[RISCV] Use shiftMaskXLen for one of the BCLR patterns. (llvm#164206)

This allows us to remove AND from the shift amount when DAG combine
has replaced (not (shl 1, X)) with (rotl -2, X). SimplifyDemandedBits
will often simplify the rotl case on its own, but not if the masked
shift amount has multiple users.

Author: topperc

llvm/lib/Target/RISCV/RISCVInstrInfoZb.td

@@ -527,8 +527,8 @@ def : Pat<(riscv_rolw GPR:$rs1, uimm5:$rs2),
 let Predicates = [HasStdExtZbs] in {
 def : Pat<(XLenVT (and (not (shl 1, shiftMaskXLen:$rs2)), GPR:$rs1)),
           (BCLR GPR:$rs1, shiftMaskXLen:$rs2)>;
-def : Pat<(XLenVT (and (rotl -2, (XLenVT GPR:$rs2)), GPR:$rs1)),
-          (BCLR GPR:$rs1, GPR:$rs2)>;
+def : Pat<(XLenVT (and (rotl -2, shiftMaskXLen:$rs2), GPR:$rs1)),
+          (BCLR GPR:$rs1, shiftMaskXLen:$rs2)>;
 def : Pat<(XLenVT (or (shl 1, shiftMaskXLen:$rs2), GPR:$rs1)),
           (BSET GPR:$rs1, shiftMaskXLen:$rs2)>;
 def : Pat<(XLenVT (xor (shl 1, shiftMaskXLen:$rs2), GPR:$rs1)),

llvm/test/CodeGen/RISCV/rv32zbs.ll

@@ -45,6 +45,32 @@ define i32 @bclr_i32_no_mask(i32 %a, i32 %b) nounwind {
   ret i32 %and1
 }
 
+define i32 @bclr_i32_mask_multiple(i32 %a, i32 %b, i32 %shamt) nounwind {
+; RV32I-LABEL: bclr_i32_mask_multiple:
+; RV32I:       # %bb.0:
+; RV32I-NEXT:    li a3, 1
+; RV32I-NEXT:    sll a2, a3, a2
+; RV32I-NEXT:    not a3, a2
+; RV32I-NEXT:    and a0, a3, a0
+; RV32I-NEXT:    or a1, a1, a2
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    ret
+;
+; RV32ZBS-LABEL: bclr_i32_mask_multiple:
+; RV32ZBS:       # %bb.0:
+; RV32ZBS-NEXT:    bclr a0, a0, a2
+; RV32ZBS-NEXT:    bset a1, a1, a2
+; RV32ZBS-NEXT:    add a0, a0, a1
+; RV32ZBS-NEXT:    ret
+  %shamt_masked = and i32 %shamt, 63
+  %shl = shl nuw i32 1, %shamt_masked
+  %neg = xor i32 %shl, -1
+  %and = and i32 %neg, %a
+  %or = or i32 %b, %shl
+  %c = add i32 %and, %or
+  ret i32 %c
+}
+
 define i64 @bclr_i64(i64 %a, i64 %b) nounwind {
 ; RV32I-LABEL: bclr_i64:
 ; RV32I:       # %bb.0:
@@ -301,17 +327,17 @@ define i64 @bext_i64(i64 %a, i64 %b) nounwind {
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    andi a3, a2, 63
 ; CHECK-NEXT:    addi a4, a3, -32
-; CHECK-NEXT:    bltz a4, .LBB12_2
+; CHECK-NEXT:    bltz a4, .LBB13_2
 ; CHECK-NEXT:  # %bb.1:
 ; CHECK-NEXT:    srl a0, a1, a3
-; CHECK-NEXT:    j .LBB12_3
-; CHECK-NEXT:  .LBB12_2:
+; CHECK-NEXT:    j .LBB13_3
+; CHECK-NEXT:  .LBB13_2:
 ; CHECK-NEXT:    srl a0, a0, a2
 ; CHECK-NEXT:    slli a1, a1, 1
 ; CHECK-NEXT:    not a2, a3
 ; CHECK-NEXT:    sll a1, a1, a2
 ; CHECK-NEXT:    or a0, a0, a1
-; CHECK-NEXT:  .LBB12_3:
+; CHECK-NEXT:  .LBB13_3:
 ; CHECK-NEXT:    andi a0, a0, 1
 ; CHECK-NEXT:    li a1, 0
 ; CHECK-NEXT:    ret
@@ -789,17 +815,17 @@ define i64 @bset_trailing_ones_i64_mask(i64 %a) nounwind {
 ; CHECK-NEXT:    li a3, -1
 ; CHECK-NEXT:    addi a1, a2, -32
 ; CHECK-NEXT:    sll a0, a3, a0
-; CHECK-NEXT:    bltz a1, .LBB43_2
+; CHECK-NEXT:    bltz a1, .LBB44_2
 ; CHECK-NEXT:  # %bb.1:
 ; CHECK-NEXT:    sll a2, a3, a2
-; CHECK-NEXT:    j .LBB43_3
-; CHECK-NEXT:  .LBB43_2:
+; CHECK-NEXT:    j .LBB44_3
+; CHECK-NEXT:  .LBB44_2:
 ; CHECK-NEXT:    not a2, a2
 ; 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:  .LBB44_3:
 ; CHECK-NEXT:    srai a1, a1, 31
 ; CHECK-NEXT:    and a0, a1, a0
 ; CHECK-NEXT:    not a1, a2
@@ -817,17 +843,17 @@ define i64 @bset_trailing_ones_i64_no_mask(i64 %a) nounwind {
 ; 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:    bltz a2, .LBB45_2
 ; CHECK-NEXT:  # %bb.1:
 ; CHECK-NEXT:    mv a0, a1
-; CHECK-NEXT:    j .LBB44_3
-; CHECK-NEXT:  .LBB44_2:
+; CHECK-NEXT:    j .LBB45_3
+; CHECK-NEXT:  .LBB45_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:  .LBB45_3:
 ; CHECK-NEXT:    srai a2, a2, 31
 ; CHECK-NEXT:    and a2, a2, a1
 ; CHECK-NEXT:    not a1, a0

llvm/test/CodeGen/RISCV/rv64zbs.ll

@@ -110,6 +110,32 @@ define i64 @bclr_i64_no_mask(i64 %a, i64 %b) nounwind {
   ret i64 %and1
 }
 
+define i64 @bclr_i64_mask_multiple(i64 %a, i64 %b, i64 %shamt) nounwind {
+; RV64I-LABEL: bclr_i64_mask_multiple:
+; RV64I:       # %bb.0:
+; RV64I-NEXT:    li a3, 1
+; RV64I-NEXT:    sll a2, a3, a2
+; RV64I-NEXT:    not a3, a2
+; RV64I-NEXT:    and a0, a3, a0
+; RV64I-NEXT:    or a1, a1, a2
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    ret
+;
+; RV64ZBS-LABEL: bclr_i64_mask_multiple:
+; RV64ZBS:       # %bb.0:
+; RV64ZBS-NEXT:    bclr a0, a0, a2
+; RV64ZBS-NEXT:    bset a1, a1, a2
+; RV64ZBS-NEXT:    add a0, a0, a1
+; RV64ZBS-NEXT:    ret
+  %shamt_masked = and i64 %shamt, 63
+  %shl = shl nuw i64 1, %shamt_masked
+  %neg = xor i64 %shl, -1
+  %and = and i64 %neg, %a
+  %or = or i64 %b, %shl
+  %c = add i64 %and, %or
+  ret i64 %c
+}
+
 define signext i32 @bset_i32(i32 signext %a, i32 signext %b) nounwind {
 ; RV64I-LABEL: bset_i32:
 ; RV64I:       # %bb.0:
@@ -372,19 +398,19 @@ define void @bext_i32_trunc(i32 signext %0, i32 signext %1) {
 ; RV64I:       # %bb.0:
 ; RV64I-NEXT:    srlw a0, a0, a1
 ; RV64I-NEXT:    andi a0, a0, 1
-; RV64I-NEXT:    beqz a0, .LBB19_2
+; RV64I-NEXT:    beqz a0, .LBB20_2
 ; RV64I-NEXT:  # %bb.1:
 ; RV64I-NEXT:    ret
-; RV64I-NEXT:  .LBB19_2:
+; RV64I-NEXT:  .LBB20_2:
 ; RV64I-NEXT:    tail bar
 ;
 ; RV64ZBS-LABEL: bext_i32_trunc:
 ; RV64ZBS:       # %bb.0:
 ; RV64ZBS-NEXT:    bext a0, a0, a1
-; RV64ZBS-NEXT:    beqz a0, .LBB19_2
+; RV64ZBS-NEXT:    beqz a0, .LBB20_2
 ; RV64ZBS-NEXT:  # %bb.1:
 ; RV64ZBS-NEXT:    ret
-; RV64ZBS-NEXT:  .LBB19_2:
+; RV64ZBS-NEXT:  .LBB20_2:
 ; RV64ZBS-NEXT:    tail bar
   %3 = shl i32 1, %1
   %4 = and i32 %3, %0