
Expand All @@ -4129,6 +4150,7 @@ MachineInstr *SIInstrInfo::convertToThreeAddress(MachineInstr &MI,
.add(*Src2)
.setMIFlags(MI.getFlags());
updateLiveVariables(LV, MI, *MIB);

if (LIS)
LIS->ReplaceMachineInstrInMaps(MI, *MIB);
if (DefMI)
Expand Down
142 changes: 125 additions & 17 deletions llvm/test/CodeGen/AMDGPU/gfx10-twoaddr-fma.mir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,33 +1,138 @@
# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 %s -run-pass twoaddressinstruction -verify-machineinstrs -o - | FileCheck --check-prefixes=GFX10 %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 %s --passes=two-address-instruction -verify-each -o - | FileCheck --check-prefixes=GFX10 %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=twoaddressinstruction -verify-machineinstrs -o - %s | FileCheck --check-prefixes=GFX10,GFX10-NOLIS %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -passes=two-address-instruction -verify-each -o - %s | FileCheck --check-prefixes=GFX10,GFX10-NOLIS %s
# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=liveintervals,twoaddressinstruction -verify-machineinstrs -o - %s | FileCheck --check-prefixes=GFX10,GFX10-LIS %s


# GFX10-LABEL: name: test_fmamk_reg_imm_f16
# GFX10: %2:vgpr_32 = IMPLICIT_DEF
# GFX10: dead %2:vgpr_32 = IMPLICIT_DEF
# GFX10-NOT: V_MOV_B32
# GFX10: V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-NOLIS: V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-LIS: V_FMAMK_F16 %0.sub0, 1078523331, %1, implicit $mode, implicit $exec
---
name: test_fmamk_reg_imm_f16
registers:
- { id: 0, class: vreg_64 }
- { id: 1, class: vgpr_32 }
- { id: 2, class: vgpr_32 }
- { id: 3, class: vgpr_32 }
tracksRegLiveness: true
body: |
bb.0:

%0 = IMPLICIT_DEF
%1 = COPY %0.sub1
%2 = V_MOV_B32_e32 1078523331, implicit $exec
%3 = V_FMAC_F16_e32 killed %0.sub0, %2, killed %1, implicit $mode, implicit $exec
%0:vreg_64 = IMPLICIT_DEF
%1:vgpr_32 = COPY %0.sub1
%2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
%3:vgpr_32 = V_FMAC_F16_e32 killed %0.sub0, %2, killed %1, implicit $mode, implicit $exec

...

# GFX10-LABEL: name: test_fmamk_reg_imm_f16__imm_is_subreg
# GFX10: %0:vreg_64 = IMPLICIT_DEF
# GFX10: %1:vgpr_32 = COPY %0.sub1
# GFX10: dead undef %2.sub0:vreg_64 = IMPLICIT_DEF
# GFX10-NOLIS: %3:vgpr_32 = V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-LIS: %3:vgpr_32 = V_FMAMK_F16 %0.sub0, 1078523331, %1, implicit $mode, implicit $exec
---
name: test_fmamk_reg_imm_f16__imm_is_subreg
tracksRegLiveness: true
body: |
bb.0:

%0:vreg_64 = IMPLICIT_DEF
%1:vgpr_32 = COPY %0.sub1
undef %2.sub0:vreg_64 = V_MOV_B32_e32 1078523331, implicit $exec
%3:vgpr_32 = V_FMAC_F16_e32 killed %0.sub0, %2.sub0, killed %1, implicit $mode, implicit $exec

...

# GFX10-LABEL: name: test_fmamk_reg_imm_f16__imm_is_subreg_fully_defined
# GFX10: %0:vreg_64 = IMPLICIT_DEF
# GFX10: %1:vgpr_32 = COPY %0.sub1
# GFX10: undef %2.sub1:vreg_64 = V_MOV_B32_e32 9999, implicit $exec
# GFX10: %2.sub0:vreg_64 = V_MOV_B32_e32 1078523331, implicit $exec
# GFX10-NOLIS: %3:vgpr_32 = V_FMA_F16_gfx9_e64 0, killed %0.sub0, 0, %2.sub0, 0, killed %1, 0, 0, 0, implicit $mode, implicit $e
# GFX10-LIS: %3:vgpr_32 = V_FMA_F16_gfx9_e64 0, %0.sub0, 0, %2.sub0, 0, %1, 0, 0, 0, implicit $mode, implicit $e
---
name: test_fmamk_reg_imm_f16__imm_is_subreg_fully_defined
tracksRegLiveness: true
body: |
bb.0:
%0:vreg_64 = IMPLICIT_DEF
%1:vgpr_32 = COPY %0.sub1
undef %2.sub1 = V_MOV_B32_e32 9999, implicit $exec
%2.sub0:vreg_64 = V_MOV_B32_e32 1078523331, implicit $exec
%3:vgpr_32 = V_FMAC_F16_e32 killed %0.sub0, %2.sub0, killed %1, implicit $mode, implicit $exec

...

# GFX10-LABEL: name: test_fmamk_reg_imm_f16__use_imm_before_mac
# GFX10: %0:vreg_64 = IMPLICIT_DEF
# GFX10: %1:vgpr_32 = COPY %0.sub1
# GFX10: %2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
# GFX10: S_NOP 0, implicit %2
# GFX10-NOLIS: %3:vgpr_32 = V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-LIS: %3:vgpr_32 = V_FMAMK_F16 %0.sub0, 1078523331, %1, implicit $mode, implicit $exec
---
name: test_fmamk_reg_imm_f16__use_imm_before_mac
tracksRegLiveness: true
body: |
bb.0:

%0:vreg_64 = IMPLICIT_DEF
%1:vgpr_32 = COPY %0.sub1
%2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
S_NOP 0, implicit %2
%3:vgpr_32 = V_FMAC_F16_e32 killed %0.sub0, %2, killed %1, implicit $mode, implicit $exec

...

# GFX10-LABEL: name: test_fmamk_reg_imm_f16__use_imm_after_mac
# GFX10: %0:vreg_64 = IMPLICIT_DEF
# GFX10: %1:vgpr_32 = COPY %0.sub1
# GFX10: %2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
# GFX10-NOLIS: %3:vgpr_32 = V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-LIS: %3:vgpr_32 = V_FMAMK_F16 %0.sub0, 1078523331, %1, implicit $mode, implicit $exec
---
name: test_fmamk_reg_imm_f16__use_imm_after_mac
tracksRegLiveness: true
body: |
bb.0:

%0:vreg_64 = IMPLICIT_DEF
%1:vgpr_32 = COPY %0.sub1
%2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
%3:vgpr_32 = V_FMAC_F16_e32 killed %0.sub0, %2, killed %1, implicit $mode, implicit $exec
S_NOP 0, implicit %2

...

# GFX10-LABEL: name: test_fmamk_reg_imm_f16__use_imm_before_after_mac
# GFX10: %0:vreg_64 = IMPLICIT_DEF
# GFX10: %1:vgpr_32 = COPY %0.sub1
# GFX10: %2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
# GFX10: S_NOP 0, implicit %2
# GFX10-NOLIS: %3:vgpr_32 = V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-LIS: %3:vgpr_32 = V_FMAMK_F16 %0.sub0, 1078523331, %1, implicit $mode, implicit $exec
# GFX10: S_NOP 0, implicit %2

---
name: test_fmamk_reg_imm_f16__use_imm_before_after_mac
tracksRegLiveness: true
body: |
bb.0:

%0:vreg_64 = IMPLICIT_DEF
%1:vgpr_32 = COPY %0.sub1
%2:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
S_NOP 0, implicit %2
%3:vgpr_32 = V_FMAC_F16_e32 killed %0.sub0, %2, killed %1, implicit $mode, implicit $exec
S_NOP 0, implicit %2

...

# GFX10-LABEL: name: test_fmamk_imm_reg_f16
# GFX10: %2:vgpr_32 = IMPLICIT_DEF
# GFX10: dead %2:vgpr_32 = IMPLICIT_DEF
# GFX10-NOT: V_MOV_B32
# GFX10: V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-NOLIS: V_FMAMK_F16 killed %0.sub0, 1078523331, killed %1, implicit $mode, implicit $exec
# GFX10-LIS: V_FMAMK_F16 %0.sub0, 1078523331, %1, implicit $mode, implicit $exec
---
name: test_fmamk_imm_reg_f16
tracksRegLiveness: true
registers:
- { id: 0, class: vreg_64 }
- { id: 1, class: vgpr_32 }
Expand All @@ -46,9 +151,11 @@ body: |
# GFX10-LABEL: name: test_fmaak_f16
# GFX10: %1:vgpr_32 = V_MOV_B32_e32 1078523331, implicit $exec
# GFX10-NOT: V_MOV_B32
# GFX10: V_FMAAK_F16 killed %0.sub0, %0.sub1, 1078523331, implicit $mode, implicit $exec
# GFX10-NOLIS: V_FMAAK_F16 killed %0.sub0, %0.sub1, 1078523331, implicit $mode, implicit $exec
# GFX10-LIS: V_FMAAK_F16 %0.sub0, %0.sub1, 1078523331, implicit $mode, implicit $exec
---
name: test_fmaak_f16
tracksRegLiveness: true
registers:
- { id: 0, class: vreg_64 }
- { id: 1, class: vgpr_32 }
Expand All @@ -65,7 +172,8 @@ body: |
# GFX10-LABEL: name: test_fmaak_inline_literal_f16
# GFX10: %1:vgpr_32 = V_MOV_B32_e32 49664, implicit $exec
# GFX10-NOT: V_MOV_B32
# GFX10: %2:vgpr_32 = V_FMAAK_F16 16384, killed %0, 49664, implicit $mode, implicit $exec
# GFX10-NOLIS: %2:vgpr_32 = V_FMAAK_F16 16384, killed %0, 49664, implicit $mode, implicit $exec
# GFX10-LIS: %2:vgpr_32 = V_FMAAK_F16 16384, %0, 49664, implicit $mode, implicit $exec

---
name: test_fmaak_inline_literal_f16
Expand Down
Loading