[RISCV] Optimize multiplication with constant

[benshi001]

i64 mul on riscv32 is omitted, since 6-10 extra instruction are generated, not sure if it is a win against calling __muldi3.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

3829

What about 128-bit on riscv64?

3834

Can we just use countTrailingZeros here?

[benshi001]

This is the best form I can tune now, the best trade among int128, int64 and int32. It generated optimized code for most of them,

except some cases are left not covered,

1. rv32 with M extension: not sure if is a real win since some HW multiplier has quite low lantency, even to 2 cycles.
2. i128 mul on rv32: the generated code is too large, not sure if it is better than calling __multi3
3. -(1<<m)-(1<<n): more NEGS are generated than other (1<<m)±(1<<n), so it may not be optimal than quick HW multipliers.

Shall we commit current form first? Then I will go on with the hard cases in the future.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

3834

Is the first part of this just !Imm.isSignedIntN(12)? And 0 != (Imm & High) is !Imm.isIntN(11)?

[benshi001]

ping ...

[benshi001]

ping

[craig.topper]

llvm/test/CodeGen/RISCV/mul.ll

933

Please pre-commit these test cases and rebase this patch.

[benshi001]

llvm/test/CodeGen/RISCV/mul.ll

933

The test cases are commited to

https://reviews.llvm.org/D94062

[jrtc27]

llvm/test/CodeGen/RISCV/mul.ll

933

Please rebase and mark that as a parent of this revision.

[jrtc27]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

3821

Please refrain from using unicode

[craig.topper]

LGTM. Please write a description of the changes in the the commit message when you commit this.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

 bool RISCVTargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT,
                                                  SDValue C) const {
   // Check integral scalar types.
   if (VT.isScalarInteger()) {
     // Do not perform the transformation on riscv32 with the M extension.
     if (!Subtarget.is64Bit() && Subtarget.hasStdExtM())
       return false;
     if (auto *ConstNode = dyn_cast<ConstantSDNode>(C.getNode())) {
       if (ConstNode->getAPIntValue().getBitWidth() > 8 * sizeof(int64_t))
         return false;
       int64_t Imm = ConstNode->getSExtValue();
+      // Optimize to ±(1<<m)±1 if applicable.
       if (isPowerOf2_64(Imm + 1) || isPowerOf2_64(Imm - 1) ||
           isPowerOf2_64(1 - Imm) || isPowerOf2_64(-1 - Imm))
         return true;
+      // Omit 64-bit multiplication on riscv32.
+      if (!Subtarget.is64Bit() && VT.getSizeInBits() > 32)
+        return false;
+      // Optimize to ±(1<<m)±(1<<n) if Imm needs a pair of LUI/ADDI.
+      if ((Imm < -2048 || 2047 < Imm) && (Imm & 0x7ff) != 0) {
+        int64_t Tmp = Imm;
+        while (Tmp != 0 && (Tmp & 1) == 0)
+          Tmp >>= 1;
+        if (isPowerOf2_64(Tmp + 1) || isPowerOf2_64(Tmp - 1) ||
+            isPowerOf2_64(1 - Tmp) || isPowerOf2_64(-1 - Tmp))
+          return true;
+      }
     }
   }
 
   return false;
 }
 
 #define GET_REGISTER_MATCHER
 #include "RISCVGenAsmMatcher.inc"

llvm/test/CodeGen/RISCV/mul.ll

 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, 12288
   ret i32 %1
 }
 
 define i32 @muli32_p4352(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_p4352:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    lui a1, 1
-; RV32I-NEXT:    addi a1, a1, 256
-; RV32I-NEXT:    call __mulsi3@plt
-; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 8
+; RV32I-NEXT:    slli a0, a0, 12
+; RV32I-NEXT:    add a0, a0, a1
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_p4352:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    lui a1, 1
 ; RV32IM-NEXT:    addi a1, a1, 256
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_p4352:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1
-; RV64I-NEXT:    addiw a1, a1, 256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 8
+; RV64I-NEXT:    slli a0, a0, 12
+; RV64I-NEXT:    addw a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_p4352:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1
-; RV64IM-NEXT:    addiw a1, a1, 256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 8
+; RV64IM-NEXT:    slli a0, a0, 12
+; RV64IM-NEXT:    addw a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, 4352
   ret i32 %1
 }
 
 define i32 @muli32_p3840(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_p3840:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    lui a1, 1
-; RV32I-NEXT:    addi a1, a1, -256
-; RV32I-NEXT:    call __mulsi3@plt
-; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 8
+; RV32I-NEXT:    slli a0, a0, 12
+; RV32I-NEXT:    sub a0, a0, a1
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_p3840:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    lui a1, 1
 ; RV32IM-NEXT:    addi a1, a1, -256
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_p3840:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1
-; RV64I-NEXT:    addiw a1, a1, -256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 8
+; RV64I-NEXT:    slli a0, a0, 12
+; RV64I-NEXT:    subw a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_p3840:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1
-; RV64IM-NEXT:    addiw a1, a1, -256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 8
+; RV64IM-NEXT:    slli a0, a0, 12
+; RV64IM-NEXT:    subw a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, 3840
   ret i32 %1
 }
 
 define i32 @muli32_m3840(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_m3840:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    lui a1, 1048575
-; RV32I-NEXT:    addi a1, a1, 256
-; RV32I-NEXT:    call __mulsi3@plt
-; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 12
+; RV32I-NEXT:    slli a0, a0, 8
+; RV32I-NEXT:    sub a0, a0, a1
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_m3840:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    lui a1, 1048575
 ; RV32IM-NEXT:    addi a1, a1, 256
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_m3840:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1048575
-; RV64I-NEXT:    addiw a1, a1, 256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 12
+; RV64I-NEXT:    slli a0, a0, 8
+; RV64I-NEXT:    subw a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_m3840:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1048575
-; RV64IM-NEXT:    addiw a1, a1, 256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 12
+; RV64IM-NEXT:    slli a0, a0, 8
+; RV64IM-NEXT:    subw a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, -3840
   ret i32 %1
 }
 
 define i32 @muli32_m4352(i32 %a) nounwind {
 ; RV32I-LABEL: muli32_m4352:
 ; RV32I:       # %bb.0:
-; RV32I-NEXT:    addi sp, sp, -16
-; RV32I-NEXT:    sw ra, 12(sp) # 4-byte Folded Spill
-; RV32I-NEXT:    lui a1, 1048575
-; RV32I-NEXT:    addi a1, a1, -256
-; RV32I-NEXT:    call __mulsi3@plt
-; RV32I-NEXT:    lw ra, 12(sp) # 4-byte Folded Reload
-; RV32I-NEXT:    addi sp, sp, 16
+; RV32I-NEXT:    slli a1, a0, 8
+; RV32I-NEXT:    slli a0, a0, 12
+; RV32I-NEXT:    add a0, a0, a1
+; RV32I-NEXT:    neg a0, a0
 ; RV32I-NEXT:    ret
 ;
 ; RV32IM-LABEL: muli32_m4352:
 ; RV32IM:       # %bb.0:
 ; RV32IM-NEXT:    lui a1, 1048575
 ; RV32IM-NEXT:    addi a1, a1, -256
 ; RV32IM-NEXT:    mul a0, a0, a1
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli32_m4352:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1048575
-; RV64I-NEXT:    addiw a1, a1, -256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 8
+; RV64I-NEXT:    slli a0, a0, 12
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    negw a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli32_m4352:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1048575
-; RV64IM-NEXT:    addiw a1, a1, -256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 8
+; RV64IM-NEXT:    slli a0, a0, 12
+; RV64IM-NEXT:    add a0, a0, a1
+; RV64IM-NEXT:    negw a0, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i32 %a, -4352
   ret i32 %1
 }
 
 define i64 @muli64_p4352(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_p4352:
 ; RV32I:       # %bb.0:

 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_p4352:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1
-; RV64I-NEXT:    addiw a1, a1, 256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 8
+; RV64I-NEXT:    slli a0, a0, 12
+; RV64I-NEXT:    add a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_p4352:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1
-; RV64IM-NEXT:    addiw a1, a1, 256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 8
+; RV64IM-NEXT:    slli a0, a0, 12
+; RV64IM-NEXT:    add a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, 4352
   ret i64 %1
 }
 
 define i64 @muli64_p3840(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_p3840:
 ; RV32I:       # %bb.0:

 ; RV32IM-NEXT:    mul a1, a1, a2
 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_p3840:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1
-; RV64I-NEXT:    addiw a1, a1, -256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 8
+; RV64I-NEXT:    slli a0, a0, 12
+; RV64I-NEXT:    sub a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_p3840:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1
-; RV64IM-NEXT:    addiw a1, a1, -256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 8
+; RV64IM-NEXT:    slli a0, a0, 12
+; RV64IM-NEXT:    sub a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, 3840
   ret i64 %1
 }
 
 define i64 @muli64_m4352(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_m4352:
 ; RV32I:       # %bb.0:

 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    sub a3, a3, a0
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_m4352:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1048575
-; RV64I-NEXT:    addiw a1, a1, -256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 8
+; RV64I-NEXT:    slli a0, a0, 12
+; RV64I-NEXT:    add a0, a0, a1
+; RV64I-NEXT:    neg a0, a0
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_m4352:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1048575
-; RV64IM-NEXT:    addiw a1, a1, -256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 8
+; RV64IM-NEXT:    slli a0, a0, 12
+; RV64IM-NEXT:    add a0, a0, a1
+; RV64IM-NEXT:    neg a0, a0
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, -4352
   ret i64 %1
 }
 
 define i64 @muli64_m3840(i64 %a) nounwind {
 ; RV32I-LABEL: muli64_m3840:
 ; RV32I:       # %bb.0:

 ; RV32IM-NEXT:    mulhu a3, a0, a2
 ; RV32IM-NEXT:    sub a3, a3, a0
 ; RV32IM-NEXT:    add a1, a3, a1
 ; RV32IM-NEXT:    mul a0, a0, a2
 ; RV32IM-NEXT:    ret
 ;
 ; RV64I-LABEL: muli64_m3840:
 ; RV64I:       # %bb.0:
-; RV64I-NEXT:    addi sp, sp, -16
-; RV64I-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
-; RV64I-NEXT:    lui a1, 1048575
-; RV64I-NEXT:    addiw a1, a1, 256
-; RV64I-NEXT:    call __muldi3@plt
-; RV64I-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
-; RV64I-NEXT:    addi sp, sp, 16
+; RV64I-NEXT:    slli a1, a0, 12
+; RV64I-NEXT:    slli a0, a0, 8
+; RV64I-NEXT:    sub a0, a0, a1
 ; RV64I-NEXT:    ret
 ;
 ; RV64IM-LABEL: muli64_m3840:
 ; RV64IM:       # %bb.0:
-; RV64IM-NEXT:    lui a1, 1048575
-; RV64IM-NEXT:    addiw a1, a1, 256
-; RV64IM-NEXT:    mul a0, a0, a1
+; RV64IM-NEXT:    slli a1, a0, 12
+; RV64IM-NEXT:    slli a0, a0, 8
+; RV64IM-NEXT:    sub a0, a0, a1
 ; RV64IM-NEXT:    ret
   %1 = mul i64 %a, -3840
   ret i64 %1
 }