[NVPTX] expand mul_lohi to mul_lo and mul_hi

[jingyue]

Summary

Fixes PR26186.

[jholewinski]

LGTM! Thanks for fixing this; I just saw the PR go by. :)

lib/Target/NVPTX/NVPTXISelLowering.cpp

NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
   setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
   setOperationAction(ISD::CTPOP, MVT::i16, Legal);
   setOperationAction(ISD::CTPOP, MVT::i32, Legal);
   setOperationAction(ISD::CTPOP, MVT::i64, Legal);
 
   // PTX does not directly support SELP of i1, so promote to i32 first
   setOperationAction(ISD::SELECT, MVT::i1, Custom);
 
+  // PTX cannot multiply two i64s in a single instruction.
+  setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
+  setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
+
   // We have some custom DAG combine patterns for these nodes
   setTargetDAGCombine(ISD::ADD);
   setTargetDAGCombine(ISD::AND);
   setTargetDAGCombine(ISD::FADD);
   setTargetDAGCombine(ISD::MUL);
   setTargetDAGCombine(ISD::SHL);
   setTargetDAGCombine(ISD::SELECT);
 

test/CodeGen/NVPTX/arithmetic-int.ll

 
 define i64 @mul_i64(i64 %a, i64 %b) {
 ; CHECK: mul.lo.s64 %rd{{[0-9]+}}, %rd{{[0-9]+}}, %rd{{[0-9]+}}
 ; CHECK: ret
   %ret = mul i64 %a, %b
   ret i64 %ret
 }
 
+define i64 @umul_lohi_i64(i64 %a) {
+; CHECK-LABEL: umul_lohi_i64(
+entry:
+  %0 = zext i64 %a to i128
+  %1 = mul i128 %0, 288
+; CHECK: mul.lo.{{u|s}}64
+; CHECK: mul.hi.{{u|s}}64
+  %2 = lshr i128 %1, 1
+  %3 = trunc i128 %2 to i64
+  ret i64 %3
+}
+
+define i64 @smul_lohi_i64(i64 %a) {
+; CHECK-LABEL: smul_lohi_i64(
+entry:
+  %0 = sext i64 %a to i128
+  %1 = mul i128 %0, 288
+; CHECK: mul.lo.{{u|s}}64
+; CHECK: mul.hi.{{u|s}}64
+  %2 = ashr i128 %1, 1
+  %3 = trunc i128 %2 to i64
+  ret i64 %3
+}
+
 define i64 @sdiv_i64(i64 %a, i64 %b) {
 ; CHECK: div.s64 %rd{{[0-9]+}}, %rd{{[0-9]+}}, %rd{{[0-9]+}}
 ; CHECK: ret
   %ret = sdiv i64 %a, %b
   ret i64 %ret
 }
 
 define i64 @udiv_i64(i64 %a, i64 %b) {