[LoongArch] Add codegen support for multiplication operations

[SixWeining]

Summary

Reference:
https://llvm.org/docs/LangRef.html#mul-instruction

[SixWeining]

Use ";;" prefix for non-RUN and non-CHECK comment line.

[xen0n]

LGTM for generated LA64 assembly; I can't verify the LA32 assembly with gcc output due to gcc not currently supporting LoongArch ILP32 ABIs at all, but they look reasonable at a first glance. They should be results of generic codegen thus probably correct.

I'd like someone else to review the substantive changes (which should be trivial enough).

[MaskRay]

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

2

Use llvm/utils/update_llc_test_checks.py if this checks so many lines.

[SixWeining]

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

2

Thanks. Actually these check lines are generated by that python script with the first “Note:” line deleted manually. This is because LoongArch support for this script hasn’t been upstreamed since the lacking of function calls to write full test for this tool.

[SixWeining]

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

2

Just now we request review for this tool in D128433. After that been merged we can directly use it.

[SixWeining]

Address @MaskRay's comment. Rebase to D128433 since we need to call update_llc_test_checks.py.

[SixWeining]

@MaskRay I have rebased to D128433 which add LoongArch support to update_llc_test_checks. Do you have any futher comments? Thanks.

[SixWeining]

rebase

[SixWeining]

Rebase and merged first since this change has been approved previously. If @MaskRay and others have futher comments, we can revert or submit new revision. Thanks.

llvm/lib/Target/LoongArch/LoongArchISelLowering.cpp

if (Subtarget.hasBasicD()) {
     setCondCodeAction(FPCCToExpand, MVT::f64, Expand);
     setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
     setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
   }
 
   setOperationAction(ISD::BR_CC, GRLenVT, Expand);
   setOperationAction(ISD::SELECT_CC, GRLenVT, Expand);
   setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
+  setOperationAction({ISD::SMUL_LOHI, ISD::UMUL_LOHI}, GRLenVT, Expand);
+
+  if (!Subtarget.is64Bit())
+    setLibcallName(RTLIB::MUL_I128, nullptr);
 
   // Compute derived properties from the register classes.
   computeRegisterProperties(STI.getRegisterInfo());
 
   setStackPointerRegisterToSaveRestore(LoongArch::R3);
 
   setBooleanContents(ZeroOrOneBooleanContent);
 

llvm/lib/Target/LoongArch/LoongArchInstrInfo.td

 let Predicates = [IsLA32] in {
 def : PatGprGpr<add, ADD_W>;
 def : PatGprImm<add, ADDI_W, simm12>;
 def : PatGprGpr<sub, SUB_W>;
 def : PatGprGpr<sdiv, DIV_W>;
 def : PatGprGpr<udiv, DIV_WU>;
 def : PatGprGpr<srem, MOD_W>;
 def : PatGprGpr<urem, MOD_WU>;
+def : PatGprGpr<mul, MUL_W>;
+def : PatGprGpr<mulhs, MULH_W>;
+def : PatGprGpr<mulhu, MULH_WU>;
 } // Predicates = [IsLA32]
 
 let Predicates = [IsLA64] in {
 def : PatGprGpr<add, ADD_D>;
 def : PatGprGpr_32<add, ADD_W>;
 def : PatGprImm<add, ADDI_D, simm12>;
 def : PatGprImm_32<add, ADDI_W, simm12>;
 def : PatGprGpr<sub, SUB_D>;
 def : PatGprGpr_32<sub, SUB_W>;
 def : PatGprGpr<sdiv, DIV_D>;
 def : PatGprGpr<udiv, DIV_DU>;
 def : PatGprGpr<srem, MOD_D>;
 def : PatGprGpr<urem, MOD_DU>;
+// TODO: Select "_W[U]" instructions for i32xi32 if only lower 32 bits of the
+// product are used.
+def : PatGprGpr<mul, MUL_D>;
+def : PatGprGpr<mulhs, MULH_D>;
+def : PatGprGpr<mulhu, MULH_DU>;
+// Select MULW_D_W for calculating the full 64 bits product of i32xi32 signed
+// multiplication.
+def : Pat<(i64 (mul (sext_inreg GPR:$rj, i32), (sext_inreg GPR:$rk, i32))),
+          (MULW_D_W GPR:$rj, GPR:$rk)>;
+// Select MULW_D_WU for calculating the full 64 bits product of i32xi32
+// unsigned multiplication.
+def : Pat<(i64 (mul (loongarch_bstrpick GPR:$rj, (i64 31), (i64 0)),
+                    (loongarch_bstrpick GPR:$rk, (i64 31), (i64 0)))),
+          (MULW_D_WU GPR:$rj, GPR:$rk)>;
 } // Predicates = [IsLA64]
 
 def : PatGprGpr<and, AND>;
 def : PatGprImm<and, ANDI, uimm12>;
 def : PatGprGpr<or, OR>;
 def : PatGprImm<or, ORI, uimm12>;
 def : PatGprGpr<xor, XOR>;
 def : PatGprImm<xor, XORI, uimm12>;

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

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc --mtriple=loongarch32 < %s | FileCheck %s --check-prefix=LA32
+; RUN: llc --mtriple=loongarch64 < %s | FileCheck %s --check-prefix=LA64
+
+;; Exercise the 'mul' LLVM IR: https://llvm.org/docs/LangRef.html#mul-instruction
+
+define i1 @mul_i1(i1 %a, i1 %b) {
+; LA32-LABEL: mul_i1:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_i1:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+entry:
+  %r = mul i1 %a, %b
+  ret i1 %r
+}
+
+define i8 @mul_i8(i8 %a, i8 %b) {
+; LA32-LABEL: mul_i8:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_i8:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+entry:
+  %r = mul i8 %a, %b
+  ret i8 %r
+}
+
+define i16 @mul_i16(i16 %a, i16 %b) {
+; LA32-LABEL: mul_i16:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_i16:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+entry:
+  %r = mul i16 %a, %b
+  ret i16 %r
+}
+
+define i32 @mul_i32(i32 %a, i32 %b) {
+; LA32-LABEL: mul_i32:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    mul.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_i32:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+entry:
+  %r = mul i32 %a, %b
+  ret i32 %r
+}
+
+define i64 @mul_i64(i64 %a, i64 %b) {
+; LA32-LABEL: mul_i64:
+; LA32:       # %bb.0: # %entry
+; LA32-NEXT:    mul.w $a3, $a0, $a3
+; LA32-NEXT:    mulh.wu $a4, $a0, $a2
+; LA32-NEXT:    add.w $a3, $a4, $a3
+; LA32-NEXT:    mul.w $a1, $a1, $a2
+; LA32-NEXT:    add.w $a1, $a3, $a1
+; LA32-NEXT:    mul.w $a0, $a0, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_i64:
+; LA64:       # %bb.0: # %entry
+; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+entry:
+  %r = mul i64 %a, %b
+  ret i64 %r
+}
+
+define i64 @mul_pow2(i64 %a) {
+; LA32-LABEL: mul_pow2:
+; LA32:       # %bb.0:
+; LA32-NEXT:    slli.w $a1, $a1, 3
+; LA32-NEXT:    srli.w $a2, $a0, 29
+; LA32-NEXT:    or $a1, $a1, $a2
+; LA32-NEXT:    slli.w $a0, $a0, 3
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_pow2:
+; LA64:       # %bb.0:
+; LA64-NEXT:    slli.d $a0, $a0, 3
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = mul i64 %a, 8
+  ret i64 %1
+}
+
+define i64 @mul_p5(i64 %a) {
+; LA32-LABEL: mul_p5:
+; LA32:       # %bb.0:
+; LA32-NEXT:    ori $a2, $zero, 5
+; 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:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mul_p5:
+; LA64:       # %bb.0:
+; LA64-NEXT:    ori $a1, $zero, 5
+; LA64-NEXT:    mul.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = mul i64 %a, 5
+  ret i64 %1
+}
+
+define i32 @mulh_w(i32 %a, i32 %b) {
+; LA32-LABEL: mulh_w:
+; LA32:       # %bb.0:
+; LA32-NEXT:    mulh.w $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mulh_w:
+; LA64:       # %bb.0:
+; LA64-NEXT:    mulw.d.w $a0, $a0, $a1
+; LA64-NEXT:    srli.d $a0, $a0, 32
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = sext i32 %a to i64
+  %2 = sext i32 %b to i64
+  %3 = mul i64 %1, %2
+  %4 = lshr i64 %3, 32
+  %5 = trunc i64 %4 to i32
+  ret i32 %5
+}
+
+define i32 @mulh_wu(i32 %a, i32 %b) {
+; LA32-LABEL: mulh_wu:
+; LA32:       # %bb.0:
+; LA32-NEXT:    mulh.wu $a0, $a0, $a1
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mulh_wu:
+; LA64:       # %bb.0:
+; LA64-NEXT:    mulw.d.wu $a0, $a0, $a1
+; LA64-NEXT:    srli.d $a0, $a0, 32
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = zext i32 %a to i64
+  %2 = zext i32 %b to i64
+  %3 = mul i64 %1, %2
+  %4 = lshr i64 %3, 32
+  %5 = trunc i64 %4 to i32
+  ret i32 %5
+}
+
+define i64 @mulh_d(i64 %a, i64 %b) {
+; LA32-LABEL: mulh_d:
+; LA32:       # %bb.0:
+; LA32-NEXT:    mulh.wu $a4, $a0, $a2
+; LA32-NEXT:    mul.w $a5, $a1, $a2
+; LA32-NEXT:    add.w $a4, $a5, $a4
+; LA32-NEXT:    sltu $a5, $a4, $a5
+; LA32-NEXT:    mulh.wu $a6, $a1, $a2
+; LA32-NEXT:    add.w $a5, $a6, $a5
+; LA32-NEXT:    mul.w $a6, $a0, $a3
+; LA32-NEXT:    add.w $a4, $a6, $a4
+; LA32-NEXT:    sltu $a4, $a4, $a6
+; LA32-NEXT:    mulh.wu $a6, $a0, $a3
+; LA32-NEXT:    add.w $a4, $a6, $a4
+; LA32-NEXT:    add.w $a4, $a5, $a4
+; LA32-NEXT:    sltu $a5, $a4, $a5
+; LA32-NEXT:    mulh.wu $a6, $a1, $a3
+; LA32-NEXT:    add.w $a5, $a6, $a5
+; LA32-NEXT:    mul.w $a6, $a1, $a3
+; LA32-NEXT:    add.w $a4, $a6, $a4
+; LA32-NEXT:    sltu $a6, $a4, $a6
+; LA32-NEXT:    add.w $a5, $a5, $a6
+; LA32-NEXT:    srai.w $a6, $a1, 31
+; LA32-NEXT:    mul.w $a7, $a2, $a6
+; LA32-NEXT:    mulh.wu $a2, $a2, $a6
+; LA32-NEXT:    add.w $a2, $a2, $a7
+; LA32-NEXT:    mul.w $a6, $a3, $a6
+; LA32-NEXT:    add.w $a2, $a2, $a6
+; LA32-NEXT:    srai.w $a3, $a3, 31
+; LA32-NEXT:    mul.w $a1, $a3, $a1
+; LA32-NEXT:    mulh.wu $a6, $a3, $a0
+; LA32-NEXT:    add.w $a1, $a6, $a1
+; LA32-NEXT:    mul.w $a0, $a3, $a0
+; LA32-NEXT:    add.w $a1, $a1, $a0
+; LA32-NEXT:    add.w $a1, $a1, $a2
+; LA32-NEXT:    add.w $a2, $a0, $a7
+; LA32-NEXT:    sltu $a0, $a2, $a0
+; LA32-NEXT:    add.w $a0, $a1, $a0
+; LA32-NEXT:    add.w $a1, $a5, $a0
+; LA32-NEXT:    add.w $a0, $a4, $a2
+; LA32-NEXT:    sltu $a2, $a0, $a4
+; LA32-NEXT:    add.w $a1, $a1, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mulh_d:
+; LA64:       # %bb.0:
+; LA64-NEXT:    mulh.d $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = sext i64 %a to i128
+  %2 = sext i64 %b to i128
+  %3 = mul i128 %1, %2
+  %4 = lshr i128 %3, 64
+  %5 = trunc i128 %4 to i64
+  ret i64 %5
+}
+
+define i64 @mulh_du(i64 %a, i64 %b) {
+; LA32-LABEL: mulh_du:
+; LA32:       # %bb.0:
+; LA32-NEXT:    mulh.wu $a4, $a0, $a2
+; LA32-NEXT:    mul.w $a5, $a1, $a2
+; LA32-NEXT:    add.w $a4, $a5, $a4
+; LA32-NEXT:    sltu $a5, $a4, $a5
+; LA32-NEXT:    mulh.wu $a2, $a1, $a2
+; LA32-NEXT:    add.w $a2, $a2, $a5
+; LA32-NEXT:    mul.w $a5, $a0, $a3
+; LA32-NEXT:    add.w $a4, $a5, $a4
+; LA32-NEXT:    sltu $a4, $a4, $a5
+; LA32-NEXT:    mulh.wu $a0, $a0, $a3
+; LA32-NEXT:    add.w $a0, $a0, $a4
+; LA32-NEXT:    mul.w $a4, $a1, $a3
+; LA32-NEXT:    mulh.wu $a1, $a1, $a3
+; LA32-NEXT:    add.w $a0, $a2, $a0
+; LA32-NEXT:    sltu $a2, $a0, $a2
+; LA32-NEXT:    add.w $a1, $a1, $a2
+; LA32-NEXT:    add.w $a0, $a4, $a0
+; LA32-NEXT:    sltu $a2, $a0, $a4
+; LA32-NEXT:    add.w $a1, $a1, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mulh_du:
+; LA64:       # %bb.0:
+; LA64-NEXT:    mulh.du $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = zext i64 %a to i128
+  %2 = zext i64 %b to i128
+  %3 = mul i128 %1, %2
+  %4 = lshr i128 %3, 64
+  %5 = trunc i128 %4 to i64
+  ret i64 %5
+}
+
+define i64 @mulw_d_w(i32 %a, i32 %b) {
+; LA32-LABEL: mulw_d_w:
+; LA32:       # %bb.0:
+; LA32-NEXT:    mul.w $a2, $a0, $a1
+; LA32-NEXT:    mulh.w $a1, $a0, $a1
+; LA32-NEXT:    move $a0, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mulw_d_w:
+; LA64:       # %bb.0:
+; LA64-NEXT:    mulw.d.w $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = sext i32 %a to i64
+  %2 = sext i32 %b to i64
+  %3 = mul i64 %1, %2
+  ret i64 %3
+}
+
+define i64 @mulw_d_wu(i32 %a, i32 %b) {
+; LA32-LABEL: mulw_d_wu:
+; LA32:       # %bb.0:
+; LA32-NEXT:    mul.w $a2, $a0, $a1
+; LA32-NEXT:    mulh.wu $a1, $a0, $a1
+; LA32-NEXT:    move $a0, $a2
+; LA32-NEXT:    jirl $zero, $ra, 0
+;
+; LA64-LABEL: mulw_d_wu:
+; LA64:       # %bb.0:
+; LA64-NEXT:    mulw.d.wu $a0, $a0, $a1
+; LA64-NEXT:    jirl $zero, $ra, 0
+  %1 = zext i32 %a to i64
+  %2 = zext i32 %b to i64
+  %3 = mul i64 %1, %2
+  ret i64 %3
+}