[RISCV] Disable constant hoisting for multiply by negated power of 2.

[craig.topper]

Summary

A mul by a negated power of 2 is a slli followed by neg. This doesn't
require any constant materialization and may be lower latency than mul.
The neg may also be foldable into other arithmetic.

[craig.topper]

Ping

[reames]

LGTM

However, I'll note the code structure here is non-obvious. When I see code talking about immediate cost, I don't expect that to be highly specific to hoisting. I expect it to instead relate to materialization cost. And yet, that doesn't appear to be what the existing code in this method does.

llvm/lib/Target/RISCV/RISCVTargetTransformInfo.cpp

if (Imm == UINT64_C(0xffff) && ST->hasStdExtZbb())
       return TTI::TCC_Free;
     // zext.w
     if (Imm == UINT64_C(0xffffffff) && ST->hasStdExtZba())
       return TTI::TCC_Free;
     LLVM_FALLTHROUGH;
   case Instruction::Add:
   case Instruction::Or:
   case Instruction::Xor:
+    Takes12BitImm = true;
+    break;
   case Instruction::Mul:
+    // Negated power of 2 is a shift and a negate.
+    if (Imm.isNegatedPowerOf2())
+      return TTI::TCC_Free;
+    // FIXME: There is no MULI instruction.
     Takes12BitImm = true;
     break;
   case Instruction::Sub:
   case Instruction::Shl:
   case Instruction::LShr:
   case Instruction::AShr:
     Takes12BitImm = true;
     ImmArgIdx = 1;

llvm/test/Transforms/ConstantHoisting/RISCV/immediates.ll

 ; Check that we don't hoist zext.w with Zba.
 define i64 @test8(i64 %a) nounwind "target-features"="+zba" {
 ; CHECK-LABEL: test8
 ; CHECK: and i64 %a, 4294967295
   %1 = and i64 %a, 4294967295
   %2 = and i64 %1, 4294967295
   ret i64 %2
 }
+
+; Check that we don't hoist mul with negated power of 2.
+define i64 @test9(i64 %a) nounwind {
+; CHECK-LABEL: test9
+; CHECK: mul i64 %a, -4294967296
+  %1 = mul i64 %a, -4294967296
+  %2 = mul i64 %1, -4294967296
+  ret i64 %2
+}