From 7cd725858b4b67ca44a8d34483c28b5e12a4c733 Mon Sep 17 00:00:00 2001
From: Craig Topper <craig.topper@sifive.com>
Date: Thu, 29 Dec 2022 16:33:18 -0800
Subject: [PATCH] [RISCV] RISCVDAGToDAGISel::selectShiftMask to shift by (sub
 size-1, X).

If the shift amount is (sub C, X) where C is -1 modulo the size of
the shift, we can replace the sub with a NOT.

We could also use XORI X, size-1, but NOT would work better with
c.not from the future Zce extension.
---
 llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp |  11 ++
 llvm/test/CodeGen/RISCV/shift-amount-mod.ll | 284 ++++++++++++++++++++++++++++
 2 files changed, 295 insertions(+)
 create mode 100644 llvm/test/CodeGen/RISCV/shift-amount-mod.ll

diff --git a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
index b147430..45c6620 100644
--- a/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
+++ b/llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp
@@ -2113,6 +2113,17 @@ bool RISCVDAGToDAGISel::selectShiftMask(SDValue N, unsigned ShiftWidth,
       ShAmt = SDValue(Neg, 0);
       return true;
     }
+    // If we are shifting by N-X where N == -1 mod Size, then just shift by ~X
+    // to generate a NOT instead of a SUB of a constant.
+    if (Imm % ShiftWidth == ShiftWidth - 1) {
+      SDLoc DL(ShAmt);
+      EVT VT = ShAmt.getValueType();
+      MachineSDNode *Not =
+          CurDAG->getMachineNode(RISCV::XORI, DL, VT, ShAmt.getOperand(1),
+                                 CurDAG->getTargetConstant(-1, DL, VT));
+      ShAmt = SDValue(Not, 0);
+      return true;
+    }
   }
 
   return true;
diff --git a/llvm/test/CodeGen/RISCV/shift-amount-mod.ll b/llvm/test/CodeGen/RISCV/shift-amount-mod.ll
new file mode 100644
index 0000000..ef69758
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/shift-amount-mod.ll
@@ -0,0 +1,284 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=RV32I
+; RUN: llc -mtriple=riscv64 -verify-machineinstrs < %s \
+; RUN:   | FileCheck %s -check-prefix=RV64I
+
+define i32 @shl_by_complemented_32(i32 %x) {
+; RV32I-LABEL: shl_by_complemented_32:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    not a1, a0
+; RV32I-NEXT:    sll a0, a0, a1
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl_by_complemented_32:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sllw a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i32 31, %x
+  %b = shl i32 %x, %a
+  ret i32 %b
+}
+
+define i64 @shl_by_complemented_64(i64 %x) {
+; RV32I-LABEL: shl_by_complemented_64:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    not a4, a0
+; RV32I-NEXT:    li a3, 31
+; RV32I-NEXT:    sub a3, a3, a0
+; RV32I-NEXT:    sll a2, a0, a4
+; RV32I-NEXT:    bltz a3, .LBB1_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    mv a1, a2
+; RV32I-NEXT:    j .LBB1_3
+; RV32I-NEXT:  .LBB1_2:
+; RV32I-NEXT:    sll a1, a1, a4
+; RV32I-NEXT:    li a4, 63
+; RV32I-NEXT:    sub a4, a4, a0
+; RV32I-NEXT:    xori a4, a4, 31
+; RV32I-NEXT:    srli a0, a0, 1
+; RV32I-NEXT:    srl a0, a0, a4
+; RV32I-NEXT:    or a1, a1, a0
+; RV32I-NEXT:  .LBB1_3:
+; RV32I-NEXT:    slti a0, a3, 0
+; RV32I-NEXT:    neg a0, a0
+; RV32I-NEXT:    and a0, a0, a2
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl_by_complemented_64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sll a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i64 63, %x
+  %b = shl i64 %x, %a
+  ret i64 %b
+}
+
+define i32 @lshr_by_complemented_32(i32 %x) {
+; RV32I-LABEL: lshr_by_complemented_32:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    not a1, a0
+; RV32I-NEXT:    srl a0, a0, a1
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: lshr_by_complemented_32:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    srlw a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i32 31, %x
+  %b = lshr i32 %x, %a
+  ret i32 %b
+}
+
+define i64 @lshr_by_complemented_64(i64 %x) {
+; RV32I-LABEL: lshr_by_complemented_64:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    not a4, a0
+; RV32I-NEXT:    li a3, 31
+; RV32I-NEXT:    sub a3, a3, a0
+; RV32I-NEXT:    srl a2, a1, a4
+; RV32I-NEXT:    bltz a3, .LBB3_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    mv a0, a2
+; RV32I-NEXT:    j .LBB3_3
+; RV32I-NEXT:  .LBB3_2:
+; RV32I-NEXT:    srl a4, a0, a4
+; RV32I-NEXT:    li a5, 63
+; RV32I-NEXT:    sub a5, a5, a0
+; RV32I-NEXT:    xori a0, a5, 31
+; RV32I-NEXT:    slli a1, a1, 1
+; RV32I-NEXT:    sll a0, a1, a0
+; RV32I-NEXT:    or a0, a4, a0
+; RV32I-NEXT:  .LBB3_3:
+; RV32I-NEXT:    slti a1, a3, 0
+; RV32I-NEXT:    neg a1, a1
+; RV32I-NEXT:    and a1, a1, a2
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: lshr_by_complemented_64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    srl a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i64 63, %x
+  %b = lshr i64 %x, %a
+  ret i64 %b
+}
+
+define i32 @ashr_by_complemented_32(i32 %x) {
+; RV32I-LABEL: ashr_by_complemented_32:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    not a1, a0
+; RV32I-NEXT:    sra a0, a0, a1
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: ashr_by_complemented_32:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sraw a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i32 31, %x
+  %b = ashr i32 %x, %a
+  ret i32 %b
+}
+
+define i64 @ashr_by_complemented_64(i64 %x) {
+; RV32I-LABEL: ashr_by_complemented_64:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    mv a2, a0
+; RV32I-NEXT:    li a0, 31
+; RV32I-NEXT:    sub a4, a0, a2
+; RV32I-NEXT:    not a3, a2
+; RV32I-NEXT:    sra a0, a1, a3
+; RV32I-NEXT:    bltz a4, .LBB5_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    srai a1, a1, 31
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB5_2:
+; RV32I-NEXT:    srl a3, a2, a3
+; RV32I-NEXT:    li a4, 63
+; RV32I-NEXT:    sub a4, a4, a2
+; RV32I-NEXT:    xori a2, a4, 31
+; RV32I-NEXT:    slli a1, a1, 1
+; RV32I-NEXT:    sll a1, a1, a2
+; RV32I-NEXT:    or a3, a3, a1
+; RV32I-NEXT:    mv a1, a0
+; RV32I-NEXT:    mv a0, a3
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: ashr_by_complemented_64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sra a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i64 63, %x
+  %b = ashr i64 %x, %a
+  ret i64 %b
+}
+
+define i32 @shl_by_masked_complemented_32(i32 %x) {
+; RV32I-LABEL: shl_by_masked_complemented_32:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    not a1, a0
+; RV32I-NEXT:    sll a0, a0, a1
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl_by_masked_complemented_32:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sllw a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i32 31, %x
+  %b = and i32 %a, 31
+  %c = shl i32 %x, %b
+  ret i32 %c
+}
+
+define i64 @shl_by_masked_complemented_64(i64 %x) {
+; RV32I-LABEL: shl_by_masked_complemented_64:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    li a2, 63
+; RV32I-NEXT:    sub a2, a2, a0
+; RV32I-NEXT:    andi a4, a2, 63
+; RV32I-NEXT:    addi a2, a4, -32
+; RV32I-NEXT:    not a3, a0
+; RV32I-NEXT:    bltz a2, .LBB7_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    sll a1, a0, a4
+; RV32I-NEXT:    j .LBB7_3
+; RV32I-NEXT:  .LBB7_2:
+; RV32I-NEXT:    sll a1, a1, a3
+; RV32I-NEXT:    xori a4, a4, 31
+; RV32I-NEXT:    srli a5, a0, 1
+; RV32I-NEXT:    srl a4, a5, a4
+; RV32I-NEXT:    or a1, a1, a4
+; RV32I-NEXT:  .LBB7_3:
+; RV32I-NEXT:    sll a0, a0, a3
+; RV32I-NEXT:    slti a2, a2, 0
+; RV32I-NEXT:    neg a2, a2
+; RV32I-NEXT:    and a0, a2, a0
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: shl_by_masked_complemented_64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sll a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i64 63, %x
+  %b = and i64 %a, 63
+  %c = shl i64 %x, %b
+  ret i64 %c
+}
+
+define i64 @lshr_by_masked_complemented_64(i64 %x) {
+; RV32I-LABEL: lshr_by_masked_complemented_64:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    li a2, 63
+; RV32I-NEXT:    sub a2, a2, a0
+; RV32I-NEXT:    andi a4, a2, 63
+; RV32I-NEXT:    addi a2, a4, -32
+; RV32I-NEXT:    not a3, a0
+; RV32I-NEXT:    bltz a2, .LBB8_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    srl a0, a1, a4
+; RV32I-NEXT:    j .LBB8_3
+; RV32I-NEXT:  .LBB8_2:
+; RV32I-NEXT:    srl a0, a0, a3
+; RV32I-NEXT:    xori a4, a4, 31
+; RV32I-NEXT:    slli a5, a1, 1
+; RV32I-NEXT:    sll a4, a5, a4
+; RV32I-NEXT:    or a0, a0, a4
+; RV32I-NEXT:  .LBB8_3:
+; RV32I-NEXT:    srl a1, a1, a3
+; RV32I-NEXT:    slti a2, a2, 0
+; RV32I-NEXT:    neg a2, a2
+; RV32I-NEXT:    and a1, a2, a1
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: lshr_by_masked_complemented_64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    srl a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i64 63, %x
+  %b = and i64 %a, 63
+  %c = lshr i64 %x, %b
+  ret i64 %c
+}
+
+define i64 @ashr_by_masked_complemented_64(i64 %x) {
+; RV32I-LABEL: ashr_by_masked_complemented_64:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    li a2, 63
+; RV32I-NEXT:    sub a2, a2, a0
+; RV32I-NEXT:    andi a3, a2, 63
+; RV32I-NEXT:    addi a4, a3, -32
+; RV32I-NEXT:    mv a2, a1
+; RV32I-NEXT:    bltz a4, .LBB9_2
+; RV32I-NEXT:  # %bb.1:
+; RV32I-NEXT:    srai a1, a2, 31
+; RV32I-NEXT:    sra a0, a2, a3
+; RV32I-NEXT:    ret
+; RV32I-NEXT:  .LBB9_2:
+; RV32I-NEXT:    not a4, a0
+; RV32I-NEXT:    sra a1, a2, a4
+; RV32I-NEXT:    srl a0, a0, a4
+; RV32I-NEXT:    xori a3, a3, 31
+; RV32I-NEXT:    slli a2, a2, 1
+; RV32I-NEXT:    sll a2, a2, a3
+; RV32I-NEXT:    or a0, a0, a2
+; RV32I-NEXT:    ret
+;
+; RV64I-LABEL: ashr_by_masked_complemented_64:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    not a1, a0
+; RV64I-NEXT:    sra a0, a0, a1
+; RV64I-NEXT:    ret
+  %a = sub i64 63, %x
+  %b = and i64 %a, 63
+  %c = ashr i64 %x, %b
+  ret i64 %c
+}
-- 
2.7.4