[LoongArch] Optimize multiplication with immediates

[benshi001]

Summary

Optimize (mul r, c) to (SLLI (ALSL r, r, i0), i1), in which
c = (1 + (1 << i0)) << i1.

[SixWeining]

About the commit message, do you mean:

Optimize (mul r, c) to (SLLI (ALSL r, r, i0), i1)

[benshi001]

In D147428#4239987, @SixWeining wrote:

About the commit message, do you mean:

Optimize (mul r, c) to (SLLI (ALSL r, r, i0), i1)

Thanks! Fixed.

[SixWeining]

LGTM except a nit. Thanks.

llvm/lib/Target/LoongArch/LoongArchInstrInfo.td

431

Nit: ALSl r, r, i0

[benshi001]

llvm/lib/Target/LoongArch/LoongArchInstrInfo.td

431

Thanks! I have fixed it before doing git push.

llvm/lib/Target/LoongArch/LoongArchInstrInfo.td

 // Return imm - (imm < 0 ? -2048 : 2047).
 def AddiPairImmSmall : SDNodeXForm<imm, [{
   int64_t Imm = N->getSExtValue();
   int64_t Adj = Imm < 0 ? -2048 : 2047;
   return CurDAG->getTargetConstant(Imm - Adj, SDLoc(N),
                                    N->getValueType(0));
 }]>;
 
+// Check if (mul r, imm) can be optimized to (SLLI (ALSL r, r, i0), i1),
+// in which imm = (1 + (1 << i0)) << i1.
+def AlslSlliImm : PatLeaf<(imm), [{
+  if (!N->hasOneUse())
+    return false;
+  uint64_t Imm = N->getZExtValue();
+  unsigned I1 = llvm::countr_zero(Imm);
+  uint64_t Rem = Imm >> I1;
+  return Rem == 3 || Rem == 5 || Rem == 9 || Rem == 17;
+}]>;
+
+def AlslSlliImmI1 : SDNodeXForm<imm, [{
+  uint64_t Imm = N->getZExtValue();
+  unsigned I1 = llvm::countr_zero(Imm);
+  return CurDAG->getTargetConstant(I1, SDLoc(N),
+                                   N->getValueType(0));
+}]>;
+
+def AlslSlliImmI0 : SDNodeXForm<imm, [{
+  uint64_t Imm = N->getZExtValue();
+  unsigned I1 = llvm::countr_zero(Imm);
+  uint64_t I0;
+  switch (Imm >> I1) {
+  case 3:  I0 = 1; break;
+  case 5:  I0 = 2; break;
+  case 9:  I0 = 3; break;
+  default: I0 = 4; break;
+  }
+  return CurDAG->getTargetConstant(I0, SDLoc(N),
+                                   N->getValueType(0));
+}]>;
+
 //===----------------------------------------------------------------------===//
 // Instruction Formats
 //===----------------------------------------------------------------------===//
 
 include "LoongArchInstrFormats.td"
 include "LoongArchFloatInstrFormats.td"
 
 //===----------------------------------------------------------------------===//
def : Pat<(sext_inreg (mul_const_oneuse GPR:$r, (i64 CImm)), i32),
                       (ALSL_W GPR:$r, GPR:$r, (i64 Idx0)), (i64 Idx1))>;
     def : Pat<(mul_const_oneuse GPR:$r, (i64 CImm)),
               (ALSL_D (ALSL_D GPR:$r, GPR:$r, (i64 Idx0)),
                       (ALSL_D GPR:$r, GPR:$r, (i64 Idx0)), (i64 Idx1))>;
   }
 }
 } // Predicates = [IsLA64]
 
+let Predicates = [IsLA32] in {
+def : Pat<(mul GPR:$rj, (AlslSlliImm:$im)),
+          (SLLI_W (ALSL_W GPR:$rj, GPR:$rj, (AlslSlliImmI0 AlslSlliImm:$im)),
+                  (AlslSlliImmI1 AlslSlliImm:$im))>;
+} // Predicates = [IsLA32]
+
+let Predicates = [IsLA64] in {
+def : Pat<(sext_inreg (mul GPR:$rj, (AlslSlliImm:$im)), i32),
+          (SLLI_W (ALSL_W GPR:$rj, GPR:$rj, (AlslSlliImmI0 AlslSlliImm:$im)),
+                  (AlslSlliImmI1 AlslSlliImm:$im))>;
+def : Pat<(mul GPR:$rj, (AlslSlliImm:$im)),
+          (SLLI_D (ALSL_D GPR:$rj, GPR:$rj, (AlslSlliImmI0 AlslSlliImm:$im)),
+                  (AlslSlliImmI1 AlslSlliImm:$im))>;
+} // Predicates = [IsLA64]
+
 foreach Idx = 1...7 in {
   defvar ShamtA = !mul(8, Idx);
   defvar ShamtB = !mul(8, !sub(8, Idx));
   def : Pat<(or (shl GPR:$rk, (i64 ShamtA)), (srl GPR:$rj, (i64 ShamtB))),
             (BYTEPICK_D GPR:$rj, GPR:$rk, Idx)>;
 }
 
 foreach Idx = 1...3 in {

llvm/test/CodeGen/LoongArch/ir-instruction/mul.ll

 ; LA64-NEXT:    ret
   %b = mul i64 %a, 4112
   ret i64 %b
 }
 
 define signext i32 @mul_i32_768(i32 %a) {
 ; LA32-LABEL: mul_i32_768:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    ori $a1, $zero, 768
-; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a0, $a0, 1
+; LA32-NEXT:    slli.w $a0, $a0, 8
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i32_768:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a1, $zero, 768
-; LA64-NEXT:    mul.d $a0, $a0, $a1
-; LA64-NEXT:    addi.w $a0, $a0, 0
+; LA64-NEXT:    alsl.w $a0, $a0, $a0, 1
+; LA64-NEXT:    slli.w $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i32 %a, 768
   ret i32 %b
 }
 
 define signext i32 @mul_i32_1280(i32 %a) {
 ; LA32-LABEL: mul_i32_1280:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    ori $a1, $zero, 1280
-; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a0, $a0, 2
+; LA32-NEXT:    slli.w $a0, $a0, 8
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i32_1280:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a1, $zero, 1280
-; LA64-NEXT:    mul.d $a0, $a0, $a1
-; LA64-NEXT:    addi.w $a0, $a0, 0
+; LA64-NEXT:    alsl.w $a0, $a0, $a0, 2
+; LA64-NEXT:    slli.w $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i32 %a, 1280
   ret i32 %b
 }
 
 define signext i32 @mul_i32_2304(i32 %a) {
 ; LA32-LABEL: mul_i32_2304:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    ori $a1, $zero, 2304
-; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a0, $a0, 3
+; LA32-NEXT:    slli.w $a0, $a0, 8
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i32_2304:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a1, $zero, 2304
-; LA64-NEXT:    mul.d $a0, $a0, $a1
-; LA64-NEXT:    addi.w $a0, $a0, 0
+; LA64-NEXT:    alsl.w $a0, $a0, $a0, 3
+; LA64-NEXT:    slli.w $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i32 %a, 2304
   ret i32 %b
 }
 
 define signext i32 @mul_i32_4352(i32 %a) {
 ; LA32-LABEL: mul_i32_4352:
 ; LA32:       # %bb.0:
-; LA32-NEXT:    lu12i.w $a1, 1
-; LA32-NEXT:    ori $a1, $a1, 256
-; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    alsl.w $a0, $a0, $a0, 4
+; LA32-NEXT:    slli.w $a0, $a0, 8
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i32_4352:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    lu12i.w $a1, 1
-; LA64-NEXT:    ori $a1, $a1, 256
-; LA64-NEXT:    mul.d $a0, $a0, $a1
-; LA64-NEXT:    addi.w $a0, $a0, 0
+; LA64-NEXT:    alsl.w $a0, $a0, $a0, 4
+; LA64-NEXT:    slli.w $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i32 %a, 4352
   ret i32 %b
 }
 
 define i64 @mul_i64_768(i64 %a) {
 ; LA32-LABEL: mul_i64_768:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    ori $a2, $zero, 768
 ; LA32-NEXT:    mul.w $a1, $a1, $a2
 ; LA32-NEXT:    mulh.wu $a3, $a0, $a2
 ; LA32-NEXT:    add.w $a1, $a3, $a1
 ; LA32-NEXT:    mul.w $a0, $a0, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i64_768:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a1, $zero, 768
-; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    alsl.d $a0, $a0, $a0, 1
+; LA64-NEXT:    slli.d $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i64 %a, 768
   ret i64 %b
 }
 
 define i64 @mul_i64_1280(i64 %a) {
 ; LA32-LABEL: mul_i64_1280:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    ori $a2, $zero, 1280
 ; LA32-NEXT:    mul.w $a1, $a1, $a2
 ; LA32-NEXT:    mulh.wu $a3, $a0, $a2
 ; LA32-NEXT:    add.w $a1, $a3, $a1
 ; LA32-NEXT:    mul.w $a0, $a0, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i64_1280:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a1, $zero, 1280
-; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    alsl.d $a0, $a0, $a0, 2
+; LA64-NEXT:    slli.d $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i64 %a, 1280
   ret i64 %b
 }
 
 define i64 @mul_i64_2304(i64 %a) {
 ; LA32-LABEL: mul_i64_2304:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    ori $a2, $zero, 2304
 ; LA32-NEXT:    mul.w $a1, $a1, $a2
 ; LA32-NEXT:    mulh.wu $a3, $a0, $a2
 ; LA32-NEXT:    add.w $a1, $a3, $a1
 ; LA32-NEXT:    mul.w $a0, $a0, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i64_2304:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    ori $a1, $zero, 2304
-; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    alsl.d $a0, $a0, $a0, 3
+; LA64-NEXT:    slli.d $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i64 %a, 2304
   ret i64 %b
 }
 
 define i64 @mul_i64_4352(i64 %a) {
 ; LA32-LABEL: mul_i64_4352:
 ; LA32:       # %bb.0:
 ; LA32-NEXT:    lu12i.w $a2, 1
 ; LA32-NEXT:    ori $a2, $a2, 256
 ; LA32-NEXT:    mul.w $a1, $a1, $a2
 ; LA32-NEXT:    mulh.wu $a3, $a0, $a2
 ; LA32-NEXT:    add.w $a1, $a3, $a1
 ; LA32-NEXT:    mul.w $a0, $a0, $a2
 ; LA32-NEXT:    ret
 ;
 ; LA64-LABEL: mul_i64_4352:
 ; LA64:       # %bb.0:
-; LA64-NEXT:    lu12i.w $a1, 1
-; LA64-NEXT:    ori $a1, $a1, 256
-; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    alsl.d $a0, $a0, $a0, 4
+; LA64-NEXT:    slli.d $a0, $a0, 8
 ; LA64-NEXT:    ret
   %b = mul i64 %a, 4352
   ret i64 %b
 }