[llvm] 6261adf - [InstCombine] Fold more intrinsics over selects
Nikita Popov via llvm-commits
llvm-commits at lists.llvm.org
Fri Mar 31 07:32:31 PDT 2023
Author: Nikita Popov
Date: 2023-03-31T16:32:21+02:00
New Revision: 6261adfa51f2d93c8258e7b2d9811d607f3e50a5
URL: https://github.com/llvm/llvm-project/commit/6261adfa51f2d93c8258e7b2d9811d607f3e50a5
DIFF: https://github.com/llvm/llvm-project/commit/6261adfa51f2d93c8258e7b2d9811d607f3e50a5.diff
LOG: [InstCombine] Fold more intrinsics over selects
Move this handling to a centralized place and extend it to handle
saturating add/sub intrinsics.
I originally wanted to make this fully generic rather than
whitelist based, because this is legal and likely profitable for all
speculatable intrinsics. The caveat is that for vector selects,
the intrinsic can't perform cross-lane operations like a shuffle
or reduction, which we don't really expose as a generic property
right now. So for now I'm just extending the list.
Added:
Modified:
llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
llvm/test/Transforms/InstCombine/intrinsic-select.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 0708fb44b982b..028792895ebdc 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -531,11 +531,6 @@ static Instruction *foldCttzCtlz(IntrinsicInst &II, InstCombinerImpl &IC) {
return IC.replaceInstUsesWith(II, ConstantInt::getNullValue(II.getType()));
}
- // If the operand is a select with constant arm(s), try to hoist ctlz/cttz.
- if (auto *Sel = dyn_cast<SelectInst>(Op0))
- if (Instruction *R = IC.FoldOpIntoSelect(II, Sel))
- return R;
-
if (IsTZ) {
// cttz(-x) -> cttz(x)
if (match(Op0, m_Neg(m_Value(X))))
@@ -654,11 +649,6 @@ static Instruction *foldCtpop(IntrinsicInst &II, InstCombinerImpl &IC) {
return CastInst::Create(Instruction::ZExt, NarrowPop, Ty);
}
- // If the operand is a select with constant arm(s), try to hoist ctpop.
- if (auto *Sel = dyn_cast<SelectInst>(Op0))
- if (Instruction *R = IC.FoldOpIntoSelect(II, Sel))
- return R;
-
KnownBits Known(BitWidth);
IC.computeKnownBits(Op0, Known, 0, &II);
@@ -1615,11 +1605,6 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
if (Instruction *SAdd = matchSAddSubSat(*II))
return SAdd;
- if (match(I1, m_ImmConstant()))
- if (auto *Sel = dyn_cast<SelectInst>(I0))
- if (Instruction *R = FoldOpIntoSelect(*II, Sel))
- return R;
-
if (Value *NewMinMax = reassociateMinMaxWithConstants(II, Builder))
return replaceInstUsesWith(*II, NewMinMax);
@@ -3070,6 +3055,30 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
}
}
+ // Try to fold intrinsic into select operands. This is legal if:
+ // * The intrinsic is speculatable.
+ // * The select condition is not a vector, or the intrinsic does not
+ // perform cross-lane operations.
+ switch (IID) {
+ case Intrinsic::ctlz:
+ case Intrinsic::cttz:
+ case Intrinsic::ctpop:
+ case Intrinsic::umin:
+ case Intrinsic::umax:
+ case Intrinsic::smin:
+ case Intrinsic::smax:
+ case Intrinsic::usub_sat:
+ case Intrinsic::uadd_sat:
+ case Intrinsic::ssub_sat:
+ case Intrinsic::sadd_sat:
+ for (Value *Op : II->args())
+ if (auto *Sel = dyn_cast<SelectInst>(Op))
+ if (Instruction *R = FoldOpIntoSelect(*II, Sel))
+ return R;
+ default:
+ break;
+ }
+
if (Instruction *Shuf = foldShuffledIntrinsicOperands(II, Builder))
return Shuf;
diff --git a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
index de653a5b05123..7cae08df1171a 100644
--- a/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstructionCombining.cpp
@@ -1039,7 +1039,7 @@ static Constant *constantFoldOperationIntoSelectOperand(
else if (auto *C = dyn_cast<Constant>(Op))
ConstOps.push_back(C);
else
- llvm_unreachable("Operands should be select or constant");
+ return nullptr;
}
return ConstantFoldInstOperands(&I, ConstOps, I.getModule()->getDataLayout());
}
diff --git a/llvm/test/Transforms/InstCombine/intrinsic-select.ll b/llvm/test/Transforms/InstCombine/intrinsic-select.ll
index f057bbd0f4e77..a203b28bcb82a 100644
--- a/llvm/test/Transforms/InstCombine/intrinsic-select.ll
+++ b/llvm/test/Transforms/InstCombine/intrinsic-select.ll
@@ -151,8 +151,7 @@ define i32 @ctpop_sel_const_true_false_extra_use(i1 %b) {
define i32 @usub_sat_rhs_const_select_all_const(i1 %b) {
; CHECK-LABEL: @usub_sat_rhs_const_select_all_const(
-; CHECK-NEXT: [[S:%.*]] = select i1 [[B:%.*]], i32 5, i32 10
-; CHECK-NEXT: [[C:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[S]], i32 7)
+; CHECK-NEXT: [[C:%.*]] = select i1 [[B:%.*]], i32 0, i32 3
; CHECK-NEXT: ret i32 [[C]]
;
%s = select i1 %b, i32 5, i32 10
@@ -173,8 +172,8 @@ define i32 @usub_sat_rhs_var_select_all_const(i1 %b, i32 %x) {
define i32 @usub_sat_rhs_const_select_one_const(i1 %b, i32 %x) {
; CHECK-LABEL: @usub_sat_rhs_const_select_one_const(
-; CHECK-NEXT: [[S:%.*]] = select i1 [[B:%.*]], i32 5, i32 [[X:%.*]]
-; CHECK-NEXT: [[C:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[S]], i32 7)
+; CHECK-NEXT: [[TMP1:%.*]] = call i32 @llvm.usub.sat.i32(i32 [[X:%.*]], i32 7)
+; CHECK-NEXT: [[C:%.*]] = select i1 [[B:%.*]], i32 0, i32 [[TMP1]]
; CHECK-NEXT: ret i32 [[C]]
;
%s = select i1 %b, i32 5, i32 %x
@@ -195,8 +194,7 @@ define i32 @usub_sat_rhs_const_select_no_const(i1 %b, i32 %x, i32 %y) {
define i32 @usub_sat_lhs_const_select_all_const(i1 %b) {
; CHECK-LABEL: @usub_sat_lhs_const_select_all_const(
-; CHECK-NEXT: [[S:%.*]] = select i1 [[B:%.*]], i32 5, i32 10
-; CHECK-NEXT: [[C:%.*]] = call i32 @llvm.usub.sat.i32(i32 7, i32 [[S]])
+; CHECK-NEXT: [[C:%.*]] = select i1 [[B:%.*]], i32 2, i32 0
; CHECK-NEXT: ret i32 [[C]]
;
%s = select i1 %b, i32 5, i32 10
More information about the llvm-commits
mailing list