Skip to content

[NVPTX][AutoUpgrade] Use integer min/max intrinsics instead of icmp, select #173097

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
AlexMaclean wants to merge 1 commit into
base: main
Choose a base branch
from
+25 −46
Open

[NVPTX][AutoUpgrade] Use integer min/max intrinsics instead of icmp, select #173097

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
35 changes: 13 additions & 22 deletions llvm/lib/IR/AutoUpgrade.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1538,6 +1538,19 @@ static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn,
{F->getReturnType()});
return true;
}
} else if (F->arg_size() == 2) {
Intrinsic::ID IID =
StringSwitch<Intrinsic::ID>(Name)
.Cases({"max.s", "max.i", "max.ll"}, Intrinsic::smax)
.Cases({"min.s", "min.i", "min.ll"}, Intrinsic::smin)
.Cases({"max.us", "max.ui", "max.ull"}, Intrinsic::umax)
.Cases({"min.us", "min.ui", "min.ull"}, Intrinsic::umin)
.Default(Intrinsic::not_intrinsic);
if (IID != Intrinsic::not_intrinsic) {
NewFn = Intrinsic::getOrInsertDeclaration(F->getParent(), IID,
{F->getReturnType()});
return true;
}
}

// Check for nvvm intrinsics that need a return type adjustment.
Expand Down Expand Up @@ -1581,10 +1594,6 @@ static bool upgradeIntrinsicFunction1(Function *F, Function *&NewFn,
// nvvm.ex2.approx.{f,ftz.f,d,f16x2}
Expand =
Name == "f" || Name == "ftz.f" || Name == "d" || Name == "f16x2";
else if (Name.consume_front("max.") || Name.consume_front("min."))
// nvvm.{min,max}.{i,ii,ui,ull}
Expand = Name == "s" || Name == "i" || Name == "ll" || Name == "us" ||
Name == "ui" || Name == "ull";
else if (Name.consume_front("atomic.load."))
// nvvm.atomic.load.add.{f32,f64}.p
// nvvm.atomic.load.{inc,dec}.32.p
Expand Down Expand Up @@ -2649,24 +2658,6 @@ static Value *upgradeNVVMIntrinsicCall(StringRef Name, CallBase *CI,
: AtomicRMWInst::UDecWrap;
Rep = Builder.CreateAtomicRMW(Op, Ptr, Val, MaybeAlign(),
AtomicOrdering::SequentiallyConsistent);
} else if (Name.consume_front("max.") &&
(Name == "s" || Name == "i" || Name == "ll" || Name == "us" ||
Name == "ui" || Name == "ull")) {
Value *Arg0 = CI->getArgOperand(0);
Value *Arg1 = CI->getArgOperand(1);
Value *Cmp = Name.starts_with("u")
? Builder.CreateICmpUGE(Arg0, Arg1, "max.cond")
: Builder.CreateICmpSGE(Arg0, Arg1, "max.cond");
Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "max");
} else if (Name.consume_front("min.") &&
(Name == "s" || Name == "i" || Name == "ll" || Name == "us" ||
Name == "ui" || Name == "ull")) {
Value *Arg0 = CI->getArgOperand(0);
Value *Arg1 = CI->getArgOperand(1);
Value *Cmp = Name.starts_with("u")
? Builder.CreateICmpULE(Arg0, Arg1, "min.cond")
: Builder.CreateICmpSLE(Arg0, Arg1, "min.cond");
Rep = Builder.CreateSelect(Cmp, Arg0, Arg1, "min");
} else if (Name == "clz.ll") {
// llvm.nvvm.clz.ll returns an i32, but llvm.ctlz.i64 returns an i64.
Value *Arg = CI->getArgOperand(0);
Expand Down
36 changes: 12 additions & 24 deletions llvm/test/Assembler/auto_upgrade_nvvm_intrinsics.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,52 +158,40 @@ define void @tanh(float %a) {

; CHECK-LABEL: @min_max
define void @min_max(i16 %a1, i16 %a2, i32 %b1, i32 %b2, i64 %c1, i64 %c2) {
; CHECK: [[maxs:%[a-zA-Z0-9.]+]] = icmp sge i16 %a1, %a2
; CHECK: select i1 [[maxs]], i16 %a1, i16 %a2
; CHECK: %r1 = call i16 @llvm.smax.i16(i16 %a1, i16 %a2)
%r1 = call i16 @llvm.nvvm.max.s(i16 %a1, i16 %a2)

; CHECK: [[maxi:%[a-zA-Z0-9.]+]] = icmp sge i32 %b1, %b2
; CHECK: select i1 [[maxi]], i32 %b1, i32 %b2
; CHECK: %r2 = call i32 @llvm.smax.i32(i32 %b1, i32 %b2)
%r2 = call i32 @llvm.nvvm.max.i(i32 %b1, i32 %b2)

; CHECK: [[maxll:%[a-zA-Z0-9.]+]] = icmp sge i64 %c1, %c2
; CHECK: select i1 [[maxll]], i64 %c1, i64 %c2
; CHECK: %r3 = call i64 @llvm.smax.i64(i64 %c1, i64 %c2)
%r3 = call i64 @llvm.nvvm.max.ll(i64 %c1, i64 %c2)

; CHECK: [[maxus:%[a-zA-Z0-9.]+]] = icmp uge i16 %a1, %a2
; CHECK: select i1 [[maxus]], i16 %a1, i16 %a2
; CHECK: %r4 = call i16 @llvm.umax.i16(i16 %a1, i16 %a2)
%r4 = call i16 @llvm.nvvm.max.us(i16 %a1, i16 %a2)

; CHECK: [[maxui:%[a-zA-Z0-9.]+]] = icmp uge i32 %b1, %b2
; CHECK: select i1 [[maxui]], i32 %b1, i32 %b2
; CHECK: %r5 = call i32 @llvm.umax.i32(i32 %b1, i32 %b2)
%r5 = call i32 @llvm.nvvm.max.ui(i32 %b1, i32 %b2)

; CHECK: [[maxull:%[a-zA-Z0-9.]+]] = icmp uge i64 %c1, %c2
; CHECK: select i1 [[maxull]], i64 %c1, i64 %c2
; CHECK: %r6 = call i64 @llvm.umax.i64(i64 %c1, i64 %c2)
%r6 = call i64 @llvm.nvvm.max.ull(i64 %c1, i64 %c2)

; CHECK: [[mins:%[a-zA-Z0-9.]+]] = icmp sle i16 %a1, %a2
; CHECK: select i1 [[mins]], i16 %a1, i16 %a2
; CHECK: %r7 = call i16 @llvm.smin.i16(i16 %a1, i16 %a2)
%r7 = call i16 @llvm.nvvm.min.s(i16 %a1, i16 %a2)

; CHECK: [[mini:%[a-zA-Z0-9.]+]] = icmp sle i32 %b1, %b2
; CHECK: select i1 [[mini]], i32 %b1, i32 %b2
; CHECK: %r8 = call i32 @llvm.smin.i32(i32 %b1, i32 %b2)
%r8 = call i32 @llvm.nvvm.min.i(i32 %b1, i32 %b2)

; CHECK: [[minll:%[a-zA-Z0-9.]+]] = icmp sle i64 %c1, %c2
; CHECK: select i1 [[minll]], i64 %c1, i64 %c2
; CHECK: %r9 = call i64 @llvm.smin.i64(i64 %c1, i64 %c2)
%r9 = call i64 @llvm.nvvm.min.ll(i64 %c1, i64 %c2)

; CHECK: [[minus:%[a-zA-Z0-9.]+]] = icmp ule i16 %a1, %a2
; CHECK: select i1 [[minus]], i16 %a1, i16 %a2
; CHECK: %r10 = call i16 @llvm.umin.i16(i16 %a1, i16 %a2)
%r10 = call i16 @llvm.nvvm.min.us(i16 %a1, i16 %a2)

; CHECK: [[minui:%[a-zA-Z0-9.]+]] = icmp ule i32 %b1, %b2
; CHECK: select i1 [[minui]], i32 %b1, i32 %b2
; CHECK: %r11 = call i32 @llvm.umin.i32(i32 %b1, i32 %b2)
%r11 = call i32 @llvm.nvvm.min.ui(i32 %b1, i32 %b2)

; CHECK: [[minull:%[a-zA-Z0-9.]+]] = icmp ule i64 %c1, %c2
; CHECK: select i1 [[minull]], i64 %c1, i64 %c2
; CHECK: %r12 = call i64 @llvm.umin.i64(i64 %c1, i64 %c2)
%r12 = call i64 @llvm.nvvm.min.ull(i64 %c1, i64 %c2)

ret void
Expand Down
Loading