[flang] 1.5x performance regression on gas_dyn2

[vzakhari]

This reproduces on zen4 after #121544

After InlineHLFIRAssign we now have this in chozdt_ routine:

  %23 = hlfir.minloc %19#0 {fastmath = #arith.fastmath<fast>} : (!fir.box<!fir.array<?xf32>>) -> !hlfir.expr<1xi32>
  fir.do_loop %arg6 = %c1 to %c1 step %c1 unordered {
    %55 = hlfir.apply %23, %arg6 : (!hlfir.expr<1xi32>, index) -> i32
    %56 = hlfir.designate %7#0 (%arg6)  : (!fir.ref<!fir.array<1xi32>>, index) -> !fir.ref<i32>
    hlfir.assign %55 to %56 : i32, !fir.ref<i32>
  }

The do-loop is the result of inlining of:

  hlfir.assign %23 to %7#0 : !hlfir.expr<1xi32>, !fir.ref<!fir.array<1xi32>>

After LLVM inlining, and other optimizations we have the following minloc loop:

.lr.ph.i:                                         ; preds = %.lr.ph.i.preheader, %.lr.ph.i
  %42 = phi i32 [ %51, %.lr.ph.i ], [ 1, %.lr.ph.i.preheader ]
  %43 = phi float [ %52, %.lr.ph.i ], [ %41, %.lr.ph.i.preheader ]
  %44 = phi i64 [ %53, %.lr.ph.i ], [ 1, %.lr.ph.i.preheader ]
  %45 = shl nsw i64 %44, 2
  %46 = getelementptr i8, ptr %10, i64 %45
  %47 = load float, ptr %46, align 4, !tbaa !60
  %48 = fcmp fast uge float %47, %43
  %49 = trunc i64 %44 to i32
  %50 = add nuw i32 %49, 1
  %51 = select i1 %48, i32 %42, i32 %50
  %52 = select i1 %48, float %43, float %47
  %53 = add nuw nsw i64 %44, 1
  %exitcond.not.i = icmp eq i64 %53, %9
  br i1 %exitcond.not.i, label %_FortranAMinlocReal4x1_i32_fast_simplified.exit, label %.lr.ph.i, !llvm.loop !64

The select operations form the cyclic dependency, and they are later transformed to cmov/minss instructions. Before my change, the SimplifyCFG pass could not produce selects, presumably, because the result of the integer select was stored into the temporary <1 x i32> array.

The cyclic dependency introduces by the integer cmov seems to limit performance of the loop. The loop would be better off with compare and jump.

Performance restores with -disable-select-optimize=false, unfortunately, it is disabled by default on X86.

Possible solutions:

• Try to enable the select optimization for X86, but I guess it is disabled for a reason.
• We can try to trick SimplifyCFG to not optimize the compare-jump into selects by setting the following probability to the jump instruction:

  %35 = load float, ptr %34, align 4, !tbaa !1609
  %36 = fcmp fast olt float %35, %23
  %37 = fcmp fast une float %23, %23
  %38 = fcmp fast oeq float %35, %35
  %39 = and i1 %37, %38
  %40 = or i1 %36, %39
  %41 = trunc i32 %27 to i1
  %42 = xor i1 %41, true
  %43 = or i1 %40, %42
  br i1 %43, label %44, label %47, !prof !2000

44:                                               ; preds = %26
  store i32 1, ptr %5, align 4, !tbaa !1609
  %45 = trunc i64 %22 to i32
  %46 = add i32 %45, 1
  store i32 %46, ptr %14, align 4, !tbaa !1609
  br label %47

47:                                               ; preds = %44, %26
...

!2000 = !{!"branch_weights", !"expected",  i32 1, i32 99}

This is just a trick though, and the right solution should be allowing the select optimization to use its heuristics.