[llvm] [RISCV] Lower SELECT's with one constant more efficiently using Zicond (PR #143581)
Ryan Buchner via llvm-commits
llvm-commits at lists.llvm.org
Thu Jun 12 12:36:52 PDT 2025
https://github.com/bababuck updated https://github.com/llvm/llvm-project/pull/143581
>From b04addc15112f00048a26d5b7e933bea36bd20b3 Mon Sep 17 00:00:00 2001
From: bababuck <buchner.ryan at gmail.com>
Date: Thu, 12 Jun 2025 11:56:02 -0700
Subject: [PATCH] [RISCV] Efficiently lower select %cond, %constant, %variable
using zicond extension
Provides the following lowerings, which reduces instruction count by 1 for rv64:
(select cond, const, t) -> (add (czero_nez t - c1, cond), const)
(select cond, t, const) -> (add (czero_eqz t - c1, cond), const)
For the special case of c1 == -2048
(select cond, -2048, t) -> (xor (czero_nez (xor t, -2048), cond), -2048)
(select cond, t, -2048) -> (xor (czero_eqz (xor t, -2048), cond), -2048)
---
llvm/lib/Target/RISCV/RISCVISelLowering.cpp | 26 +++++++++++
.../CodeGen/RISCV/short-forward-branch-opt.ll | 46 ++++++++++++-------
llvm/test/CodeGen/RISCV/zicond-opts.ll | 38 +++++++--------
3 files changed, 72 insertions(+), 38 deletions(-)
diff --git a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
index a157c94849f37..bb9951ea15355 100644
--- a/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelLowering.cpp
@@ -9096,6 +9096,32 @@ SDValue RISCVTargetLowering::lowerSELECT(SDValue Op, SelectionDAG &DAG) const {
return DAG.getNode(ISD::ADD, DL, VT, CMOV, RHSVal);
}
+ // (select c, c1, t) -> (add (czero_nez t - c1, c), c1)
+ // (select c, t, c1) -> (add (czero_eqz t - c1, c), c1)
+ if (isa<ConstantSDNode>(TrueV) != isa<ConstantSDNode>(FalseV)) {
+ bool IsCZERO_NEZ = isa<ConstantSDNode>(TrueV);
+ SDValue ConstVal = IsCZERO_NEZ ? TrueV : FalseV;
+ SDValue RegV = IsCZERO_NEZ ? FalseV : TrueV;
+ int64_t RawConstVal = cast<ConstantSDNode>(ConstVal)->getSExtValue();
+ // Fall back to XORI if Const == -0x800
+ if (RawConstVal == -0x800) {
+ SDValue XorOp = DAG.getNode(ISD::XOR, DL, VT, RegV, ConstVal);
+ SDValue CMOV =
+ DAG.getNode(IsCZERO_NEZ ? RISCVISD::CZERO_NEZ : RISCVISD::CZERO_EQZ,
+ DL, VT, XorOp, CondV);
+ return DAG.getNode(ISD::XOR, DL, VT, CMOV, ConstVal);
+ }
+ // Efficient only if the constant and its negation fit into `ADDI`
+ // Prefer Add/Sub over Xor since can be compressed for small immediates
+ if (isInt<12>(RawConstVal)) {
+ SDValue SubOp = DAG.getNode(ISD::SUB, DL, VT, RegV, ConstVal);
+ SDValue CMOV =
+ DAG.getNode(IsCZERO_NEZ ? RISCVISD::CZERO_NEZ : RISCVISD::CZERO_EQZ,
+ DL, VT, SubOp, CondV);
+ return DAG.getNode(ISD::ADD, DL, VT, CMOV, ConstVal);
+ }
+ }
+
// (select c, t, f) -> (or (czero_eqz t, c), (czero_nez f, c))
// Unless we have the short forward branch optimization.
if (!Subtarget.hasConditionalMoveFusion())
diff --git a/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll b/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll
index b7b88584f3bdb..13c43a3875a08 100644
--- a/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll
+++ b/llvm/test/CodeGen/RISCV/short-forward-branch-opt.ll
@@ -173,14 +173,21 @@ define signext i32 @test6(i32 signext %x, i32 signext %z) {
; NOSFB-NEXT: or a0, a0, a1
; NOSFB-NEXT: ret
;
-; SFB-LABEL: test6:
-; SFB: # %bb.0:
-; SFB-NEXT: li a2, -1
-; SFB-NEXT: beqz a1, .LBB5_2
-; SFB-NEXT: # %bb.1:
-; SFB-NEXT: mv a0, a2
-; SFB-NEXT: .LBB5_2:
-; SFB-NEXT: ret
+; NOZICOND-LABEL: test6:
+; NOZICOND: # %bb.0:
+; NOZICOND-NEXT: li a2, -1
+; NOZICOND-NEXT: beqz a1, .LBB5_2
+; NOZICOND-NEXT: # %bb.1:
+; NOZICOND-NEXT: mv a0, a2
+; NOZICOND-NEXT: .LBB5_2:
+; NOZICOND-NEXT: ret
+;
+; ZICOND-LABEL: test6:
+; ZICOND: # %bb.0:
+; ZICOND-NEXT: addi a0, a0, 1
+; ZICOND-NEXT: czero.nez a0, a0, a1
+; ZICOND-NEXT: addi a0, a0, -1
+; ZICOND-NEXT: ret
%c = icmp eq i32 %z, 0
%b = select i1 %c, i32 %x, i32 -1
ret i32 %b
@@ -195,14 +202,21 @@ define signext i32 @test7(i32 signext %x, i32 signext %z) {
; NOSFB-NEXT: or a0, a0, a1
; NOSFB-NEXT: ret
;
-; SFB-LABEL: test7:
-; SFB: # %bb.0:
-; SFB-NEXT: li a2, -1
-; SFB-NEXT: bnez a1, .LBB6_2
-; SFB-NEXT: # %bb.1:
-; SFB-NEXT: mv a0, a2
-; SFB-NEXT: .LBB6_2:
-; SFB-NEXT: ret
+; NOZICOND-LABEL: test7:
+; NOZICOND: # %bb.0:
+; NOZICOND-NEXT: li a2, -1
+; NOZICOND-NEXT: bnez a1, .LBB6_2
+; NOZICOND-NEXT: # %bb.1:
+; NOZICOND-NEXT: mv a0, a2
+; NOZICOND-NEXT: .LBB6_2:
+; NOZICOND-NEXT: ret
+;
+; ZICOND-LABEL: test7:
+; ZICOND: # %bb.0:
+; ZICOND-NEXT: addi a0, a0, 1
+; ZICOND-NEXT: czero.eqz a0, a0, a1
+; ZICOND-NEXT: addi a0, a0, -1
+; ZICOND-NEXT: ret
%c = icmp eq i32 %z, 0
%b = select i1 %c, i32 -1, i32 %x
ret i32 %b
diff --git a/llvm/test/CodeGen/RISCV/zicond-opts.ll b/llvm/test/CodeGen/RISCV/zicond-opts.ll
index f5a25868bd12b..2512ba803cf48 100644
--- a/llvm/test/CodeGen/RISCV/zicond-opts.ll
+++ b/llvm/test/CodeGen/RISCV/zicond-opts.ll
@@ -146,20 +146,18 @@ define i64 @select_imm_reg(i64 %t, i1 %cond) {
; RV32ZICOND-LABEL: select_imm_reg:
; RV32ZICOND: # %bb.0:
; RV32ZICOND-NEXT: andi a2, a2, 1
-; RV32ZICOND-NEXT: li a3, 3
-; RV32ZICOND-NEXT: czero.nez a0, a0, a2
-; RV32ZICOND-NEXT: czero.eqz a3, a3, a2
-; RV32ZICOND-NEXT: or a0, a3, a0
+; RV32ZICOND-NEXT: addi a0, a0, -3
; RV32ZICOND-NEXT: czero.nez a1, a1, a2
+; RV32ZICOND-NEXT: czero.nez a0, a0, a2
+; RV32ZICOND-NEXT: addi a0, a0, 3
; RV32ZICOND-NEXT: ret
;
; RV64ZICOND-LABEL: select_imm_reg:
; RV64ZICOND: # %bb.0:
; RV64ZICOND-NEXT: andi a1, a1, 1
-; RV64ZICOND-NEXT: li a2, 3
+; RV64ZICOND-NEXT: addi a0, a0, -3
; RV64ZICOND-NEXT: czero.nez a0, a0, a1
-; RV64ZICOND-NEXT: czero.eqz a1, a2, a1
-; RV64ZICOND-NEXT: or a0, a1, a0
+; RV64ZICOND-NEXT: addi a0, a0, 3
; RV64ZICOND-NEXT: ret
%4 = select i1 %cond, i64 3, i64 %t
ret i64 %4
@@ -170,20 +168,18 @@ define i64 @select_reg_imm(i64 %t, i1 %cond) {
; RV32ZICOND-LABEL: select_reg_imm:
; RV32ZICOND: # %bb.0:
; RV32ZICOND-NEXT: andi a2, a2, 1
-; RV32ZICOND-NEXT: li a3, 3
-; RV32ZICOND-NEXT: czero.nez a3, a3, a2
-; RV32ZICOND-NEXT: czero.eqz a0, a0, a2
-; RV32ZICOND-NEXT: or a0, a0, a3
+; RV32ZICOND-NEXT: addi a0, a0, -3
; RV32ZICOND-NEXT: czero.eqz a1, a1, a2
+; RV32ZICOND-NEXT: czero.eqz a0, a0, a2
+; RV32ZICOND-NEXT: addi a0, a0, 3
; RV32ZICOND-NEXT: ret
;
; RV64ZICOND-LABEL: select_reg_imm:
; RV64ZICOND: # %bb.0:
; RV64ZICOND-NEXT: andi a1, a1, 1
-; RV64ZICOND-NEXT: li a2, 3
-; RV64ZICOND-NEXT: czero.nez a2, a2, a1
+; RV64ZICOND-NEXT: addi a0, a0, -3
; RV64ZICOND-NEXT: czero.eqz a0, a0, a1
-; RV64ZICOND-NEXT: or a0, a0, a2
+; RV64ZICOND-NEXT: addi a0, a0, 3
; RV64ZICOND-NEXT: ret
%4 = select i1 %cond, i64 %t, i64 3
ret i64 %4
@@ -194,21 +190,19 @@ define i64 @select_imm_reg_neg_2048(i64 %t, i1 %cond) {
; RV32ZICOND-LABEL: select_imm_reg_neg_2048:
; RV32ZICOND: # %bb.0:
; RV32ZICOND-NEXT: andi a2, a2, 1
-; RV32ZICOND-NEXT: li a3, -2048
+; RV32ZICOND-NEXT: xori a0, a0, -2048
+; RV32ZICOND-NEXT: neg a3, a2
; RV32ZICOND-NEXT: czero.nez a0, a0, a2
-; RV32ZICOND-NEXT: czero.eqz a3, a3, a2
-; RV32ZICOND-NEXT: neg a2, a2
-; RV32ZICOND-NEXT: or a0, a3, a0
-; RV32ZICOND-NEXT: or a1, a2, a1
+; RV32ZICOND-NEXT: or a1, a3, a1
+; RV32ZICOND-NEXT: xori a0, a0, -2048
; RV32ZICOND-NEXT: ret
;
; RV64ZICOND-LABEL: select_imm_reg_neg_2048:
; RV64ZICOND: # %bb.0:
; RV64ZICOND-NEXT: andi a1, a1, 1
-; RV64ZICOND-NEXT: li a2, -2048
+; RV64ZICOND-NEXT: xori a0, a0, -2048
; RV64ZICOND-NEXT: czero.nez a0, a0, a1
-; RV64ZICOND-NEXT: czero.eqz a1, a2, a1
-; RV64ZICOND-NEXT: or a0, a1, a0
+; RV64ZICOND-NEXT: xori a0, a0, -2048
; RV64ZICOND-NEXT: ret
%4 = select i1 %cond, i64 -2048, i64 %t
ret i64 %4
More information about the llvm-commits
mailing list