[PATCH] D124109: [RISCV] Add a DAG combine to pre-promote (i32 (and (srl X, Y), 1)) with Zbs on RV64.
Craig Topper via Phabricator via llvm-commits
llvm-commits at lists.llvm.org
Wed Apr 20 10:30:56 PDT 2022
craig.topper created this revision.
craig.topper added reviewers: asb, luismarques, kito-cheng, jrtc27.
Herald added subscribers: sunshaoce, VincentWu, luke957, StephenFan, vkmr, frasercrmck, evandro, apazos, sameer.abuasal, s.egerton, Jim, benna, psnobl, jocewei, PkmX, the_o, brucehoult, MartinMosbeck, rogfer01, edward-jones, zzheng, niosHD, sabuasal, simoncook, johnrusso, rbar, hiraditya, arichardson.
Herald added a project: All.
craig.topper requested review of this revision.
Herald added subscribers: pcwang-thead, eopXD, MaskRay.
Herald added a project: LLVM.
Type legalization will want to turn (srl X, Y) into RISCVISD::SRLW,
which will prevent us from using a BEXT instruction.
I don't think there is any precedent for type promotion checking
users to decide how to promote. Instead, I've added this DAG combine to
do it before type legalization.
Repository:
rG LLVM Github Monorepo
https://reviews.llvm.org/D124109
Files:
llvm/lib/Target/RISCV/RISCVISelLowering.cpp
llvm/test/CodeGen/RISCV/rv64zbs.ll
Index: llvm/test/CodeGen/RISCV/rv64zbs.ll
===================================================================
--- llvm/test/CodeGen/RISCV/rv64zbs.ll
+++ llvm/test/CodeGen/RISCV/rv64zbs.ll
@@ -347,8 +347,8 @@
;
; RV64ZBS-LABEL: bext_i32:
; RV64ZBS: # %bb.0:
-; RV64ZBS-NEXT: srlw a0, a0, a1
-; RV64ZBS-NEXT: andi a0, a0, 1
+; RV64ZBS-NEXT: andi a1, a1, 31
+; RV64ZBS-NEXT: bext a0, a0, a1
; RV64ZBS-NEXT: ret
%and = and i32 %b, 31
%shr = lshr i32 %a, %and
@@ -365,8 +365,7 @@
;
; RV64ZBS-LABEL: bext_i32_no_mask:
; RV64ZBS: # %bb.0:
-; RV64ZBS-NEXT: srlw a0, a0, a1
-; RV64ZBS-NEXT: andi a0, a0, 1
+; RV64ZBS-NEXT: bext a0, a0, a1
; RV64ZBS-NEXT: ret
%shr = lshr i32 %a, %b
%and1 = and i32 %shr, 1
Index: llvm/lib/Target/RISCV/RISCVISelLowering.cpp
===================================================================
--- llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -7922,7 +7922,26 @@
return combineSelectAndUse(N, N1, N0, DAG, /*AllOnes*/ false);
}
-static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue performANDCombine(SDNode *N, SelectionDAG &DAG,
+ const RISCVSubtarget &Subtarget) {
+ SDValue N0 = N->getOperand(0);
+ // Pre-promote (i32 (and (srl X, Y), 1)) on RV64 with Zbs without zero
+ // extending X. This is safe since we only need the LSB after the shift and
+ // shift amounts larger than 31 would produce poison. If we wait until
+ // type legalization, we'll create RISCVISD::SRLW and we can't recover it
+ // to use a BEXT instruction.
+ if (Subtarget.is64Bit() && Subtarget.hasStdExtZbs() &&
+ N->getValueType(0) == MVT::i32 && N0.getOpcode() == ISD::SRL &&
+ isOneConstant(N->getOperand(1)) && N0.hasOneUse()) {
+ SDLoc DL(N);
+ SDValue Op0 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, N0.getOperand(0));
+ SDValue Op1 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, N0.getOperand(1));
+ SDValue Srl = DAG.getNode(ISD::SRL, DL, MVT::i64, Op0, Op1);
+ SDValue And = DAG.getNode(ISD::AND, DL, MVT::i64, Srl,
+ DAG.getConstant(1, DL, MVT::i64));
+ return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, And);
+ }
+
// fold (and (select lhs, rhs, cc, -1, y), x) ->
// (select lhs, rhs, cc, x, (and x, y))
return combineSelectAndUseCommutative(N, DAG, /*AllOnes*/ true);
@@ -8564,7 +8583,7 @@
case ISD::SUB:
return performSUBCombine(N, DAG);
case ISD::AND:
- return performANDCombine(N, DAG);
+ return performANDCombine(N, DAG, Subtarget);
case ISD::OR:
return performORCombine(N, DAG, Subtarget);
case ISD::XOR:
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D124109.423954.patch
Type: text/x-patch
Size: 2701 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220420/3f03ac1a/attachment.bin>
More information about the llvm-commits
mailing list