}

	// Indicate that this span is cached and prevent asynchronous
	// sweeping in the next sweep phase.
	s.sweepgen = mheap_.sweepgen + 3

	// Store the current alloc count for accounting later.
	s.allocCountBeforeCache = s.allocCount

	// Update heapLive and flush scanAlloc.
	//
	// We have not yet allocated anything new into the span, but we

		target <- i

		// Check to make sure the memory limit is maintained.
		// We're just sampling here so if it transiently goes over we might miss it.
		// The internal accounting is inconsistent anyway, so going over by a few
		// pages is certainly possible. Just make sure we're within some bound.
		// Note that to avoid flakiness due to #52433 (especially since we're allocating

  std::string left_subs;
  for (const char s : input_subs) {
    if (!reduced_label_set.contains(s)) {
      left_subs.push_back(s);
    }
  }

  // Compute the gradient wrt the input, without accounting for the operation
  // "abc->ac". So, now we have the VJP of the operation "ac,cd->ad".
  Output grad_reduced =
      Einsum(scope, {output_grad, other_operand},

		//
		// The reason we do this instead of just waiting for those events to finish and push updates
		// is to ensure that all the time we're accounting for happened sometime between lastUpdate
		// and now. This dramatically simplifies reasoning about the limiter because we're not at
		// risk of extra time being accounted for in this window than actually happened in this window,

//
// If typ.manual() == false, allocSpan allocates a heap span of class spanclass
// and updates heap accounting. If manual == true, allocSpan allocates a
// manually-managed span (spanclass is ignored), and the caller is
// responsible for any accounting related to its use of the span. Either
// way, allocSpan will atomically add the bytes in the newly allocated
// span to *sysStat.
//

		// Failed to allocate.
		mp.mallocing = 0
		releasem(mp)
		return nil
	}
	if s.needzero != 0 {
		throw("arena chunk needs zeroing, but should already be zeroed")
	}
	// Set up heap bitmap and do extra accounting.
	if typ.Pointers() {
		if cap >= 0 {
			userArenaHeapBitsSetSliceType(typ, cap, ptr, s)
		} else {
			userArenaHeapBitsSetType(typ, ptr, s)
		}
		c := getMCache(mp)
		if c == nil {

			//
			// Instead, check against a much more reasonable upper-bound: the amount of
			// mapped heap memory. We can't possibly overcount to the point of exceeding
			// total mapped heap memory, except if there's an accounting bug.
			if live := samples[i].Value.Uint64(); live > mstats.HeapSys {
				t.Errorf("live bytes: %d > heap sys: %d", live, mstats.HeapSys)
			} else if live == 0 {
				// Might happen if we don't call runtime.GC() above.

	estimate := int64(8 * BatchSize * ArraySize * RecursionDepth) // 8 is to reduce flakiness.
	if consumed > estimate {
		t.Fatalf("Stack mem: want %v, got %v", estimate, consumed)
	}
	// Due to broken stack memory accounting (https://golang.org/issue/7468),
	// StackInuse can decrease during function execution, so we cast the values to int64.
	inuse := int64(s1.StackInuse) - int64(s0.StackInuse)
	t.Logf("Inuse %vMB for stack mem", inuse>>20)

	pending []whitespace // pending whitespace
	lastTok token        // last token (after any pending semi) processed by print
}

// write is a thin wrapper around p.output.Write
// that takes care of accounting and error handling.
func (p *printer) write(data []byte) {
	n, err := p.output.Write(data)
	p.written += n
	if err != nil {
		panic(writeError{err})
	}
}

var (
	tabBytes    = []byte("\t\t\t\t\t\t\t\t")

			// Optimization: we've encountered an ASCII character that's not a letter
			// or number. Avoid the call into s.next() and corresponding set up.
			//
			// Note that s.next() does some line accounting if s.ch is '\n', so this
			// shortcut is only possible because we know that the preceding character
			// is not '\n'.
			s.ch = rune(b)
			s.offset = s.rdOffset
			s.rdOffset++
			goto exit
		}