[Hexagon] Lowering saturating subtraction #158726
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-backend-hexagon Author: Alexey Karyakin (quic-akaryaki) ChangesSaturating arithmetic can be expressed by llvm.uadd/usub.sat generic intrinsics. Full diff: https://github.com/llvm/llvm-project/pull/158726.diff 3 Files Affected:
diff --git a/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp b/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
index f1fa40c1b9036..adfcddf035011 100644
--- a/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
+++ b/llvm/lib/Target/Hexagon/HexagonISelLoweringHVX.cpp
@@ -208,6 +208,8 @@ HexagonTargetLowering::initializeHVXLowering() {
setOperationAction(ISD::SPLAT_VECTOR, T, Legal);
setOperationAction(ISD::UADDSAT, T, Legal);
setOperationAction(ISD::SADDSAT, T, Legal);
+ setOperationAction(ISD::USUBSAT, T, Legal);
+ setOperationAction(ISD::SSUBSAT, T, Legal);
if (T != ByteV) {
setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, T, Legal);
setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, T, Legal);
@@ -302,6 +304,8 @@ HexagonTargetLowering::initializeHVXLowering() {
setOperationAction(ISD::UADDSAT, T, Legal);
setOperationAction(ISD::SADDSAT, T, Legal);
setOperationAction(ISD::SUB, T, Legal);
+ setOperationAction(ISD::USUBSAT, T, Legal);
+ setOperationAction(ISD::SSUBSAT, T, Legal);
setOperationAction(ISD::MUL, T, Custom);
setOperationAction(ISD::MULHS, T, Custom);
setOperationAction(ISD::MULHU, T, Custom);
diff --git a/llvm/lib/Target/Hexagon/HexagonPatternsHVX.td b/llvm/lib/Target/Hexagon/HexagonPatternsHVX.td
index fb2ef59d99ef1..1637b91f1fa12 100644
--- a/llvm/lib/Target/Hexagon/HexagonPatternsHVX.td
+++ b/llvm/lib/Target/Hexagon/HexagonPatternsHVX.td
@@ -441,6 +441,21 @@ let Predicates = [UseHVX] in {
def: OpR_RR_pat_sat<V6_vaddwsat_dv, saddsat, VecPI32, HWI32>;
}
+let Predicates = [UseHVX] in {
+ def: OpR_RR_pat_sat<V6_vsububsat, usubsat, VecI8, HVI8>;
+ def: OpR_RR_pat_sat<V6_vsubuhsat, usubsat, VecI16, HVI16>;
+ def: OpR_RR_pat_sat<V6_vsubuwsat, usubsat, VecI32, HVI32>;
+ def: OpR_RR_pat_sat<V6_vsubbsat, ssubsat, VecI8, HVI8>;
+ def: OpR_RR_pat_sat<V6_vsubhsat, ssubsat, VecI16, HVI16>;
+ def: OpR_RR_pat_sat<V6_vsubwsat, ssubsat, VecI32, HVI32>;
+ def: OpR_RR_pat_sat<V6_vsububsat_dv, usubsat, VecPI8, HWI8>;
+ def: OpR_RR_pat_sat<V6_vsubuhsat_dv, usubsat, VecPI16, HWI16>;
+ def: OpR_RR_pat_sat<V6_vsubuwsat_dv, usubsat, VecPI32, HWI32>;
+ def: OpR_RR_pat_sat<V6_vsubbsat_dv, ssubsat, VecPI8, HWI8>;
+ def: OpR_RR_pat_sat<V6_vsubhsat_dv, ssubsat, VecPI16, HWI16>;
+ def: OpR_RR_pat_sat<V6_vsubwsat_dv, ssubsat, VecPI32, HWI32>;
+}
+
// For now, we always deal with vector floating point in SF mode.
class OpR_RR_pat_conv<InstHexagon MI, PatFrag Op, ValueType ResType,
PatFrag RsPred, PatFrag RtPred = RsPred>
diff --git a/llvm/test/CodeGen/Hexagon/vsubsat.ll b/llvm/test/CodeGen/Hexagon/vsubsat.ll
new file mode 100644
index 0000000000000..bb65aff166e23
--- /dev/null
+++ b/llvm/test/CodeGen/Hexagon/vsubsat.ll
@@ -0,0 +1,99 @@
+; RUN: llc -march=hexagon -mattr=+hvxv73,+hvx-length128b < %s | FileCheck %s
+
+;; Saturating subtraction.
+
+; CHECK-LABEL: vsububsat
+; CHECK: v[[#]].ub = vsub(v[[#]].ub,v[[#]].ub):sat
+define dso_local <128 x i8> @vsububsat(<128 x i8> %x, <128 x i8> %y) {
+entry:
+ %0 = tail call <128 x i8> @llvm.usub.sat.v128i8(<128 x i8> %x, <128 x i8> %y)
+ ret <128 x i8> %0
+}
+
+; CHECK-LABEL: vsubuhsat
+; CHECK: v[[#]].uh = vsub(v[[#]].uh,v[[#]].uh):sat
+define dso_local <64 x i16> @vsubuhsat(<64 x i16> %x, <64 x i16> %y) {
+entry:
+ %0 = tail call <64 x i16> @llvm.usub.sat.v64i16(<64 x i16> %x, <64 x i16> %y)
+ ret <64 x i16> %0
+}
+
+; CHECK-LABEL: vsubuwsat
+; CHECK: v[[#]].uw = vsub(v[[#]].uw,v[[#]].uw):sat
+define dso_local <32 x i32> @vsubuwsat(<32 x i32> %x, <32 x i32> %y) {
+entry:
+ %0 = tail call <32 x i32> @llvm.usub.sat.v32i32(<32 x i32> %x, <32 x i32> %y)
+ ret <32 x i32> %0
+}
+
+; CHECK-LABEL: vsubbsat
+; CHECK: v[[#]].b = vsub(v[[#]].b,v[[#]].b):sat
+define dso_local <128 x i8> @vsubbsat(<128 x i8> %x, <128 x i8> %y) {
+entry:
+ %0 = tail call <128 x i8> @llvm.ssub.sat.v128i8(<128 x i8> %x, <128 x i8> %y)
+ ret <128 x i8> %0
+}
+
+; CHECK-LABEL: vsubhsat
+; CHECK: v[[#]].h = vsub(v[[#]].h,v[[#]].h):sat
+define dso_local <64 x i16> @vsubhsat(<64 x i16> %x, <64 x i16> %y) {
+entry:
+ %0 = tail call <64 x i16> @llvm.ssub.sat.v64i16(<64 x i16> %x, <64 x i16> %y)
+ ret <64 x i16> %0
+}
+
+; CHECK-LABEL: vsubwsat
+; CHECK: v[[#]].w = vsub(v[[#]].w,v[[#]].w):sat
+define dso_local <32 x i32> @vsubwsat(<32 x i32> %x, <32 x i32> %y) {
+entry:
+ %0 = tail call <32 x i32> @llvm.ssub.sat.v32i32(<32 x i32> %x, <32 x i32> %y)
+ ret <32 x i32> %0
+}
+
+; CHECK-LABEL: vsububsat_dv
+; CHECK: v[[#]]:[[#]].ub = vsub(v[[#]]:[[#]].ub,v[[#]]:[[#]].ub):sat
+define dso_local <256 x i8> @vsububsat_dv(<256 x i8> %x, <256 x i8> %y) {
+entry:
+ %0 = tail call <256 x i8> @llvm.usub.sat.v256i8(<256 x i8> %x, <256 x i8> %y)
+ ret <256 x i8> %0
+}
+
+; CHECK-LABEL: vsubuhsat_dv
+; CHECK: v[[#]]:[[#]].uh = vsub(v[[#]]:[[#]].uh,v[[#]]:[[#]].uh):sat
+define dso_local <128 x i16> @vsubuhsat_dv(<128 x i16> %x, <128 x i16> %y) {
+entry:
+ %0 = tail call <128 x i16> @llvm.usub.sat.v128i16(<128 x i16> %x, <128 x i16> %y)
+ ret <128 x i16> %0
+}
+
+; CHECK-LABEL: vsubuwsat_dv
+; CHECK: v[[#]]:[[#]].uw = vsub(v[[#]]:[[#]].uw,v[[#]]:[[#]].uw):sat
+define dso_local <64 x i32> @vsubuwsat_dv(<64 x i32> %x, <64 x i32> %y) {
+entry:
+ %0 = tail call <64 x i32> @llvm.usub.sat.v64i32(<64 x i32> %x, <64 x i32> %y)
+ ret <64 x i32> %0
+}
+
+; CHECK-LABEL: vsubbsat_dv
+; CHECK: v[[#]]:[[#]].b = vsub(v[[#]]:[[#]].b,v[[#]]:[[#]].b):sat
+define dso_local <256 x i8> @vsubbsat_dv(<256 x i8> %x, <256 x i8> %y) {
+entry:
+ %0 = tail call <256 x i8> @llvm.ssub.sat.v256i8(<256 x i8> %x, <256 x i8> %y)
+ ret <256 x i8> %0
+}
+
+; CHECK-LABEL: vsubhsat_dv
+; CHECK: v[[#]]:[[#]].h = vsub(v[[#]]:[[#]].h,v[[#]]:[[#]].h):sat
+define dso_local <128 x i16> @vsubhsat_dv(<128 x i16> %x, <128 x i16> %y) {
+entry:
+ %0 = tail call <128 x i16> @llvm.ssub.sat.v128i16(<128 x i16> %x, <128 x i16> %y)
+ ret <128 x i16> %0
+}
+
+; CHECK-LABEL: vsubwsat_dv
+; CHECK: v[[#]]:[[#]].w = vsub(v[[#]]:[[#]].w,v[[#]]:[[#]].w):sat
+define dso_local <64 x i32> @vsubwsat_dv(<64 x i32> %x, <64 x i32> %y) {
+entry:
+ %0 = tail call <64 x i32> @llvm.ssub.sat.v64i32(<64 x i32> %x, <64 x i32> %y)
+ ret <64 x i32> %0
+}
|
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
Saturating arithmetic can be expressed by llvm.uadd/usub.sat generic intrinsics. Change-Id: Iead2ed8c6cc1378dc2a08eb33a6950aec5b8d9be JIRA: https://jira-dc.qualcomm.com/jira/browse/Hexagon Change-Id: I86c4734d81a5c581777763a4b8d224d3727e2bba
quic-akaryaki
force-pushed
the
hexagon-satsub
branch
from
a2ce907 to
ea8f914
Compare
androm3da
approved these changes
Sep 15, 2025
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Saturating arithmetic can be expressed by llvm.uadd/usub.sat generic intrinsics.