Skip to content

[AMDGPU] Implement IR variant of isFMAFasterThanFMulAndFAdd #121465

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 1 commit into from
Jan 10, 2025
Merged

[AMDGPU] Implement IR variant of isFMAFasterThanFMulAndFAdd #121465

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
56 changes: 56 additions & 0 deletions llvm/lib/Target/AMDGPU/SIISelLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5731,6 +5731,35 @@ bool SITargetLowering::isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
return false;
}

// Refer to comments added to the MIR variant of isFMAFasterThanFMulAndFAdd for
// specific details.
bool SITargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F,
Type *Ty) const {
switch (Ty->getScalarSizeInBits()) {
case 16: {
SIModeRegisterDefaults Mode = SIModeRegisterDefaults(F, *Subtarget);
Copy link
Contributor

@arsenm arsenm Jan 9, 2025

Choose a reason for hiding this comment

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

Should probably try to defer parsing the attribute as late as possible (i.e. check the features first, and only check this if the mode matters for the subtarget)

Copy link
Contributor Author

@chinmaydd chinmaydd Jan 10, 2025

Choose a reason for hiding this comment

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

Will handle this in a follow-up PR (handling modifiers) soon

return Subtarget->has16BitInsts() &&
Mode.FP64FP16Denormals != DenormalMode::getPreserveSign();
}
case 32: {
if (!Subtarget->hasMadMacF32Insts())
return Subtarget->hasFastFMAF32();

SIModeRegisterDefaults Mode = SIModeRegisterDefaults(F, *Subtarget);
if (Mode.FP32Denormals != DenormalMode::getPreserveSign())
return Subtarget->hasFastFMAF32() || Subtarget->hasDLInsts();

return Subtarget->hasFastFMAF32() && Subtarget->hasDLInsts();
}
case 64:
return true;
default:
break;
}

return false;
}

bool SITargetLowering::isFMADLegal(const MachineInstr &MI, LLT Ty) const {
if (!Ty.isScalar())
return false;
Expand Down Expand Up @@ -16992,6 +17021,33 @@ bool SITargetLowering::checkForPhysRegDependency(
return false;
}

/// Check if it is profitable to hoist instruction in then/else to if.
bool SITargetLowering::isProfitableToHoist(Instruction *I) const {
if (!I->hasOneUse())
return true;

Instruction *User = I->user_back();
// TODO: Add more patterns that are not profitable to hoist and
// handle modifiers such as fabs and fneg
switch (I->getOpcode()) {
case Instruction::FMul: {
if (User->getOpcode() != Instruction::FSub &&
User->getOpcode() != Instruction::FAdd)
return true;

const TargetOptions &Options = getTargetMachine().Options;

return ((!I->hasAllowContract() || !User->hasAllowContract()) &&
Options.AllowFPOpFusion != FPOpFusion::Fast &&
!Options.UnsafeFPMath) ||
!isFMAFasterThanFMulAndFAdd(*I->getFunction(), User->getType());
}
default:
return true;
}
return true;
}

void SITargetLowering::emitExpandAtomicAddrSpacePredicate(
Instruction *AI) const {
// Given: atomicrmw fadd ptr %addr, float %val ordering
Expand Down
3 changes: 3 additions & 0 deletions llvm/lib/Target/AMDGPU/SIISelLowering.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,6 +459,7 @@ class SITargetLowering final : public AMDGPUTargetLowering {
EVT VT) const override;
bool isFMAFasterThanFMulAndFAdd(const MachineFunction &MF,
const LLT Ty) const override;
bool isFMAFasterThanFMulAndFAdd(const Function &F, Type *Ty) const override;
bool isFMADLegal(const SelectionDAG &DAG, const SDNode *N) const override;
bool isFMADLegal(const MachineInstr &MI, const LLT Ty) const override;

Expand Down Expand Up @@ -538,6 +539,8 @@ class SITargetLowering final : public AMDGPUTargetLowering {
const TargetInstrInfo *TII, unsigned &PhysReg,
int &Cost) const override;

bool isProfitableToHoist(Instruction *I) const override;

bool isKnownNeverNaNForTargetNode(SDValue Op,
const SelectionDAG &DAG,
bool SNaN = false,
Expand Down
108 changes: 73 additions & 35 deletions llvm/test/CodeGen/AMDGPU/prevent-fmul-hoist-ir.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,16 +11,17 @@ define double @is_profitable_f64_contract(ptr dereferenceable(8) %ptr_x, ptr der
; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
; GFX-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
; GFX-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
; GFX-NEXT: [[MUL:%.*]] = fmul contract double [[X]], [[A_1]]
; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; GFX: [[COMMON_RET:.*]]:
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: ret double [[COMMON_RET_OP]]
; GFX: [[IF_THEN]]:
; GFX-NEXT: [[MUL:%.*]] = fmul contract double [[X]], [[A_1]]
; GFX-NEXT: [[ADD]] = fadd contract double 1.000000e+00, [[MUL]]
; GFX-NEXT: br label %[[COMMON_RET]]
; GFX: [[IF_ELSE]]:
; GFX-NEXT: [[SUB]] = fsub contract double [[MUL]], [[Y]]
; GFX-NEXT: [[MUL1:%.*]] = fmul contract double [[X]], [[A_1]]
; GFX-NEXT: [[SUB]] = fsub contract double [[MUL1]], [[Y]]
; GFX-NEXT: br label %[[COMMON_RET]]
;
entry:
Expand Down Expand Up @@ -93,16 +94,17 @@ define float @is_profitable_f32(ptr dereferenceable(8) %ptr_x, ptr dereferenceab
; GFX-NEXT: [[CMP:%.*]] = fcmp oeq float [[Y]], 0.000000e+00
; GFX-NEXT: [[X:%.*]] = load float, ptr [[PTR_X]], align 8
; GFX-NEXT: [[A_1:%.*]] = load float, ptr [[PTR_A]], align 8
; GFX-NEXT: [[MUL:%.*]] = fmul contract float [[X]], [[A_1]]
; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; GFX: [[COMMON_RET:.*]]:
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi float [ [[MUL]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi float [ [[MUL:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: ret float [[COMMON_RET_OP]]
; GFX: [[IF_THEN]]:
; GFX-NEXT: [[MUL]] = fmul contract float [[X]], [[A_1]]
; GFX-NEXT: [[ADD:%.*]] = fadd contract float 1.000000e+00, [[MUL]]
; GFX-NEXT: br label %[[COMMON_RET]]
; GFX: [[IF_ELSE]]:
; GFX-NEXT: [[SUB]] = fsub contract float [[MUL]], [[Y]]
; GFX-NEXT: [[MUL1:%.*]] = fmul contract float [[X]], [[A_1]]
; GFX-NEXT: [[SUB]] = fsub contract float [[MUL1]], [[Y]]
; GFX-NEXT: br label %[[COMMON_RET]]
;
entry:
Expand All @@ -111,7 +113,6 @@ entry:
%x = load float, ptr %ptr_x, align 8
br i1 %cmp, label %if.then, label %if.else


if.then: ; preds = %entry
%a_1 = load float, ptr %ptr_a, align 8
%mul = fmul contract float %x, %a_1
Expand Down Expand Up @@ -172,16 +173,17 @@ define half @is_profitable_f16_ieee(ptr dereferenceable(8) %ptr_x, ptr dereferen
; GFX-NEXT: [[CMP:%.*]] = fcmp oeq half [[Y]], 0xH0000
; GFX-NEXT: [[X:%.*]] = load half, ptr [[PTR_X]], align 8
; GFX-NEXT: [[A_1:%.*]] = load half, ptr [[PTR_A]], align 8
; GFX-NEXT: [[MUL:%.*]] = fmul contract half [[X]], [[A_1]]
; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; GFX: [[COMMON_RET:.*]]:
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi half [ [[MUL]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi half [ [[MUL:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: ret half [[COMMON_RET_OP]]
; GFX: [[IF_THEN]]:
; GFX-NEXT: [[MUL]] = fmul contract half [[X]], [[A_1]]
; GFX-NEXT: [[ADD:%.*]] = fadd contract half [[Y]], [[MUL]]
; GFX-NEXT: br label %[[COMMON_RET]]
; GFX: [[IF_ELSE]]:
; GFX-NEXT: [[SUB]] = fsub contract half [[MUL]], [[Y]]
; GFX-NEXT: [[MUL1:%.*]] = fmul contract half [[X]], [[A_1]]
; GFX-NEXT: [[SUB]] = fsub contract half [[MUL1]], [[Y]]
; GFX-NEXT: br label %[[COMMON_RET]]
;
entry:
Expand Down Expand Up @@ -250,16 +252,17 @@ define bfloat @is_profitable_bfloat_ieee(ptr dereferenceable(8) %ptr_x, ptr dere
; GFX-NEXT: [[CMP:%.*]] = fcmp oeq bfloat [[Y]], 0xR0000
; GFX-NEXT: [[X:%.*]] = load bfloat, ptr [[PTR_X]], align 8
; GFX-NEXT: [[A_1:%.*]] = load bfloat, ptr [[PTR_A]], align 8
; GFX-NEXT: [[MUL:%.*]] = fmul contract bfloat [[X]], [[A_1]]
; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; GFX: [[COMMON_RET:.*]]:
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi bfloat [ [[MUL]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi bfloat [ [[MUL:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: ret bfloat [[COMMON_RET_OP]]
; GFX: [[IF_THEN]]:
; GFX-NEXT: [[MUL]] = fmul contract bfloat [[X]], [[A_1]]
; GFX-NEXT: [[ADD:%.*]] = fadd contract bfloat 0xR3F80, [[MUL]]
; GFX-NEXT: br label %[[COMMON_RET]]
; GFX: [[IF_ELSE]]:
; GFX-NEXT: [[SUB]] = fsub contract bfloat [[MUL]], [[Y]]
; GFX-NEXT: [[MUL1:%.*]] = fmul contract bfloat [[X]], [[A_1]]
; GFX-NEXT: [[SUB]] = fsub contract bfloat [[MUL1]], [[Y]]
; GFX-NEXT: br label %[[COMMON_RET]]
;
entry:
Expand Down Expand Up @@ -330,16 +333,17 @@ define <8 x half> @is_profitable_vector(ptr dereferenceable(8) %ptr_x, ptr deref
; GFX-NEXT: [[V1:%.*]] = load <8 x half>, ptr addrspace(3) @v1_ptr, align 16
; GFX-NEXT: [[V2:%.*]] = load <8 x half>, ptr addrspace(3) @v2_ptr, align 16
; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
; GFX-NEXT: [[MUL:%.*]] = fmul contract <8 x half> [[V1]], [[X]]
; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; GFX: [[COMMON_RET:.*]]:
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi <8 x half> [ [[ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: ret <8 x half> [[COMMON_RET_OP]]
; GFX: [[IF_THEN]]:
; GFX-NEXT: [[MUL:%.*]] = fmul contract <8 x half> [[V1]], [[X]]
; GFX-NEXT: [[ADD]] = fadd contract <8 x half> [[V2]], [[MUL]]
; GFX-NEXT: br label %[[COMMON_RET]]
; GFX: [[IF_ELSE]]:
; GFX-NEXT: [[SUB]] = fsub contract <8 x half> [[MUL]], [[V2]]
; GFX-NEXT: [[MUL1:%.*]] = fmul contract <8 x half> [[V1]], [[X]]
; GFX-NEXT: [[SUB]] = fsub contract <8 x half> [[MUL1]], [[V2]]
; GFX-NEXT: br label %[[COMMON_RET]]
;
entry:
Expand All @@ -362,23 +366,61 @@ if.else: ; preds = %entry
}

define double @is_profitable_f64_nocontract(ptr dereferenceable(8) %ptr_x, ptr dereferenceable(8) %ptr_y, ptr dereferenceable(8) %ptr_a) #0 {
; GFX-LABEL: define double @is_profitable_f64_nocontract(
; GFX-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
; GFX-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
; GFX-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
; GFX-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
; GFX-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
; GFX-NEXT: [[MUL:%.*]] = fmul double [[X]], [[A_1]]
; GFX-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; GFX: [[COMMON_RET:.*]]:
; GFX-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[PTR_ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; GFX-NEXT: ret double [[COMMON_RET_OP]]
; GFX: [[IF_THEN]]:
; GFX-NEXT: [[PTR_ADD]] = fadd double 1.000000e+00, [[MUL]]
; GFX-NEXT: br label %[[COMMON_RET]]
; GFX: [[IF_ELSE]]:
; GFX-NEXT: [[SUB]] = fsub double [[MUL]], [[Y]]
; GFX-NEXT: br label %[[COMMON_RET]]
; FP-CONTRACT-FAST-LABEL: define double @is_profitable_f64_nocontract(
; FP-CONTRACT-FAST-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
; FP-CONTRACT-FAST-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
; FP-CONTRACT-FAST-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
; FP-CONTRACT-FAST-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
; FP-CONTRACT-FAST-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
; FP-CONTRACT-FAST-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; FP-CONTRACT-FAST: [[COMMON_RET:.*]]:
; FP-CONTRACT-FAST-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[PTR_ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; FP-CONTRACT-FAST-NEXT: ret double [[COMMON_RET_OP]]
; FP-CONTRACT-FAST: [[IF_THEN]]:
; FP-CONTRACT-FAST-NEXT: [[MUL:%.*]] = fmul double [[X]], [[A_1]]
; FP-CONTRACT-FAST-NEXT: [[PTR_ADD]] = fadd double 1.000000e+00, [[MUL]]
; FP-CONTRACT-FAST-NEXT: br label %[[COMMON_RET]]
; FP-CONTRACT-FAST: [[IF_ELSE]]:
; FP-CONTRACT-FAST-NEXT: [[MUL1:%.*]] = fmul double [[X]], [[A_1]]
; FP-CONTRACT-FAST-NEXT: [[SUB]] = fsub double [[MUL1]], [[Y]]
; FP-CONTRACT-FAST-NEXT: br label %[[COMMON_RET]]
;
; UNSAFE-FP-MATH-LABEL: define double @is_profitable_f64_nocontract(
; UNSAFE-FP-MATH-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
; UNSAFE-FP-MATH-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
; UNSAFE-FP-MATH-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
; UNSAFE-FP-MATH-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
; UNSAFE-FP-MATH-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
; UNSAFE-FP-MATH-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; UNSAFE-FP-MATH: [[COMMON_RET:.*]]:
; UNSAFE-FP-MATH-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[PTR_ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; UNSAFE-FP-MATH-NEXT: ret double [[COMMON_RET_OP]]
; UNSAFE-FP-MATH: [[IF_THEN]]:
; UNSAFE-FP-MATH-NEXT: [[MUL:%.*]] = fmul double [[X]], [[A_1]]
; UNSAFE-FP-MATH-NEXT: [[PTR_ADD]] = fadd double 1.000000e+00, [[MUL]]
; UNSAFE-FP-MATH-NEXT: br label %[[COMMON_RET]]
; UNSAFE-FP-MATH: [[IF_ELSE]]:
; UNSAFE-FP-MATH-NEXT: [[MUL1:%.*]] = fmul double [[X]], [[A_1]]
; UNSAFE-FP-MATH-NEXT: [[SUB]] = fsub double [[MUL1]], [[Y]]
; UNSAFE-FP-MATH-NEXT: br label %[[COMMON_RET]]
;
; NO-UNSAFE-FP-MATH-LABEL: define double @is_profitable_f64_nocontract(
; NO-UNSAFE-FP-MATH-SAME: ptr dereferenceable(8) [[PTR_X:%.*]], ptr dereferenceable(8) [[PTR_Y:%.*]], ptr dereferenceable(8) [[PTR_A:%.*]]) #[[ATTR0]] {
; NO-UNSAFE-FP-MATH-NEXT: [[Y:%.*]] = load double, ptr [[PTR_Y]], align 8
; NO-UNSAFE-FP-MATH-NEXT: [[CMP:%.*]] = fcmp oeq double [[Y]], 0.000000e+00
; NO-UNSAFE-FP-MATH-NEXT: [[X:%.*]] = load double, ptr [[PTR_X]], align 8
; NO-UNSAFE-FP-MATH-NEXT: [[A_1:%.*]] = load double, ptr [[PTR_A]], align 8
; NO-UNSAFE-FP-MATH-NEXT: [[MUL:%.*]] = fmul double [[X]], [[A_1]]
; NO-UNSAFE-FP-MATH-NEXT: br i1 [[CMP]], label %[[IF_THEN:.*]], label %[[IF_ELSE:.*]]
; NO-UNSAFE-FP-MATH: [[COMMON_RET:.*]]:
; NO-UNSAFE-FP-MATH-NEXT: [[COMMON_RET_OP:%.*]] = phi double [ [[PTR_ADD:%.*]], %[[IF_THEN]] ], [ [[SUB:%.*]], %[[IF_ELSE]] ]
; NO-UNSAFE-FP-MATH-NEXT: ret double [[COMMON_RET_OP]]
; NO-UNSAFE-FP-MATH: [[IF_THEN]]:
; NO-UNSAFE-FP-MATH-NEXT: [[PTR_ADD]] = fadd double 1.000000e+00, [[MUL]]
; NO-UNSAFE-FP-MATH-NEXT: br label %[[COMMON_RET]]
; NO-UNSAFE-FP-MATH: [[IF_ELSE]]:
; NO-UNSAFE-FP-MATH-NEXT: [[SUB]] = fsub double [[MUL]], [[Y]]
; NO-UNSAFE-FP-MATH-NEXT: br label %[[COMMON_RET]]
;
%y = load double, ptr %ptr_y, align 8
%cmp = fcmp oeq double %y, 0.000000e+00
Expand All @@ -400,7 +442,3 @@ if.else: ; preds = %entry

attributes #0 = { nounwind "denormal-fp-math"="preserve-sign,preserve-sign" }
attributes #1 = { nounwind "denormal-fp-math"="ieee,ieee" }
;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
; FP-CONTRACT-FAST: {{.*}}
; NO-UNSAFE-FP-MATH: {{.*}}
; UNSAFE-FP-MATH: {{.*}}
Loading