[llvm] [RISCV] Use bset+addi for (not (sll -1, X)). (PR #72549)
Craig Topper via llvm-commits
llvm-commits at lists.llvm.org
Thu Nov 16 10:21:01 PST 2023
https://github.com/topperc created https://github.com/llvm/llvm-project/pull/72549
This is an alternative to #71420 that handles i32 on RV64 safely by pre-promoting the pattern in DAG combine.
>From 93bf2acfb5c2ef0249922bec38e8734a2fdeb452 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Thu, 16 Nov 2023 10:01:42 -0800
Subject: [PATCH 1/2] [RISCV] Add test cases for (not (sll -1, X)) for Zbs. NFC
We can use (ADDI (BSET X0, X), -1).
---
llvm/test/CodeGen/RISCV/rv32zbs.ll | 81 ++++++++++++++++++++++++++++++
llvm/test/CodeGen/RISCV/rv64zbs.ll | 50 ++++++++++++++++++
2 files changed, 131 insertions(+)
diff --git a/llvm/test/CodeGen/RISCV/rv32zbs.ll b/llvm/test/CodeGen/RISCV/rv32zbs.ll
index 460d15991788238..55c27e69d38fd41 100644
--- a/llvm/test/CodeGen/RISCV/rv32zbs.ll
+++ b/llvm/test/CodeGen/RISCV/rv32zbs.ll
@@ -744,3 +744,84 @@ define i32 @or_i32_66901(i32 %a) nounwind {
%or = or i32 %a, 66901
ret i32 %or
}
+
+define i32 @bset_trailing_ones_i32_mask(i32 %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i32_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: sll a0, a1, a0
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %and = and i32 %a, 31
+ %shift = shl nsw i32 -1, %and
+ %not = xor i32 %shift, -1
+ ret i32 %not
+}
+
+define i32 @bset_trailing_ones_i32_no_mask(i32 %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i32_no_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: sll a0, a1, a0
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %shift = shl nsw i32 -1, %a
+ %not = xor i32 %shift, -1
+ ret i32 %not
+}
+
+define i64 @bset_trailing_ones_i64_mask(i64 %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i64_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a2, -1
+; CHECK-NEXT: andi a3, a0, 63
+; CHECK-NEXT: addi a1, a3, -32
+; CHECK-NEXT: sll a0, a2, a0
+; CHECK-NEXT: bltz a1, .LBB43_2
+; CHECK-NEXT: # %bb.1:
+; CHECK-NEXT: sll a2, a2, a3
+; CHECK-NEXT: j .LBB43_3
+; CHECK-NEXT: .LBB43_2:
+; CHECK-NEXT: not a2, a3
+; CHECK-NEXT: lui a3, 524288
+; CHECK-NEXT: addi a3, a3, -1
+; CHECK-NEXT: srl a2, a3, a2
+; CHECK-NEXT: or a2, a0, a2
+; CHECK-NEXT: .LBB43_3:
+; CHECK-NEXT: srai a1, a1, 31
+; CHECK-NEXT: and a0, a1, a0
+; CHECK-NEXT: not a1, a2
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %and = and i64 %a, 63
+ %shift = shl nsw i64 -1, %and
+ %not = xor i64 %shift, -1
+ ret i64 %not
+}
+
+define i64 @bset_trailing_ones_i64_no_mask(i64 %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i64_no_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: addi a2, a0, -32
+; CHECK-NEXT: sll a1, a1, a0
+; CHECK-NEXT: bltz a2, .LBB44_2
+; CHECK-NEXT: # %bb.1:
+; CHECK-NEXT: mv a0, a1
+; CHECK-NEXT: j .LBB44_3
+; CHECK-NEXT: .LBB44_2:
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: lui a3, 524288
+; CHECK-NEXT: addi a3, a3, -1
+; CHECK-NEXT: srl a0, a3, a0
+; CHECK-NEXT: or a0, a1, a0
+; CHECK-NEXT: .LBB44_3:
+; CHECK-NEXT: srai a2, a2, 31
+; CHECK-NEXT: and a2, a2, a1
+; CHECK-NEXT: not a1, a0
+; CHECK-NEXT: not a0, a2
+; CHECK-NEXT: ret
+ %shift = shl nsw i64 -1, %a
+ %not = xor i64 %shift, -1
+ ret i64 %not
+}
diff --git a/llvm/test/CodeGen/RISCV/rv64zbs.ll b/llvm/test/CodeGen/RISCV/rv64zbs.ll
index b30b3c15196076b..bf822caa37341c7 100644
--- a/llvm/test/CodeGen/RISCV/rv64zbs.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zbs.ll
@@ -1071,3 +1071,53 @@ define i64 @or_i64_66901(i64 %a) nounwind {
%or = or i64 %a, 66901
ret i64 %or
}
+
+define signext i32 @bset_trailing_ones_i32_mask(i32 signext %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i32_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: sllw a0, a1, a0
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %and = and i32 %a, 31
+ %shift = shl nsw i32 -1, %and
+ %not = xor i32 %shift, -1
+ ret i32 %not
+}
+
+define signext i32 @bset_trailing_ones_i32_no_mask(i32 signext %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i32_no_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: sllw a0, a1, a0
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %shift = shl nsw i32 -1, %a
+ %not = xor i32 %shift, -1
+ ret i32 %not
+}
+
+define signext i64 @bset_trailing_ones_i64_mask(i64 signext %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i64_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: sll a0, a1, a0
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %and = and i64 %a, 63
+ %shift = shl nsw i64 -1, %and
+ %not = xor i64 %shift, -1
+ ret i64 %not
+}
+
+define signext i64 @bset_trailing_ones_i64_no_mask(i64 signext %a) nounwind {
+; CHECK-LABEL: bset_trailing_ones_i64_no_mask:
+; CHECK: # %bb.0:
+; CHECK-NEXT: li a1, -1
+; CHECK-NEXT: sll a0, a1, a0
+; CHECK-NEXT: not a0, a0
+; CHECK-NEXT: ret
+ %shift = shl nsw i64 -1, %a
+ %not = xor i64 %shift, -1
+ ret i64 %not
+}
>From c8cebfce007eb43b9f6e3f38f05348f0caced6b4 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper at sifive.com>
Date: Thu, 16 Nov 2023 10:03:52 -0800
Subject: [PATCH 2/2] [RISCV] Use bset+addi for (not (sll -1, X)).
This is an alternative to #71420 that handles i32 on RV64 safely
by pre-promoting the pattern in DAG combine.
---
llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 15 +++++
llvm/lib/Target/RISCV/RISCVInstrInfoZb.td | 2 +
llvm/test/CodeGen/RISCV/rv32zbs.ll | 36 ++++++----
llvm/test/CodeGen/RISCV/rv64zbs.ll | 73 ++++++++++++++-------
4 files changed, 90 insertions(+), 36 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index fc0f59a09315b76..f89f300a4e9e50c 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -12362,6 +12362,21 @@ static SDValue performXORCombine(SDNode *N, SelectionDAG &DAG,
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
+ // Pre-promote (i32 (xor (shl -1, X), ~0)) on RV64 with Zbs so we can use
+ // (ADDI (BSET X0, X), -1). If we wait until/ type legalization, we'll create
+ // RISCVISD:::SLLW and we can't recover it to use a BSET instruction.
+ if (!RV64LegalI32 && Subtarget.is64Bit() && Subtarget.hasStdExtZbs() &&
+ N->getValueType(0) == MVT::i32 && isAllOnesConstant(N1) &&
+ N0.getOpcode() == ISD::SHL && isAllOnesConstant(N0.getOperand(0)) &&
+ !isa<ConstantSDNode>(N0.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 Shl = DAG.getNode(ISD::SHL, DL, MVT::i64, Op0, Op1);
+ SDValue And = DAG.getNOT(DL, Shl, MVT::i64);
+ return DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, And);
+ }
+
// fold (xor (sllw 1, x), -1) -> (rolw ~1, x)
// NOTE: Assumes ROL being legal means ROLW is legal.
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
diff --git a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
index d1c29842b85d38a..f7f8560b57b5c2d 100644
--- a/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
+++ b/llvm/lib/Target/RISCV/RISCVInstrInfoZb.td
@@ -554,6 +554,8 @@ def : Pat<(XLenVT (and (shiftop<srl> GPR:$rs1, (XLenVT GPR:$rs2)), 1)),
def : Pat<(XLenVT (shiftop<shl> 1, (XLenVT GPR:$rs2))),
(BSET (XLenVT X0), GPR:$rs2)>;
+def : Pat<(XLenVT (not (shiftop<shl> -1, (XLenVT GPR:$rs2)))),
+ (ADDI (BSET (XLenVT X0), GPR:$rs2), -1)>;
def : Pat<(XLenVT (and GPR:$rs1, BCLRMask:$mask)),
(BCLRI GPR:$rs1, BCLRMask:$mask)>;
diff --git a/llvm/test/CodeGen/RISCV/rv32zbs.ll b/llvm/test/CodeGen/RISCV/rv32zbs.ll
index 55c27e69d38fd41..ccda8f4e5dd0595 100644
--- a/llvm/test/CodeGen/RISCV/rv32zbs.ll
+++ b/llvm/test/CodeGen/RISCV/rv32zbs.ll
@@ -746,12 +746,18 @@ define i32 @or_i32_66901(i32 %a) nounwind {
}
define i32 @bset_trailing_ones_i32_mask(i32 %a) nounwind {
-; CHECK-LABEL: bset_trailing_ones_i32_mask:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, -1
-; CHECK-NEXT: sll a0, a1, a0
-; CHECK-NEXT: not a0, a0
-; CHECK-NEXT: ret
+; RV32I-LABEL: bset_trailing_ones_i32_mask:
+; RV32I: # %bb.0:
+; RV32I-NEXT: li a1, -1
+; RV32I-NEXT: sll a0, a1, a0
+; RV32I-NEXT: not a0, a0
+; RV32I-NEXT: ret
+;
+; RV32ZBS-LABEL: bset_trailing_ones_i32_mask:
+; RV32ZBS: # %bb.0:
+; RV32ZBS-NEXT: bset a0, zero, a0
+; RV32ZBS-NEXT: addi a0, a0, -1
+; RV32ZBS-NEXT: ret
%and = and i32 %a, 31
%shift = shl nsw i32 -1, %and
%not = xor i32 %shift, -1
@@ -759,12 +765,18 @@ define i32 @bset_trailing_ones_i32_mask(i32 %a) nounwind {
}
define i32 @bset_trailing_ones_i32_no_mask(i32 %a) nounwind {
-; CHECK-LABEL: bset_trailing_ones_i32_no_mask:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, -1
-; CHECK-NEXT: sll a0, a1, a0
-; CHECK-NEXT: not a0, a0
-; CHECK-NEXT: ret
+; RV32I-LABEL: bset_trailing_ones_i32_no_mask:
+; RV32I: # %bb.0:
+; RV32I-NEXT: li a1, -1
+; RV32I-NEXT: sll a0, a1, a0
+; RV32I-NEXT: not a0, a0
+; RV32I-NEXT: ret
+;
+; RV32ZBS-LABEL: bset_trailing_ones_i32_no_mask:
+; RV32ZBS: # %bb.0:
+; RV32ZBS-NEXT: bset a0, zero, a0
+; RV32ZBS-NEXT: addi a0, a0, -1
+; RV32ZBS-NEXT: ret
%shift = shl nsw i32 -1, %a
%not = xor i32 %shift, -1
ret i32 %not
diff --git a/llvm/test/CodeGen/RISCV/rv64zbs.ll b/llvm/test/CodeGen/RISCV/rv64zbs.ll
index bf822caa37341c7..016b0924eaf1993 100644
--- a/llvm/test/CodeGen/RISCV/rv64zbs.ll
+++ b/llvm/test/CodeGen/RISCV/rv64zbs.ll
@@ -1073,12 +1073,19 @@ define i64 @or_i64_66901(i64 %a) nounwind {
}
define signext i32 @bset_trailing_ones_i32_mask(i32 signext %a) nounwind {
-; CHECK-LABEL: bset_trailing_ones_i32_mask:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, -1
-; CHECK-NEXT: sllw a0, a1, a0
-; CHECK-NEXT: not a0, a0
-; CHECK-NEXT: ret
+; RV64I-LABEL: bset_trailing_ones_i32_mask:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, -1
+; RV64I-NEXT: sllw a0, a1, a0
+; RV64I-NEXT: not a0, a0
+; RV64I-NEXT: ret
+;
+; RV64ZBS-LABEL: bset_trailing_ones_i32_mask:
+; RV64ZBS: # %bb.0:
+; RV64ZBS-NEXT: andi a0, a0, 31
+; RV64ZBS-NEXT: bset a0, zero, a0
+; RV64ZBS-NEXT: addi a0, a0, -1
+; RV64ZBS-NEXT: ret
%and = and i32 %a, 31
%shift = shl nsw i32 -1, %and
%not = xor i32 %shift, -1
@@ -1086,24 +1093,36 @@ define signext i32 @bset_trailing_ones_i32_mask(i32 signext %a) nounwind {
}
define signext i32 @bset_trailing_ones_i32_no_mask(i32 signext %a) nounwind {
-; CHECK-LABEL: bset_trailing_ones_i32_no_mask:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, -1
-; CHECK-NEXT: sllw a0, a1, a0
-; CHECK-NEXT: not a0, a0
-; CHECK-NEXT: ret
+; RV64I-LABEL: bset_trailing_ones_i32_no_mask:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, -1
+; RV64I-NEXT: sllw a0, a1, a0
+; RV64I-NEXT: not a0, a0
+; RV64I-NEXT: ret
+;
+; RV64ZBS-LABEL: bset_trailing_ones_i32_no_mask:
+; RV64ZBS: # %bb.0:
+; RV64ZBS-NEXT: bset a0, zero, a0
+; RV64ZBS-NEXT: addiw a0, a0, -1
+; RV64ZBS-NEXT: ret
%shift = shl nsw i32 -1, %a
%not = xor i32 %shift, -1
ret i32 %not
}
define signext i64 @bset_trailing_ones_i64_mask(i64 signext %a) nounwind {
-; CHECK-LABEL: bset_trailing_ones_i64_mask:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, -1
-; CHECK-NEXT: sll a0, a1, a0
-; CHECK-NEXT: not a0, a0
-; CHECK-NEXT: ret
+; RV64I-LABEL: bset_trailing_ones_i64_mask:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, -1
+; RV64I-NEXT: sll a0, a1, a0
+; RV64I-NEXT: not a0, a0
+; RV64I-NEXT: ret
+;
+; RV64ZBS-LABEL: bset_trailing_ones_i64_mask:
+; RV64ZBS: # %bb.0:
+; RV64ZBS-NEXT: bset a0, zero, a0
+; RV64ZBS-NEXT: addi a0, a0, -1
+; RV64ZBS-NEXT: ret
%and = and i64 %a, 63
%shift = shl nsw i64 -1, %and
%not = xor i64 %shift, -1
@@ -1111,12 +1130,18 @@ define signext i64 @bset_trailing_ones_i64_mask(i64 signext %a) nounwind {
}
define signext i64 @bset_trailing_ones_i64_no_mask(i64 signext %a) nounwind {
-; CHECK-LABEL: bset_trailing_ones_i64_no_mask:
-; CHECK: # %bb.0:
-; CHECK-NEXT: li a1, -1
-; CHECK-NEXT: sll a0, a1, a0
-; CHECK-NEXT: not a0, a0
-; CHECK-NEXT: ret
+; RV64I-LABEL: bset_trailing_ones_i64_no_mask:
+; RV64I: # %bb.0:
+; RV64I-NEXT: li a1, -1
+; RV64I-NEXT: sll a0, a1, a0
+; RV64I-NEXT: not a0, a0
+; RV64I-NEXT: ret
+;
+; RV64ZBS-LABEL: bset_trailing_ones_i64_no_mask:
+; RV64ZBS: # %bb.0:
+; RV64ZBS-NEXT: bset a0, zero, a0
+; RV64ZBS-NEXT: addi a0, a0, -1
+; RV64ZBS-NEXT: ret
%shift = shl nsw i64 -1, %a
%not = xor i64 %shift, -1
ret i64 %not
More information about the llvm-commits
mailing list