This is an archive of the discontinued LLVM Phabricator instance.

Differential D83671

GlobalISel: Implement widenScalar for saturating add/sub
ClosedPublic

Authored by arsenm on Jul 13 2020, 5:21 AM.

Details

Reviewers
foad
aemerson
aditya_nandakumar
paquette
kerbowa
Summary

Add a placeholder legality rule for AMDGPU until the rest of the
actions are handled.

Diff Detail

Event Timeline

arsenm created this revision.Jul 13 2020, 5:21 AM
Herald added a project: Restricted Project. · View Herald TranscriptJul 13 2020, 5:21 AM
Herald added subscribers: hiraditya, tpr, rovka and 3 others. · View Herald Transcript
foad accepted this revision.Jul 13 2020, 6:32 AM

Looks fine but can be simplified.

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
1644–1645

This isn't needed if you make the changes below.

1660–1663

These can both be anyext.

1671

There's no need to make this conditional on IsSigned if we're just about to truncate anyway.

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
1431

"selectively"

This revision is now accepted and ready to land.Jul 13 2020, 6:32 AM
arsenm marked an inline comment as done.Jul 13 2020, 6:36 AM
arsenm added inline comments.
llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
1671

How else would I know whether to prefer ashr or lshr?

foad added inline comments.Jul 13 2020, 6:37 AM
llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
1671

Why would it make any difference? Can't you just pick one arbitrarily?

arsenm marked an inline comment as done.Jul 13 2020, 6:46 AM
arsenm added inline comments.
llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
1671

First I don't know whether the target has a reason to prefer one over the other. I also would assume using the right shift to preserve the expected number of sign bits for the signed result would be more useful for downstream simplifications once the trunc is folded out

foad added inline comments.Jul 13 2020, 6:50 AM
llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp
1671

OK, I won't object if you want to leave it like that. The other comment about anyext still stands.

arsenm closed this revision.Jul 13 2020, 11:46 AM

6a8c11a11f6cf8d4ef1fa373c9546c33c69f8cb0

Revision Contents

PathSize
llvm/
include/
llvm/
CodeGen/
GlobalISel/
2 lines
lib/
CodeGen/
GlobalISel/
43 lines
Target/
AMDGPU/
6 lines
test/
CodeGen/
AMDGPU/
GlobalISel/
72 lines
72 lines
72 lines
72 lines

Diff 277392

llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h

Show First 20 Lines • Show All 157 Lines • ▼ Show 20 Linesprivate:
LegalizeResult LegalizeResult
widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx, LLT WideTy); widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
LegalizeResult LegalizeResult
widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx, LLT WideTy); widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
LegalizeResult LegalizeResult
widenScalarExtract(MachineInstr &MI, unsigned TypeIdx, LLT WideTy); widenScalarExtract(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
LegalizeResult LegalizeResult
widenScalarInsert(MachineInstr &MI, unsigned TypeIdx, LLT WideTy); widenScalarInsert(MachineInstr &MI, unsigned TypeIdx, LLT WideTy);
LegalizeResult widenScalarAddSubSat(MachineInstr &MI, unsigned TypeIdx,
LLT WideTy);
/// Helper function to split a wide generic register into bitwise blocks with /// Helper function to split a wide generic register into bitwise blocks with
/// the given Type (which implies the number of blocks needed). The generic /// the given Type (which implies the number of blocks needed). The generic
/// registers created are appended to Ops, starting at bit 0 of Reg. /// registers created are appended to Ops, starting at bit 0 of Reg.
void extractParts(Register Reg, LLT Ty, int NumParts, void extractParts(Register Reg, LLT Ty, int NumParts,
SmallVectorImpl<Register> &VRegs); SmallVectorImpl<Register> &VRegs);
/// Version which handles irregular splits. /// Version which handles irregular splits.
▲ Show 20 Lines • Show All 169 Lines • Show Last 20 Lines

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

Show First 20 Lines • Show All 1,633 Lines • ▼ Show 20 LinesLegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx,
Observer.changingInstr(MI); Observer.changingInstr(MI);
widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT); widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
widenScalarDst(MI, WideTy); widenScalarDst(MI, WideTy);
Observer.changedInstr(MI); Observer.changedInstr(MI);
return Legalized; return Legalized;
} }
LegalizerHelper::LegalizeResult LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarAddSubSat(MachineInstr &MI, unsigned TypeIdx,
LLT WideTy) {
bool IsSigned = MI.getOpcode() == TargetOpcode::G_SADDSAT ||
MI.getOpcode() == TargetOpcode::G_SSUBSAT;
foadUnsubmitted
Not Done
ReplyInline Actions

This isn't needed if you make the changes below.

foad: This isn't needed if you make the changes below.
// We can convert this to:
// 1. Extend iN to iM
// 2. SHL by M-N
// 3. [US][ADD|SUB]SAT
// 4. L/ASHR by M-N
//
// It may be more efficient to lower this to a min and a max operation in
// the higher precision arithmetic if the promoted operation isn't legal,
// but this decision is up to the target's lowering request.
Register DstReg = MI.getOperand(0).getReg();
unsigned NewBits = WideTy.getScalarSizeInBits();
unsigned SHLAmount = NewBits - MRI.getType(DstReg).getScalarSizeInBits();
auto LHS = IsSigned ? MIRBuilder.buildSExt(WideTy, MI.getOperand(1))
: MIRBuilder.buildZExt(WideTy, MI.getOperand(1));
auto RHS = IsSigned ? MIRBuilder.buildSExt(WideTy, MI.getOperand(2))
: MIRBuilder.buildZExt(WideTy, MI.getOperand(2));
foadUnsubmitted
Not Done
ReplyInline Actions

These can both be anyext.

foad: These can both be anyext.
auto ShiftK = MIRBuilder.buildConstant(WideTy, SHLAmount);
auto ShiftL = MIRBuilder.buildShl(WideTy, LHS, ShiftK);
auto ShiftR = MIRBuilder.buildShl(WideTy, RHS, ShiftK);
auto WideInst = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy},
{ShiftL, ShiftR}, MI.getFlags());
auto Result = IsSigned ? MIRBuilder.buildAShr(WideTy, WideInst, ShiftK)
foadUnsubmitted
Not Done
ReplyInline Actions

There's no need to make this conditional on IsSigned if we're just about to truncate anyway.

foad: There's no need to make this conditional on IsSigned if we're just about to truncate anyway.
arsenmAuthorUnsubmitted
Done
ReplyInline Actions

How else would I know whether to prefer ashr or lshr?

arsenm: How else would I know whether to prefer ashr or lshr?
foadUnsubmitted
Not Done
ReplyInline Actions

Why would it make any difference? Can't you just pick one arbitrarily?

foad: Why would it make any difference? Can't you just pick one arbitrarily?
arsenmAuthorUnsubmitted
Done
ReplyInline Actions

First I don't know whether the target has a reason to prefer one over the other. I also would assume using the right shift to preserve the expected number of sign bits for the signed result would be more useful for downstream simplifications once the trunc is folded out

arsenm: First I don't know whether the target has a reason to prefer one over the other. I also would…
foadUnsubmitted
Not Done
ReplyInline Actions

OK, I won't object if you want to leave it like that. The other comment about anyext still stands.

foad: OK, I won't object if you want to leave it like that. The other comment about anyext still…
: MIRBuilder.buildLShr(WideTy, WideInst, ShiftK);
MIRBuilder.buildTrunc(DstReg, Result);
MI.eraseFromParent();
return Legalized;
}
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) { LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
switch (MI.getOpcode()) { switch (MI.getOpcode()) {
default: default:
return UnableToLegalize; return UnableToLegalize;
case TargetOpcode::G_EXTRACT: case TargetOpcode::G_EXTRACT:
return widenScalarExtract(MI, TypeIdx, WideTy); return widenScalarExtract(MI, TypeIdx, WideTy);
case TargetOpcode::G_INSERT: case TargetOpcode::G_INSERT:
return widenScalarInsert(MI, TypeIdx, WideTy); return widenScalarInsert(MI, TypeIdx, WideTy);
Show All 19 Lines auto AndOp = MIRBuilder.buildAnd(
WideTy, NewOp, MIRBuilder.buildConstant(WideTy, Mask)); WideTy, NewOp, MIRBuilder.buildConstant(WideTy, Mask));
// There is no overflow if the AndOp is the same as NewOp. // There is no overflow if the AndOp is the same as NewOp.
MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1), NewOp, AndOp); MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1), NewOp, AndOp);
// Now trunc the NewOp to the original result. // Now trunc the NewOp to the original result.
MIRBuilder.buildTrunc(MI.getOperand(0), NewOp); MIRBuilder.buildTrunc(MI.getOperand(0), NewOp);
MI.eraseFromParent(); MI.eraseFromParent();
return Legalized; return Legalized;
} }
case TargetOpcode::G_SADDSAT:
case TargetOpcode::G_SSUBSAT:
case TargetOpcode::G_UADDSAT:
case TargetOpcode::G_USUBSAT:
return widenScalarAddSubSat(MI, TypeIdx, WideTy);
case TargetOpcode::G_CTTZ: case TargetOpcode::G_CTTZ:
case TargetOpcode::G_CTTZ_ZERO_UNDEF: case TargetOpcode::G_CTTZ_ZERO_UNDEF:
case TargetOpcode::G_CTLZ: case TargetOpcode::G_CTLZ:
case TargetOpcode::G_CTLZ_ZERO_UNDEF: case TargetOpcode::G_CTLZ_ZERO_UNDEF:
case TargetOpcode::G_CTPOP: { case TargetOpcode::G_CTPOP: {
if (TypeIdx == 0) { if (TypeIdx == 0) {
Observer.changingInstr(MI); Observer.changingInstr(MI);
widenScalarDst(MI, WideTy, 0); widenScalarDst(MI, WideTy, 0);
▲ Show 20 Lines • Show All 3,639 Lines • Show Last 20 Lines

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

Show First 20 Lines • Show All 1,421 Lines • ▼ Show 20 Lines
} else if (ST.has16BitInsts()) { } else if (ST.has16BitInsts()) {
SextInReg.lowerFor({{S32}, {S64}, {S16}}); SextInReg.lowerFor({{S32}, {S64}, {S16}});
} else { } else {
// Prefer to promote to s32 before lowering if we don't have 16-bit // Prefer to promote to s32 before lowering if we don't have 16-bit
// shifts. This avoid a lot of intermediate truncate and extend operations. // shifts. This avoid a lot of intermediate truncate and extend operations.
SextInReg.lowerFor({{S32}, {S64}}); SextInReg.lowerFor({{S32}, {S64}});
} }
// FIXME: Placeholder rule. Really depends on whether the clamp modifier is
// available, and is selectivly legal for s16, s32, v2s16.
foadUnsubmitted
Not Done
ReplyInline Actions

"selectively"

foad: "selectively"
getActionDefinitionsBuilder({G_SADDSAT, G_SSUBSAT, G_UADDSAT, G_USUBSAT})
.scalarize(0)
.clampScalar(0, S16, S32);
SextInReg SextInReg
.scalarize(0) .scalarize(0)
.clampScalar(0, S32, S64) .clampScalar(0, S32, S64)
.lower(); .lower();
getActionDefinitionsBuilder(G_FSHR) getActionDefinitionsBuilder(G_FSHR)
.legalFor({{S32, S32}}) .legalFor({{S32, S32}})
.scalarize(0) .scalarize(0)
▲ Show 20 Lines • Show All 2,903 Lines • Show Last 20 Lines

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-saddsat.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9
--- ---
name: saddsat_s7 name: saddsat_s7
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: saddsat_s7 ; GFX6-LABEL: name: saddsat_s7
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX6: [[SADDSAT:%[0-9]+]]:_(s7) = G_SADDSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s7)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SADDSAT]](s7) ; GFX6: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s7)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX6: [[SADDSAT:%[0-9]+]]:_(s16) = G_SADDSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[SEXT2:%[0-9]+]]:_(s32) = G_SEXT [[SADDSAT]](s16)
; GFX6: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[SEXT2]], [[ZEXT]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[ASHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s7) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s7)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: saddsat_s7 ; GFX8-LABEL: name: saddsat_s7
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX8: [[SADDSAT:%[0-9]+]]:_(s7) = G_SADDSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s7)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SADDSAT]](s7) ; GFX8: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s7)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX8: [[SADDSAT:%[0-9]+]]:_(s16) = G_SADDSAT [[SHL]], [[SHL1]]
; GFX8: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SADDSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[ASHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: saddsat_s7 ; GFX9-LABEL: name: saddsat_s7
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX9: [[SADDSAT:%[0-9]+]]:_(s7) = G_SADDSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s7)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SADDSAT]](s7) ; GFX9: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s7)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX9: [[SADDSAT:%[0-9]+]]:_(s16) = G_SADDSAT [[SHL]], [[SHL1]]
; GFX9: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SADDSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[ASHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s7) = G_TRUNC %0 %2:_(s7) = G_TRUNC %0
%3:_(s7) = G_TRUNC %1 %3:_(s7) = G_TRUNC %1
%4:_(s7) = G_SADDSAT %2, %3 %4:_(s7) = G_SADDSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
... ...
--- ---
name: saddsat_s8 name: saddsat_s8
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: saddsat_s8 ; GFX6-LABEL: name: saddsat_s8
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX6: [[SADDSAT:%[0-9]+]]:_(s8) = G_SADDSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s8)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SADDSAT]](s8) ; GFX6: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s8)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX6: [[SADDSAT:%[0-9]+]]:_(s16) = G_SADDSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[SEXT2:%[0-9]+]]:_(s32) = G_SEXT [[SADDSAT]](s16)
; GFX6: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[SEXT2]], [[ZEXT]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[ASHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s8)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: saddsat_s8 ; GFX8-LABEL: name: saddsat_s8
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX8: [[SADDSAT:%[0-9]+]]:_(s8) = G_SADDSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s8)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SADDSAT]](s8) ; GFX8: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s8)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX8: [[SADDSAT:%[0-9]+]]:_(s16) = G_SADDSAT [[SHL]], [[SHL1]]
; GFX8: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SADDSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[ASHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: saddsat_s8 ; GFX9-LABEL: name: saddsat_s8
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX9: [[SADDSAT:%[0-9]+]]:_(s8) = G_SADDSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s8)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SADDSAT]](s8) ; GFX9: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s8)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX9: [[SADDSAT:%[0-9]+]]:_(s16) = G_SADDSAT [[SHL]], [[SHL1]]
; GFX9: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SADDSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[ASHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s8) = G_TRUNC %0 %2:_(s8) = G_TRUNC %0
%3:_(s8) = G_TRUNC %1 %3:_(s8) = G_TRUNC %1
%4:_(s8) = G_SADDSAT %2, %3 %4:_(s8) = G_SADDSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
▲ Show 20 Lines • Show All 294 Lines • Show Last 20 Lines

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-ssubsat.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9
--- ---
name: ssubsat_s7 name: ssubsat_s7
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: ssubsat_s7 ; GFX6-LABEL: name: ssubsat_s7
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX6: [[SSUBSAT:%[0-9]+]]:_(s7) = G_SSUBSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s7)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SSUBSAT]](s7) ; GFX6: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s7)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX6: [[SSUBSAT:%[0-9]+]]:_(s16) = G_SSUBSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[SEXT2:%[0-9]+]]:_(s32) = G_SEXT [[SSUBSAT]](s16)
; GFX6: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[SEXT2]], [[ZEXT]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[ASHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s7) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s7)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: ssubsat_s7 ; GFX8-LABEL: name: ssubsat_s7
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX8: [[SSUBSAT:%[0-9]+]]:_(s7) = G_SSUBSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s7)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SSUBSAT]](s7) ; GFX8: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s7)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX8: [[SSUBSAT:%[0-9]+]]:_(s16) = G_SSUBSAT [[SHL]], [[SHL1]]
; GFX8: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SSUBSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[ASHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: ssubsat_s7 ; GFX9-LABEL: name: ssubsat_s7
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX9: [[SSUBSAT:%[0-9]+]]:_(s7) = G_SSUBSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s7)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SSUBSAT]](s7) ; GFX9: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s7)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX9: [[SSUBSAT:%[0-9]+]]:_(s16) = G_SSUBSAT [[SHL]], [[SHL1]]
; GFX9: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SSUBSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[ASHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s7) = G_TRUNC %0 %2:_(s7) = G_TRUNC %0
%3:_(s7) = G_TRUNC %1 %3:_(s7) = G_TRUNC %1
%4:_(s7) = G_SSUBSAT %2, %3 %4:_(s7) = G_SSUBSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
... ...
--- ---
name: ssubsat_s8 name: ssubsat_s8
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: ssubsat_s8 ; GFX6-LABEL: name: ssubsat_s8
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX6: [[SSUBSAT:%[0-9]+]]:_(s8) = G_SSUBSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s8)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SSUBSAT]](s8) ; GFX6: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s8)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX6: [[SSUBSAT:%[0-9]+]]:_(s16) = G_SSUBSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[SEXT2:%[0-9]+]]:_(s32) = G_SEXT [[SSUBSAT]](s16)
; GFX6: [[ASHR:%[0-9]+]]:_(s32) = G_ASHR [[SEXT2]], [[ZEXT]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[ASHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s8)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: ssubsat_s8 ; GFX8-LABEL: name: ssubsat_s8
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX8: [[SSUBSAT:%[0-9]+]]:_(s8) = G_SSUBSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s8)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SSUBSAT]](s8) ; GFX8: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s8)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX8: [[SSUBSAT:%[0-9]+]]:_(s16) = G_SSUBSAT [[SHL]], [[SHL1]]
; GFX8: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SSUBSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[ASHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: ssubsat_s8 ; GFX9-LABEL: name: ssubsat_s8
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX9: [[SSUBSAT:%[0-9]+]]:_(s8) = G_SSUBSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[SEXT:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC]](s8)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[SSUBSAT]](s8) ; GFX9: [[SEXT1:%[0-9]+]]:_(s16) = G_SEXT [[TRUNC1]](s8)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[SEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[SEXT1]], [[C]](s16)
; GFX9: [[SSUBSAT:%[0-9]+]]:_(s16) = G_SSUBSAT [[SHL]], [[SHL1]]
; GFX9: [[ASHR:%[0-9]+]]:_(s16) = G_ASHR [[SSUBSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[ASHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s8) = G_TRUNC %0 %2:_(s8) = G_TRUNC %0
%3:_(s8) = G_TRUNC %1 %3:_(s8) = G_TRUNC %1
%4:_(s8) = G_SSUBSAT %2, %3 %4:_(s8) = G_SSUBSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
▲ Show 20 Lines • Show All 294 Lines • Show Last 20 Lines

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-uaddsat.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9
--- ---
name: uaddsat_s7 name: uaddsat_s7
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: uaddsat_s7 ; GFX6-LABEL: name: uaddsat_s7
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX6: [[UADDSAT:%[0-9]+]]:_(s7) = G_UADDSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s7)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UADDSAT]](s7) ; GFX6: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s7)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX6: [[UADDSAT:%[0-9]+]]:_(s16) = G_UADDSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[UADDSAT]](s16)
; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[ZEXT3]], [[ZEXT2]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s7) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s7)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: uaddsat_s7 ; GFX8-LABEL: name: uaddsat_s7
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX8: [[UADDSAT:%[0-9]+]]:_(s7) = G_UADDSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s7)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UADDSAT]](s7) ; GFX8: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s7)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX8: [[UADDSAT:%[0-9]+]]:_(s16) = G_UADDSAT [[SHL]], [[SHL1]]
; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[UADDSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[LSHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: uaddsat_s7 ; GFX9-LABEL: name: uaddsat_s7
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX9: [[UADDSAT:%[0-9]+]]:_(s7) = G_UADDSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s7)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UADDSAT]](s7) ; GFX9: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s7)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX9: [[UADDSAT:%[0-9]+]]:_(s16) = G_UADDSAT [[SHL]], [[SHL1]]
; GFX9: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[UADDSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[LSHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s7) = G_TRUNC %0 %2:_(s7) = G_TRUNC %0
%3:_(s7) = G_TRUNC %1 %3:_(s7) = G_TRUNC %1
%4:_(s7) = G_UADDSAT %2, %3 %4:_(s7) = G_UADDSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
... ...
--- ---
name: uaddsat_s8 name: uaddsat_s8
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: uaddsat_s8 ; GFX6-LABEL: name: uaddsat_s8
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX6: [[UADDSAT:%[0-9]+]]:_(s8) = G_UADDSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s8)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UADDSAT]](s8) ; GFX6: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s8)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX6: [[UADDSAT:%[0-9]+]]:_(s16) = G_UADDSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[UADDSAT]](s16)
; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[ZEXT3]], [[ZEXT2]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s8)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: uaddsat_s8 ; GFX8-LABEL: name: uaddsat_s8
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX8: [[UADDSAT:%[0-9]+]]:_(s8) = G_UADDSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s8)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UADDSAT]](s8) ; GFX8: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s8)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX8: [[UADDSAT:%[0-9]+]]:_(s16) = G_UADDSAT [[SHL]], [[SHL1]]
; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[UADDSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: uaddsat_s8 ; GFX9-LABEL: name: uaddsat_s8
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX9: [[UADDSAT:%[0-9]+]]:_(s8) = G_UADDSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s8)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[UADDSAT]](s8) ; GFX9: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s8)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX9: [[UADDSAT:%[0-9]+]]:_(s16) = G_UADDSAT [[SHL]], [[SHL1]]
; GFX9: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[UADDSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s8) = G_TRUNC %0 %2:_(s8) = G_TRUNC %0
%3:_(s8) = G_TRUNC %1 %3:_(s8) = G_TRUNC %1
%4:_(s8) = G_UADDSAT %2, %3 %4:_(s8) = G_UADDSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
▲ Show 20 Lines • Show All 294 Lines • Show Last 20 Lines

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-usubsat.mir

# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=tahiti -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX6
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=fiji -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX8
# RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9 # RUN: llc -global-isel-abort=0 -march=amdgcn -mcpu=gfx900 -run-pass=legalizer %s -o - | FileCheck -check-prefix=GFX9
--- ---
name: usubsat_s7 name: usubsat_s7
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: usubsat_s7 ; GFX6-LABEL: name: usubsat_s7
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX6: [[USUBSAT:%[0-9]+]]:_(s7) = G_USUBSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s7)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[USUBSAT]](s7) ; GFX6: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s7)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX6: [[USUBSAT:%[0-9]+]]:_(s16) = G_USUBSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[USUBSAT]](s16)
; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[ZEXT3]], [[ZEXT2]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s7) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s7)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: usubsat_s7 ; GFX8-LABEL: name: usubsat_s7
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX8: [[USUBSAT:%[0-9]+]]:_(s7) = G_USUBSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s7)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[USUBSAT]](s7) ; GFX8: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s7)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX8: [[USUBSAT:%[0-9]+]]:_(s16) = G_USUBSAT [[SHL]], [[SHL1]]
; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[USUBSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[LSHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: usubsat_s7 ; GFX9-LABEL: name: usubsat_s7
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s7) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s7) = G_TRUNC [[COPY1]](s32)
; GFX9: [[USUBSAT:%[0-9]+]]:_(s7) = G_USUBSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s7)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[USUBSAT]](s7) ; GFX9: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s7)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 9
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX9: [[USUBSAT:%[0-9]+]]:_(s16) = G_USUBSAT [[SHL]], [[SHL1]]
; GFX9: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[USUBSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s7) = G_TRUNC [[LSHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s7)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s7) = G_TRUNC %0 %2:_(s7) = G_TRUNC %0
%3:_(s7) = G_TRUNC %1 %3:_(s7) = G_TRUNC %1
%4:_(s7) = G_USUBSAT %2, %3 %4:_(s7) = G_USUBSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
... ...
--- ---
name: usubsat_s8 name: usubsat_s8
body: | body: |
bb.0: bb.0:
liveins: $vgpr0, $vgpr1 liveins: $vgpr0, $vgpr1
; GFX6-LABEL: name: usubsat_s8 ; GFX6-LABEL: name: usubsat_s8
; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX6: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX6: [[USUBSAT:%[0-9]+]]:_(s8) = G_USUBSAT [[TRUNC]], [[TRUNC1]] ; GFX6: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s8)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[USUBSAT]](s8) ; GFX6: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s8)
; GFX6: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX6: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX6: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX6: [[USUBSAT:%[0-9]+]]:_(s16) = G_USUBSAT [[SHL]], [[SHL1]]
; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[C]](s16)
; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[USUBSAT]](s16)
; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[ZEXT3]], [[ZEXT2]](s32)
; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
; GFX6: [[TRUNC3:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC2]](s16)
; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC3]](s8)
; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX8-LABEL: name: usubsat_s8 ; GFX8-LABEL: name: usubsat_s8
; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX8: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX8: [[USUBSAT:%[0-9]+]]:_(s8) = G_USUBSAT [[TRUNC]], [[TRUNC1]] ; GFX8: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s8)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[USUBSAT]](s8) ; GFX8: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s8)
; GFX8: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX8: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX8: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX8: [[USUBSAT:%[0-9]+]]:_(s16) = G_USUBSAT [[SHL]], [[SHL1]]
; GFX8: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[USUBSAT]], [[C]](s16)
; GFX8: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32)
; GFX9-LABEL: name: usubsat_s8 ; GFX9-LABEL: name: usubsat_s8
; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0
; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1 ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY $vgpr1
; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32) ; GFX9: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY1]](s32)
; GFX9: [[USUBSAT:%[0-9]+]]:_(s8) = G_USUBSAT [[TRUNC]], [[TRUNC1]] ; GFX9: [[ZEXT:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC]](s8)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[USUBSAT]](s8) ; GFX9: [[ZEXT1:%[0-9]+]]:_(s16) = G_ZEXT [[TRUNC1]](s8)
; GFX9: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
; GFX9: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[ZEXT]], [[C]](s16)
; GFX9: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[ZEXT1]], [[C]](s16)
; GFX9: [[USUBSAT:%[0-9]+]]:_(s16) = G_USUBSAT [[SHL]], [[SHL1]]
; GFX9: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[USUBSAT]], [[C]](s16)
; GFX9: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[TRUNC2]](s8)
; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32)
%0:_(s32) = COPY $vgpr0 %0:_(s32) = COPY $vgpr0
%1:_(s32) = COPY $vgpr1 %1:_(s32) = COPY $vgpr1
%2:_(s8) = G_TRUNC %0 %2:_(s8) = G_TRUNC %0
%3:_(s8) = G_TRUNC %1 %3:_(s8) = G_TRUNC %1
%4:_(s8) = G_USUBSAT %2, %3 %4:_(s8) = G_USUBSAT %2, %3
%5:_(s32) = G_ANYEXT %4 %5:_(s32) = G_ANYEXT %4
$vgpr0 = COPY %5 $vgpr0 = COPY %5
▲ Show 20 Lines • Show All 294 Lines • Show Last 20 Lines